Per the PMIx RFC:
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
2014-08-21 18:56:47 +00:00
|
|
|
/*
|
|
|
|
* 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) 2009-2012 Cisco Systems, Inc. All rights reserved.
|
|
|
|
* Copyright (c) 2011 Oak Ridge National Labs. All rights reserved.
|
2015-03-25 13:17:47 -07:00
|
|
|
* Copyright (c) 2013-2015 Intel, Inc. All rights reserved.
|
Per the PMIx RFC:
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
2014-08-21 18:56:47 +00:00
|
|
|
* $COPYRIGHT$
|
|
|
|
*
|
|
|
|
* Additional copyrights may follow
|
|
|
|
*
|
|
|
|
* $HEADER$
|
|
|
|
*
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include "opal_config.h"
|
|
|
|
#include "opal/types.h"
|
|
|
|
|
|
|
|
#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_stdint.h"
|
|
|
|
#include "opal/opal_socket_errno.h"
|
|
|
|
#include "opal/dss/dss.h"
|
|
|
|
#include "opal/mca/dstore/dstore.h"
|
|
|
|
#include "opal/mca/sec/sec.h"
|
|
|
|
#include "opal/runtime/opal.h"
|
|
|
|
#include "opal/util/show_help.h"
|
|
|
|
#include "opal/util/error.h"
|
|
|
|
#include "opal/util/output.h"
|
|
|
|
#include "opal/util/proc.h"
|
|
|
|
|
|
|
|
#include "opal/mca/pmix/base/base.h"
|
|
|
|
#include "pmix_native.h"
|
|
|
|
|
|
|
|
static int usock_send_blocking(char *ptr, size_t size);
|
|
|
|
static void pmix_usock_try_connect(int fd, short args, void *cbdata);
|
|
|
|
|
|
|
|
/* State machine for internal operations */
|
|
|
|
typedef struct {
|
|
|
|
opal_object_t super;
|
|
|
|
opal_event_t ev;
|
|
|
|
} pmix_usock_op_t;
|
|
|
|
static OBJ_CLASS_INSTANCE(pmix_usock_op_t,
|
|
|
|
opal_object_t,
|
|
|
|
NULL, NULL);
|
|
|
|
|
|
|
|
#define PMIX_ACTIVATE_USOCK_STATE(cbfunc) \
|
|
|
|
do { \
|
|
|
|
pmix_usock_op_t *op; \
|
|
|
|
op = OBJ_NEW(pmix_usock_op_t); \
|
|
|
|
opal_event_set(mca_pmix_native_component.evbase, &op->ev, -1, \
|
|
|
|
OPAL_EV_WRITE, (cbfunc), op); \
|
|
|
|
opal_event_set_priority(&op->ev, OPAL_EV_MSG_LO_PRI); \
|
|
|
|
opal_event_active(&op->ev, OPAL_EV_WRITE, 1); \
|
|
|
|
} while(0);
|
|
|
|
|
|
|
|
void pmix_usock_send_recv(int fd, short args, void *cbdata)
|
|
|
|
{
|
|
|
|
pmix_usock_sr_t *ms = (pmix_usock_sr_t*)cbdata;
|
|
|
|
pmix_usock_posted_recv_t *req;
|
|
|
|
pmix_usock_send_t *snd;
|
|
|
|
uint32_t tag = UINT32_MAX;
|
|
|
|
|
|
|
|
if (NULL != ms->cbfunc) {
|
|
|
|
/* if a callback msg is expected, setup a recv for it */
|
|
|
|
req = OBJ_NEW(pmix_usock_posted_recv_t);
|
|
|
|
/* take the next tag in the sequence */
|
|
|
|
if (UINT32_MAX == mca_pmix_native_component.tag) {
|
|
|
|
mca_pmix_native_component.tag = 0;
|
|
|
|
}
|
|
|
|
req->tag = mca_pmix_native_component.tag++;
|
|
|
|
tag = req->tag;
|
|
|
|
req->cbfunc = ms->cbfunc;
|
|
|
|
req->cbdata = ms->cbdata;
|
|
|
|
opal_output_verbose(5, opal_pmix_base_framework.framework_output,
|
|
|
|
"%s posting recv on tag %d",
|
|
|
|
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME), req->tag);
|
|
|
|
/* add it to the list of recvs - we cannot have unexpected messages
|
|
|
|
* in this subsystem as the server never sends us something that
|
|
|
|
* we didn't previously request */
|
|
|
|
opal_list_append(&mca_pmix_native_component.posted_recvs, &req->super);
|
|
|
|
}
|
|
|
|
|
|
|
|
