ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
6b45d475e9
openmpi/orte/mca/oob/tcp/oob_tcp_sendrecv.c

659 строки
30 KiB
C
Исходник Обычный вид История

As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
/*
* Copyright (c) 2004-2010 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2011 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2006-2013 Los Alamos National Security, LLC.
* All rights reserved.
* Copyright (c) 2009 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2011 Oak Ridge National Labs. All rights reserved.
Track the origin of a message so it can be passed across transports Refs trac:4184 This commit was SVN r30433. The following Trac tickets were found above: Ticket 4184 --> https://svn.open-mpi.org/trac/ompi/ticket/4184
2014-01-27 01:09:26 +04:00
* Copyright (c) 2013-2014 Intel, Inc. All rights reserved.
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*
* In windows, many of the socket functions return an EWOULDBLOCK
* instead of \ things like EAGAIN, EINPROGRESS, etc. It has been
* verified that this will \ not conflict with other error codes that
* are returned by these functions \ under UNIX/Linux environments
*/
#include "orte_config.h"
#ifdef HAVE_UNISTD_H
#include <unistd.h>
#endif
#include <fcntl.h>
#ifdef HAVE_SYS_UIO_H
#include <sys/uio.h>
#endif
#ifdef HAVE_NET_UIO_H
#include <net/uio.h>
#endif
#ifdef HAVE_SYS_TYPES_H
#include <sys/types.h>
#endif
#include "opal/opal_socket_errno.h"
#ifdef HAVE_NETINET_IN_H
#include <netinet/in.h>
#endif
#ifdef HAVE_ARPA_INET_H
#include <arpa/inet.h>
#endif
#ifdef HAVE_NETINET_TCP_H
#include <netinet/tcp.h>
#endif
#include "opal_stdint.h"
#include "opal/types.h"
#include "opal/mca/backtrace/backtrace.h"
#include "opal/util/output.h"
#include "opal/util/net.h"
#include "opal/util/error.h"
#include "opal/class/opal_hash_table.h"
#include "opal/mca/event/event.h"
#include "orte/util/name_fns.h"
#include "orte/runtime/orte_globals.h"
#include "orte/mca/errmgr/errmgr.h"
#include "orte/mca/ess/ess.h"
#include "orte/mca/routed/routed.h"
#include "orte/mca/state/state.h"
#include "orte/runtime/orte_wait.h"
#include "oob_tcp.h"
#include "orte/mca/oob/tcp/oob_tcp_component.h"
#include "orte/mca/oob/tcp/oob_tcp_peer.h"
#include "orte/mca/oob/tcp/oob_tcp_common.h"
#include "orte/mca/oob/tcp/oob_tcp_connection.h"
static int send_bytes(mca_oob_tcp_peer_t* peer)
{
mca_oob_tcp_send_t* msg = peer->send_msg;
int rc;
while (0 < msg->sdbytes) {
rc = write(peer->sd, msg->sdptr, msg->sdbytes);
if (rc < 0) {
if (opal_socket_errno == EINTR) {
continue;
} else if (opal_socket_errno == EAGAIN) {
/* tell the caller to keep this message on active,
* but let the event lib cycle so other messages
* can progress while this socket is busy
*/
return ORTE_ERR_RESOURCE_BUSY;
} else if (opal_socket_errno == EWOULDBLOCK) {
/* tell the caller to keep this message on active,
* but let the event lib cycle so other messages
* can progress while this socket is busy
*/
return ORTE_ERR_WOULD_BLOCK;
}
/* we hit an error and cannot progress this message */
opal_output(0, "%s->%s mca_oob_tcp_msg_send_bytes: write failed: %s (%d) [sd = %d]",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer->name)),
strerror(opal_socket_errno),
opal_socket_errno,
peer->sd);
return ORTE_ERR_COMM_FAILURE;
}
/* update location */
msg->sdbytes -= rc;
msg->sdptr += rc;
}
/* we sent the full data block */
return ORTE_SUCCESS;
}
/*
* A file descriptor is available/ready for send. Check the state
* of the socket and take the appropriate action.
*/
void mca_oob_tcp_send_handler(int sd, short flags, void *cbdata)
{
mca_oob_tcp_peer_t* peer = (mca_oob_tcp_peer_t*)cbdata;
mca_oob_tcp_send_t* msg = peer->send_msg;
mca_oob_tcp_module_t *mod = (mca_oob_tcp_module_t*)peer->mod;
int rc;
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s tcp:send_handler called to send to peer %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&peer->name));
switch (peer->state) {
case MCA_OOB_TCP_CONNECTING:
case MCA_OOB_TCP_CLOSED:
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s tcp:send_handler %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
mca_oob_tcp_state_print(peer->state));
mca_oob_tcp_peer_complete_connect(mod, peer);
/* de-activate the send event until the connection
* handshake completes
*/
if (peer->send_ev_active) {
opal_event_del(&peer->send_event);
peer->send_ev_active = false;
}
break;
case MCA_OOB_TCP_CONNECTED:
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s tcp:send_handler SENDING TO %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(NULL == peer->send_msg) ? "NULL" : ORTE_NAME_PRINT(&peer->name));
if (NULL != msg) {
/* if the header hasn't been completely sent, send it */
if (!msg->hdr_sent) {
if (ORTE_SUCCESS == (rc = send_bytes(peer))) {
/* header is completely sent */
msg->hdr_sent = true;
/* setup to send the data */
if (NULL != msg->data) {
/* relay msg - send that data */
msg->sdptr = msg->data;
msg->sdbytes = (int)ntohl(msg->hdr.nbytes);
} else if (NULL == msg->msg) {
/* this was a zero-byte relay - nothing more to do */
OBJ_RELEASE(msg);
peer->send_msg = NULL;
goto next;
} else if (NULL != msg->msg->buffer) {
/* send the buffer data as a single block */
msg->sdptr = msg->msg->buffer->base_ptr;
msg->sdbytes = msg->msg->buffer->bytes_used;
Track the origin of a message so it can be passed across transports Refs trac:4184 This commit was SVN r30433. The following Trac tickets were found above: Ticket 4184 --> https://svn.open-mpi.org/trac/ompi/ticket/4184
2014-01-27 01:09:26 +04:00
} else if (NULL != msg->msg->iov) {
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
/* start with the first iovec */
msg->sdptr = msg->msg->iov[0].iov_base;
msg->sdbytes = msg->msg->iov[0].iov_len;
msg->iovnum = 0;
Track the origin of a message so it can be passed across transports Refs trac:4184 This commit was SVN r30433. The following Trac tickets were found above: Ticket 4184 --> https://svn.open-mpi.org/trac/ompi/ticket/4184
2014-01-27 01:09:26 +04:00
} else {
/* just send the data */
msg->sdptr = msg->msg->data;
msg->sdbytes = msg->msg->count;
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
}
/* fall thru and let the send progress */
} else if (ORTE_ERR_RESOURCE_BUSY == rc ||
ORTE_ERR_WOULD_BLOCK == rc) {
/* exit this event and let the event lib progress */
return;
} else {
// report the error
opal_output(0, "%s-%s mca_oob_tcp_peer_send_handler: unable to send header",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer->name)));
opal_event_del(&peer->send_event);
msg->msg->status = rc;
ORTE_RML_SEND_COMPLETE(msg->msg);
OBJ_RELEASE(msg);
peer->send_msg = NULL;
goto next;
}
}
/* progress the data transmission */
if (msg->hdr_sent) {
if (ORTE_SUCCESS == (rc = send_bytes(peer))) {
/* this block is complete */
if (NULL != msg->data || NULL == msg->msg) {
/* the relay is complete - release the data */
opal_output_verbose(2, orte_oob_base_framework.framework_output,
Fix a race condition created when two processes attempt to send to each other at the same time. This causes both processes to start connection procedures, resulting in a c onflict that can cause messages to be lost. Add detection of this condition, and have both processes cancel their connect operations. The process with the higher rank will reconnect, while the lower rank process will simply wait for the connection to be created. Refs trac:3696 This commit was SVN r29139. The following Trac tickets were found above: Ticket 3696 --> https://svn.open-mpi.org/trac/ompi/ticket/3696
2013-09-06 09:15:25 +04:00
"%s MESSAGE RELAY COMPLETE TO %s OF %d BYTES ON SOCKET %d",
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer->name)),
Fix a race condition created when two processes attempt to send to each other at the same time. This causes both processes to start connection procedures, resulting in a c onflict that can cause messages to be lost. Add detection of this condition, and have both processes cancel their connect operations. The process with the higher rank will reconnect, while the lower rank process will simply wait for the connection to be created. Refs trac:3696 This commit was SVN r29139. The following Trac tickets were found above: Ticket 3696 --> https://svn.open-mpi.org/trac/ompi/ticket/3696
2013-09-06 09:15:25 +04:00
(int)ntohl(msg->hdr.nbytes), peer->sd);
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
OBJ_RELEASE(msg);
peer->send_msg = NULL;
} else if (NULL != msg->msg->buffer) {
/* we are done - notify the RML */
opal_output_verbose(2, orte_oob_base_framework.framework_output,
Fix a race condition created when two processes attempt to send to each other at the same time. This causes both processes to start connection procedures, resulting in a c onflict that can cause messages to be lost. Add detection of this condition, and have both processes cancel their connect operations. The process with the higher rank will reconnect, while the lower rank process will simply wait for the connection to be created. Refs trac:3696 This commit was SVN r29139. The following Trac tickets were found above: Ticket 3696 --> https://svn.open-mpi.org/trac/ompi/ticket/3696
2013-09-06 09:15:25 +04:00
"%s MESSAGE SEND COMPLETE TO %s OF %d BYTES ON SOCKET %d",
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer->name)),
Fix a race condition created when two processes attempt to send to each other at the same time. This causes both processes to start connection procedures, resulting in a c onflict that can cause messages to be lost. Add detection of this condition, and have both processes cancel their connect operations. The process with the higher rank will reconnect, while the lower rank process will simply wait for the connection to be created. Refs trac:3696 This commit was SVN r29139. The following Trac tickets were found above: Ticket 3696 --> https://svn.open-mpi.org/trac/ompi/ticket/3696
2013-09-06 09:15:25 +04:00
(int)ntohl(msg->hdr.nbytes), peer->sd);
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
msg->msg->status = ORTE_SUCCESS;
ORTE_RML_SEND_COMPLETE(msg->msg);
OBJ_RELEASE(msg);
peer->send_msg = NULL;
Track the origin of a message so it can be passed across transports Refs trac:4184 This commit was SVN r30433. The following Trac tickets were found above: Ticket 4184 --> https://svn.open-mpi.org/trac/ompi/ticket/4184
2014-01-27 01:09:26 +04:00
} else if (NULL != msg->msg->data) {
/* this was a relay we have now completed - no need to
* notify the RML as the local proc didn't initiate
* the send
*/
opal_output_verbose(2, orte_oob_base_framework.framework_output,
"%s MESSAGE RELAY COMPLETE TO %s OF %d BYTES ON SOCKET %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer->name)),
(int)ntohl(msg->hdr.nbytes), peer->sd);
msg->msg->status = ORTE_SUCCESS;
OBJ_RELEASE(msg);
peer->send_msg = NULL;
} else {
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
/* rotate to the next iovec */
msg->iovnum++;
if (msg->iovnum < msg->msg->count) {
msg->sdptr = msg->msg->iov[msg->iovnum].iov_base;
msg->sdbytes = msg->msg->iov[msg->iovnum].iov_len;
/* exit this event to give the event lib
* a chance to progress any other pending
* actions
*/
return;
} else {
/* this message is complete - notify the RML */
opal_output_verbose(2, orte_oob_base_framework.framework_output,
Fix a race condition created when two processes attempt to send to each other at the same time. This causes both processes to start connection procedures, resulting in a c onflict that can cause messages to be lost. Add detection of this condition, and have both processes cancel their connect operations. The process with the higher rank will reconnect, while the lower rank process will simply wait for the connection to be created. Refs trac:3696 This commit was SVN r29139. The following Trac tickets were found above: Ticket 3696 --> https://svn.open-mpi.org/trac/ompi/ticket/3696
2013-09-06 09:15:25 +04:00
"%s MESSAGE SEND COMPLETE TO %s OF %d BYTES ON SOCKET %d",
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer->name)),
Fix a race condition created when two processes attempt to send to each other at the same time. This causes both processes to start connection procedures, resulting in a c onflict that can cause messages to be lost. Add detection of this condition, and have both processes cancel their connect operations. The process with the higher rank will reconnect, while the lower rank process will simply wait for the connection to be created. Refs trac:3696 This commit was SVN r29139. The following Trac tickets were found above: Ticket 3696 --> https://svn.open-mpi.org/trac/ompi/ticket/3696
2013-09-06 09:15:25 +04:00
(int)ntohl(msg->hdr.nbytes), peer->sd);
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
msg->msg->status = ORTE_SUCCESS;
ORTE_RML_SEND_COMPLETE(msg->msg);
OBJ_RELEASE(msg);
peer->send_msg = NULL;
}
}
/* fall thru to queue the next message */
} else if (ORTE_ERR_RESOURCE_BUSY == rc ||
ORTE_ERR_WOULD_BLOCK == rc) {
/* exit this event and let the event lib progress */
return;
} else {
// report the error
Fix a race condition created when two processes attempt to send to each other at the same time. This causes both processes to start connection procedures, resulting in a c onflict that can cause messages to be lost. Add detection of this condition, and have both processes cancel their connect operations. The process with the higher rank will reconnect, while the lower rank process will simply wait for the connection to be created. Refs trac:3696 This commit was SVN r29139. The following Trac tickets were found above: Ticket 3696 --> https://svn.open-mpi.org/trac/ompi/ticket/3696
2013-09-06 09:15:25 +04:00
opal_output(0, "%s-%s mca_oob_tcp_peer_send_handler: unable to send message ON SOCKET %d",
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
Fix a race condition created when two processes attempt to send to each other at the same time. This causes both processes to start connection procedures, resulting in a c onflict that can cause messages to be lost. Add detection of this condition, and have both processes cancel their connect operations. The process with the higher rank will reconnect, while the lower rank process will simply wait for the connection to be created. Refs trac:3696 This commit was SVN r29139. The following Trac tickets were found above: Ticket 3696 --> https://svn.open-mpi.org/trac/ompi/ticket/3696
2013-09-06 09:15:25 +04:00
ORTE_NAME_PRINT(&(peer->name)), peer->sd);
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
opal_event_del(&peer->send_event);
msg->msg->status = rc;
ORTE_RML_SEND_COMPLETE(msg->msg);
OBJ_RELEASE(msg);
peer->send_msg = NULL;
ORTE_FORCED_TERMINATE(1);
return;
}
}
next:
/* if current message completed - progress any pending sends by
* moving the next in the queue into the "on-deck" position. Note
* that this doesn't mean we send the message right now - we will
* wait for another send_event to fire before doing so. This gives
* us a chance to service any pending recvs.
*/
peer->send_msg = (mca_oob_tcp_send_t*)
opal_list_remove_first(&peer->send_queue);
}
/* if nothing else to do unregister for send event notifications */
if (NULL == peer->send_msg && peer->send_ev_active) {
opal_event_del(&peer->send_event);
peer->send_ev_active = false;
}
break;
default:
Fix a race condition created when two processes attempt to send to each other at the same time. This causes both processes to start connection procedures, resulting in a c onflict that can cause messages to be lost. Add detection of this condition, and have both processes cancel their connect operations. The process with the higher rank will reconnect, while the lower rank process will simply wait for the connection to be created. Refs trac:3696 This commit was SVN r29139. The following Trac tickets were found above: Ticket 3696 --> https://svn.open-mpi.org/trac/ompi/ticket/3696
2013-09-06 09:15:25 +04:00
opal_output(0, "%s-%s mca_oob_tcp_peer_send_handler: invalid connection state (%d) on socket %d",
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer->name)),
Fix a race condition created when two processes attempt to send to each other at the same time. This causes both processes to start connection procedures, resulting in a c onflict that can cause messages to be lost. Add detection of this condition, and have both processes cancel their connect operations. The process with the higher rank will reconnect, while the lower rank process will simply wait for the connection to be created. Refs trac:3696 This commit was SVN r29139. The following Trac tickets were found above: Ticket 3696 --> https://svn.open-mpi.org/trac/ompi/ticket/3696
2013-09-06 09:15:25 +04:00
peer->state, peer->sd);
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
if (peer->send_ev_active) {
opal_event_del(&peer->send_event);
peer->send_ev_active = false;
}
break;
}
}
static int read_bytes(mca_oob_tcp_peer_t* peer)
{
int rc;
Remove debug - problem has been identified Refs trac:4026 This commit was SVN r30075. The following Trac tickets were found above: Ticket 4026 --> https://svn.open-mpi.org/trac/ompi/ticket/4026
2013-12-24 19:22:18 +04:00
/* read until all bytes recvd or error */
while (0 < peer->recv_msg->rdbytes) {
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
rc = read(peer->sd, peer->recv_msg->rdptr, peer->recv_msg->rdbytes);
if (rc < 0) {
if(opal_socket_errno == EINTR) {
continue;
} else if (opal_socket_errno == EAGAIN) {
/* tell the caller to keep this message on active,
* but let the event lib cycle so other messages
* can progress while this socket is busy
*/
return ORTE_ERR_RESOURCE_BUSY;
} else if (opal_socket_errno == EWOULDBLOCK) {
/* tell the caller to keep this message on active,
* but let the event lib cycle so other messages
* can progress while this socket is busy
*/
return ORTE_ERR_WOULD_BLOCK;
}
/* we hit an error and cannot progress this message - report
* the error back to the RML and let the caller know
* to abort this message
*/
Attempt to debug hang that is hitting some environments. Posting to 1.7.4 as a placeholder for the eventual solution cmr=v1.7.4:reviewer=rhc This commit was SVN r30060.
2013-12-23 23:57:05 +04:00
opal_output_verbose(OOB_TCP_DEBUG_FAIL, orte_oob_base_framework.framework_output,
"%s-%s mca_oob_tcp_msg_recv: readv failed: %s (%d)",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer->name)),
strerror(opal_socket_errno),
opal_socket_errno);
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
// mca_oob_tcp_peer_close(peer);
// if (NULL != mca_oob_tcp.oob_exception_callback) {
// mca_oob_tcp.oob_exception_callback(&peer->name, ORTE_RML_PEER_DISCONNECTED);
//}
return ORTE_ERR_COMM_FAILURE;
} else if (rc == 0) {
/* the remote peer closed the connection - report that condition
* and let the caller know
*/
Attempt to debug hang that is hitting some environments. Posting to 1.7.4 as a placeholder for the eventual solution cmr=v1.7.4:reviewer=rhc This commit was SVN r30060.
2013-12-23 23:57:05 +04:00
opal_output_verbose(OOB_TCP_DEBUG_FAIL, orte_oob_base_framework.framework_output,
"%s-%s mca_oob_tcp_msg_recv: peer closed connection",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer->name)));
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
/* stop all events */
if (peer->recv_ev_active) {
opal_event_del(&peer->recv_event);
peer->recv_ev_active = false;
}
if (peer->timer_ev_active) {
opal_event_del(&peer->timer_event);
peer->timer_ev_active = false;
}
if (peer->send_ev_active) {
opal_event_del(&peer->send_event);
peer->send_ev_active = false;
}
if (NULL != peer->recv_msg) {
OBJ_RELEASE(peer->recv_msg);
peer->recv_msg = NULL;
}
mca_oob_tcp_peer_close((mca_oob_tcp_module_t*)peer->mod, peer);
//if (NULL != mca_oob_tcp.oob_exception_callback) {
// mca_oob_tcp.oob_exception_callback(&peer->peer_name, ORTE_RML_PEER_DISCONNECTED);
//}
return ORTE_ERR_WOULD_BLOCK;
}
/* we were able to read something, so adjust counters and location */
peer->recv_msg->rdbytes -= rc;
peer->recv_msg->rdptr += rc;
}
Attempt to debug hang that is hitting some environments. Posting to 1.7.4 as a placeholder for the eventual solution cmr=v1.7.4:reviewer=rhc This commit was SVN r30060.
2013-12-23 23:57:05 +04:00
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
/* we read the full data block */
return ORTE_SUCCESS;
}
/*
* Dispatch to the appropriate action routine based on the state
* of the connection with the peer.
*/
void mca_oob_tcp_recv_handler(int sd, short flags, void *cbdata)
{
mca_oob_tcp_peer_t* peer = (mca_oob_tcp_peer_t*)cbdata;
mca_oob_tcp_module_t *mod = (mca_oob_tcp_module_t*)peer->mod;
int rc;
orte_process_name_t hop;
mca_oob_tcp_peer_t *relay;
uint64_t ui64;
if (orte_abnormal_term_ordered) {
return;
}
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s:tcp:recv:handler called for peer %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&peer->name));
switch (peer->state) {
case MCA_OOB_TCP_CONNECT_ACK:
Upgrade the security framework to avoid multiple hits against the global security server. Add support for future case where mpirun assings a global security credential for a given run, though we need to work out how to handle connect-accept from other mpirun's in that case. Remove a bunch of duplicate code in the OOB by consolidating the connection handshake code. Refs trac:4221 This commit was SVN r30554. The following Trac tickets were found above: Ticket 4221 --> https://svn.open-mpi.org/trac/ompi/ticket/4221
2014-02-04 18:47:04 +04:00
if (ORTE_SUCCESS == (rc = mca_oob_tcp_peer_recv_connect_ack(mod, peer, peer->sd, NULL))) {
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s:tcp:recv:handler starting send/recv events",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
/* we connected! Start the send/recv events */
if (!peer->recv_ev_active) {
opal_event_add(&peer->recv_event, 0);
peer->recv_ev_active = true;
}
if (peer->timer_ev_active) {
opal_event_del(&peer->timer_event);
peer->timer_ev_active = false;
}
/* if there is a message waiting to be sent, queue it */
if (NULL == peer->send_msg) {
peer->send_msg = (mca_oob_tcp_send_t*)opal_list_remove_first(&peer->send_queue);
}
if (NULL != peer->send_msg && !peer->send_ev_active) {
opal_event_add(&peer->send_event, 0);
peer->send_ev_active = true;
}
/* update our state */
peer->state = MCA_OOB_TCP_CONNECTED;
} else {
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s UNABLE TO COMPLETE CONNECT ACK WITH %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&peer->name));
opal_event_del(&peer->recv_event);
ORTE_FORCED_TERMINATE(1);
return;
}
break;
case MCA_OOB_TCP_CONNECTED:
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s:tcp:recv:handler CONNECTED",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
/* allocate a new message and setup for recv */
if (NULL == peer->recv_msg) {
Fix a race condition created when two processes attempt to send to each other at the same time. This causes both processes to start connection procedures, resulting in a c onflict that can cause messages to be lost. Add detection of this condition, and have both processes cancel their connect operations. The process with the higher rank will reconnect, while the lower rank process will simply wait for the connection to be created. Refs trac:3696 This commit was SVN r29139. The following Trac tickets were found above: Ticket 3696 --> https://svn.open-mpi.org/trac/ompi/ticket/3696
2013-09-06 09:15:25 +04:00
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s:tcp:recv:handler allocate new recv msg",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
peer->recv_msg = OBJ_NEW(mca_oob_tcp_recv_t);
if (NULL == peer->recv_msg) {
opal_output(0, "%s-%s mca_oob_tcp_peer_recv_handler: unable to allocate recv message\n",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer->name)));
return;
}
/* start by reading the header */
peer->recv_msg->rdptr = (char*)&peer->recv_msg->hdr;
peer->recv_msg->rdbytes = sizeof(mca_oob_tcp_hdr_t);
}
/* if the header hasn't been completely read, read it */
if (!peer->recv_msg->hdr_recvd) {
Fix a race condition created when two processes attempt to send to each other at the same time. This causes both processes to start connection procedures, resulting in a c onflict that can cause messages to be lost. Add detection of this condition, and have both processes cancel their connect operations. The process with the higher rank will reconnect, while the lower rank process will simply wait for the connection to be created. Refs trac:3696 This commit was SVN r29139. The following Trac tickets were found above: Ticket 3696 --> https://svn.open-mpi.org/trac/ompi/ticket/3696
2013-09-06 09:15:25 +04:00
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s:tcp:recv:handler read hdr",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
if (ORTE_SUCCESS == (rc = read_bytes(peer))) {
/* completed reading the header */
peer->recv_msg->hdr_recvd = true;
/* convert the header */
MCA_OOB_TCP_HDR_NTOH(&peer->recv_msg->hdr);
/* if this is a zero-byte message, then we are done */
if (0 == peer->recv_msg->hdr.nbytes) {
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s RECVD ZERO-BYTE MESSAGE FROM %s for tag %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&peer->name), peer->recv_msg->hdr.tag);
peer->recv_msg->data = NULL; // make sure
} else {
Fix a race condition created when two processes attempt to send to each other at the same time. This causes both processes to start connection procedures, resulting in a c onflict that can cause messages to be lost. Add detection of this condition, and have both processes cancel their connect operations. The process with the higher rank will reconnect, while the lower rank process will simply wait for the connection to be created. Refs trac:3696 This commit was SVN r29139. The following Trac tickets were found above: Ticket 3696 --> https://svn.open-mpi.org/trac/ompi/ticket/3696
2013-09-06 09:15:25 +04:00
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
Attempt to debug hang that is hitting some environments. Posting to 1.7.4 as a placeholder for the eventual solution cmr=v1.7.4:reviewer=rhc This commit was SVN r30060.
2013-12-23 23:57:05 +04:00
"%s:tcp:recv:handler allocate data region of size %lu",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), (unsigned long)peer->recv_msg->hdr.nbytes);
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
/* allocate the data region */
peer->recv_msg->data = (char*)malloc(peer->recv_msg->hdr.nbytes);
/* point to it */
peer->recv_msg->rdptr = peer->recv_msg->data;
peer->recv_msg->rdbytes = peer->recv_msg->hdr.nbytes;
}
/* fall thru and attempt to read the data */
} else if (ORTE_ERR_RESOURCE_BUSY == rc ||
ORTE_ERR_WOULD_BLOCK == rc) {
/* exit this event and let the event lib progress */
return;
} else {
/* close the connection */
Fix a race condition created when two processes attempt to send to each other at the same time. This causes both processes to start connection procedures, resulting in a c onflict that can cause messages to be lost. Add detection of this condition, and have both processes cancel their connect operations. The process with the higher rank will reconnect, while the lower rank process will simply wait for the connection to be created. Refs trac:3696 This commit was SVN r29139. The following Trac tickets were found above: Ticket 3696 --> https://svn.open-mpi.org/trac/ompi/ticket/3696
2013-09-06 09:15:25 +04:00
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s:tcp:recv:handler error reading bytes - closing connection",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
mca_oob_tcp_peer_close(mod, peer);
return;
}
}
if (peer->recv_msg->hdr_recvd) {
/* continue to read the data block - we start from
* wherever we left off, which could be at the
* beginning or somewhere in the message
*/
if (ORTE_SUCCESS == (rc = read_bytes(peer))) {
/* we recvd all of the message */
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s RECVD COMPLETE MESSAGE FROM %s (ORIGIN %s) OF %d BYTES FOR DEST %s TAG %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&peer->name),
ORTE_NAME_PRINT(&peer->recv_msg->hdr.origin),
(int)peer->recv_msg->hdr.nbytes,
ORTE_NAME_PRINT(&peer->recv_msg->hdr.dst),
peer->recv_msg->hdr.tag);
/* am I the intended recipient (header was already converted back to host order)? */
if (peer->recv_msg->hdr.dst.jobid == ORTE_PROC_MY_NAME->jobid &&
peer->recv_msg->hdr.dst.vpid == ORTE_PROC_MY_NAME->vpid) {
/* yes - post it to the RML for delivery */
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s DELIVERING TO RML",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
ORTE_RML_POST_MESSAGE(&peer->recv_msg->hdr.origin, peer->recv_msg->hdr.tag,
peer->recv_msg->data,
peer->recv_msg->hdr.nbytes);
OBJ_RELEASE(peer->recv_msg);
} else {
/* no - find the next hop in the route */
hop = orte_routed.get_route(&peer->recv_msg->hdr.dst);
if (hop.jobid == ORTE_JOBID_INVALID ||
hop.vpid == ORTE_VPID_INVALID) {
/* no hop known - post the error to the component
* and let the OOB see if there is another way
* to get there from here
*/
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s NO ROUTE TO %s FROM HERE",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&peer->name));
/* let the component know about the problem */
ORTE_ACTIVATE_TCP_MSG_ERROR(mod, NULL, peer->recv_msg, &hop, mca_oob_tcp_component_no_route);
/* cleanup */
OBJ_RELEASE(peer->recv_msg);
return;
} else {
/* does this module know how to reach the next hop? */
memcpy(&ui64, (char*)&hop, sizeof(uint64_t));
if (OPAL_SUCCESS != opal_hash_table_get_value_uint64(&mod->peers, ui64, (void**)&relay)) {
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s ADDRESS OF NEXT HOP %s TO %s IS UNKNOWN VIA MODULE %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&hop),
ORTE_NAME_PRINT(&peer->recv_msg->hdr.dst),
peer->mod->if_name);
/* let the component know about the problem */
ORTE_ACTIVATE_TCP_MSG_ERROR(mod, NULL, peer->recv_msg, &hop, mca_oob_tcp_component_hop_unknown);
/* cleanup */
OBJ_RELEASE(peer->recv_msg);
return;
}
opal_output_verbose(OOB_TCP_DEBUG_CONNECT, orte_oob_base_framework.framework_output,
"%s ROUTING TO %s FROM HERE",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&relay->name));
Extend the dpm framework API to support persistent accept/connect operations: * paccept - establish a persistent listening port for async connect requests * pconnect - async connect to remote process that has posted a paccept port. Provides a timeout mechanism, and allows the underlying implementation to retry until timeout * pclose - shuts down a prior paccept posting Includes example programs paccept.c and pconnect.c in orte/test/mpi. New MPI extension interfaces coming... This commit was SVN r29063.
2013-08-23 22:02:50 +04:00
/* if this came from a different job family, then ensure
* we know how to return
*/
if (ORTE_JOB_FAMILY(peer->recv_msg->hdr.origin.jobid) != ORTE_JOB_FAMILY(ORTE_PROC_MY_NAME->jobid)) {
orte_routed.update_route(&(peer->recv_msg->hdr.origin), &peer->name);
}
/* post the message for retransmission */
As per the RFC, bring in the ORTE async progress code and the rewrite of OOB: *** THIS RFC INCLUDES A MINOR CHANGE TO THE MPI-RTE INTERFACE *** Note: during the course of this work, it was necessary to completely separate the MPI and RTE progress engines. There were multiple places in the MPI layer where ORTE_WAIT_FOR_COMPLETION was being used. A new OMPI_WAIT_FOR_COMPLETION macro was created (defined in ompi/mca/rte/rte.h) that simply cycles across opal_progress until the provided flag becomes false. Places where the MPI layer blocked waiting for RTE to complete an event have been modified to use this macro. *************************************************************************************** I am reissuing this RFC because of the time that has passed since its original release. Since its initial release and review, I have debugged it further to ensure it fully supports tests like loop_spawn. It therefore seems ready for merge back to the trunk. Given its prior review, I have set the timeout for one week. The code is in https://bitbucket.org/rhc/ompi-oob2 WHAT: Rewrite of ORTE OOB WHY: Support asynchronous progress and a host of other features WHEN: Wed, August 21 SYNOPSIS: The current OOB has served us well, but a number of limitations have been identified over the years. Specifically: * it is only progressed when called via opal_progress, which can lead to hangs or recursive calls into libevent (which is not supported by that code) * we've had issues when multiple NICs are available as the code doesn't "shift" messages between transports - thus, all nodes had to be available via the same TCP interface. * the OOB "unloads" incoming opal_buffer_t objects during the transmission, thus preventing use of OBJ_RETAIN in the code when repeatedly sending the same message to multiple recipients * there is no failover mechanism across NICs - if the selected NIC (or its attached switch) fails, we are forced to abort * only one transport (i.e., component) can be "active" The revised OOB resolves these problems: * async progress is used for all application processes, with the progress thread blocking in the event library * each available TCP NIC is supported by its own TCP module. The ability to asynchronously progress each module independently is provided, but not enabled by default (a runtime MCA parameter turns it "on") * multi-address TCP NICs (e.g., a NIC with both an IPv4 and IPv6 address, or with virtual interfaces) are supported - reachability is determined by comparing the contact info for a peer against all addresses within the range covered by the address/mask pairs for the NIC. * a message that arrives on one TCP NIC is automatically shifted to whatever NIC that is connected to the next "hop" if that peer cannot be reached by the incoming NIC. If no TCP module will reach the peer, then the OOB attempts to send the message via all other available components - if none can reach the peer, then an "error" is reported back to the RML, which then calls the errmgr for instructions. * opal_buffer_t now conforms to standard object rules re OBJ_RETAIN as we no longer "unload" the incoming object * NIC failure is reported to the TCP component, which then tries to resend the message across any other available TCP NIC. If that doesn't work, then the message is given back to the OOB base to try using other components. If all that fails, then the error is reported to the RML, which reports to the errmgr for instructions * obviously from the above, multiple OOB components (e.g., TCP and UD) can be active in parallel * the matching code has been moved to the RML (and out of the OOB/TCP component) so it is independent of transport * routing is done by the individual OOB modules (as opposed to the RML). Thus, both routed and non-routed transports can simultaneously be active * all blocking send/recv APIs have been removed. Everything operates asynchronously. KNOWN LIMITATIONS: * although provision is made for component failover as described above, the code for doing so has not been fully implemented yet. At the moment, if all connections for a given peer fail, the errmgr is notified of a "lost connection", which by default results in termination of the job if it was a lifeline * the IPv6 code is present and compiles, but is not complete. Since the current IPv6 support in the OOB doesn't work anyway, I don't consider this a blocker * routing is performed at the individual module level, yet the active routed component is selected on a global basis. We probably should update that to reflect that different transports may need/choose to route in different ways * obviously, not every error path has been tested nor necessarily covered * determining abnormal termination is more challenging than in the old code as we now potentially have multiple ways of connecting to a process. Ideally, we would declare "connection failed" when *all* transports can no longer reach the process, but that requires some additional (possibly complex) code. For now, the code replicates the old behavior only somewhat modified - i.e., if a module sees its connection fail, it checks to see if it is a lifeline. If so, it notifies the errmgr that the lifeline is lost - otherwise, it notifies the errmgr that a non-lifeline connection was lost. * reachability is determined solely on the basis of a shared subnet address/mask - more sophisticated algorithms (e.g., the one used in the tcp btl) are required to handle routing via gateways * the RML needs to assign sequence numbers to each message on a per-peer basis. The receiving RML will then deliver messages in order, thus preventing out-of-order messaging in the case where messages travel across different transports or a message needs to be redirected/resent due to failure of a NIC This commit was SVN r29058.
2013-08-22 20:37:40 +04:00
MCA_OOB_TCP_QUEUE_RELAY(peer->recv_msg, relay);
OBJ_RELEASE(peer->recv_msg);
}
}
peer->recv_msg = NULL;
return;
} else if (ORTE_ERR_RESOURCE_BUSY == rc ||
ORTE_ERR_WOULD_BLOCK == rc) {
/* exit this event and let the event lib progress */
return;
} else {
// report the error
opal_output(0, "%s-%s mca_oob_tcp_peer_recv_handler: unable to recv message",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer->name)));
/* turn off the recv event */
opal_event_del(&peer->recv_event);
ORTE_FORCED_TERMINATE(1);
return;
}
}
break;
default:
opal_output(0, "%s-%s mca_oob_tcp_peer_recv_handler: invalid socket state(%d)",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&(peer->name)),
peer->state);
// mca_oob_tcp_peer_close(peer);
break;
}
}
static void snd_cons(mca_oob_tcp_send_t *ptr)
{
ptr->msg = NULL;
ptr->data = NULL;
ptr->hdr_sent = false;
ptr->iovnum = 0;
ptr->sdptr = NULL;
ptr->sdbytes = 0;
}
/* we don't destruct any RML msg that is
* attached to our send as the RML owns
* that memory. However, if we relay a
* msg, the data in the relay belongs to
* us and must be free'd
*/
static void snd_des(mca_oob_tcp_send_t *ptr)
{
if (NULL != ptr->data) {
free(ptr->data);
}
}
OBJ_CLASS_INSTANCE(mca_oob_tcp_send_t,
opal_list_item_t,
snd_cons, snd_des);
static void rcv_cons(mca_oob_tcp_recv_t *ptr)
{
ptr->hdr_recvd = false;
ptr->rdptr = NULL;
ptr->rdbytes = 0;
}
OBJ_CLASS_INSTANCE(mca_oob_tcp_recv_t,
opal_list_item_t,
rcv_cons, NULL);
static void err_cons(mca_oob_tcp_msg_error_t *ptr)
{
ptr->rmsg = NULL;
ptr->snd = NULL;
}
OBJ_CLASS_INSTANCE(mca_oob_tcp_msg_error_t,
opal_object_t,
err_cons, NULL);
Копировать постоянную ссылку