snd = OBJ_NEW(pmix_usock_send_t);
|
|
|
|
snd->hdr.id = mca_pmix_native_component.id;
|
|
|
|
snd->hdr.type = PMIX_USOCK_USER;
|
|
|
|
snd->hdr.tag = tag;
|
|
|
|
snd->hdr.nbytes = ms->bfr->bytes_used;
|
|
|
|
snd->data = ms->bfr->base_ptr;
|
|
|
|
/* always start with the header */
|
|
|
|
snd->sdptr = (char*)&snd->hdr;
|
|
|
|
snd->sdbytes = sizeof(pmix_usock_hdr_t);
|
|
|
|
|
|
|
|
/* add the msg to the send queue if we are already connected*/
|
|
|
|
if (PMIX_USOCK_CONNECTED == mca_pmix_native_component.state) {
|
|
|
|
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
|
|
|
|
"%s usock:send_nb: already connected to server - queueing for send",
|
|
|
|
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
|
|
|
|
/* if there is no message on-deck, put this one there */
|
|
|
|
if (NULL == mca_pmix_native_component.send_msg) {
|
|
|
|
mca_pmix_native_component.send_msg = snd;
|
|
|
|
} else {
|
|
|
|
/* add it to the queue */
|
|
|
|
opal_list_append(&mca_pmix_native_component.send_queue, &snd->super);
|
|
|
|
}
|
|
|
|
/* ensure the send event is active */
|
|
|
|
if (!mca_pmix_native_component.send_ev_active) {
|
|
|
|
opal_event_add(&mca_pmix_native_component.send_event, 0);
|
|
|
|
mca_pmix_native_component.send_ev_active = true;
|
|
|
|
}
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* add the message to the queue for sending after the
|
|
|
|
* connection is formed
|
|
|
|
*/
|
|
|
|
opal_list_append(&mca_pmix_native_component.send_queue, &snd->super);
|
|
|
|
|
|
|
|
if (PMIX_USOCK_CONNECTING != mca_pmix_native_component.state &&
|
|
|
|
PMIX_USOCK_CONNECT_ACK != mca_pmix_native_component.state) {
|
|
|
|
/* we have to initiate the connection - again, we do not
|
|
|
|
* want to block while the connection is created.
|
|
|
|
* So throw us into an event that will create
|
|
|
|
* the connection via a mini-state-machine :-)
|
|
|
|
*/
|
|
|
|
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
|
|
|
|
"%s usock:send_nb: initiating connection to server",
|
|
|
|
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
|
|
|
|
mca_pmix_native_component.state = PMIX_USOCK_CONNECTING;
|
|
|
|
PMIX_ACTIVATE_USOCK_STATE(pmix_usock_try_connect);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void pmix_usock_process_msg(int fd, short flags, void *cbdata)
|
|
|
|
{
|
|
|
|
pmix_usock_recv_t *msg = (pmix_usock_recv_t*)cbdata;
|
|
|
|
pmix_usock_posted_recv_t *rcv;
|
|
|
|
opal_buffer_t buf;
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((5, opal_pmix_base_framework.framework_output,
|
|
|
|
"%s message received %d bytes for tag %u",
|
|
|
|
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
|
|
|
|
(int)msg->hdr.nbytes, msg->hdr.tag));
|
|
|
|
|
|
|
|
/* see if we have a waiting recv for this message */
|
|
|
|
OPAL_LIST_FOREACH(rcv, &mca_pmix_native_component.posted_recvs, pmix_usock_posted_recv_t) {
|
|
|
|
opal_output_verbose(5, opal_pmix_base_framework.framework_output,
|
|
|
|
"%s checking msg on tag %u for tag %u",
|
|
|
|
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
|
|
|
|
msg->hdr.tag, rcv->tag);
|
|
|
|
|
|
|
|
if (msg->hdr.tag == rcv->tag) {
|
|
|
|
if (NULL != rcv->cbfunc) {
|
|
|
|
/* construct and load the buffer */
|
|
|
|
OBJ_CONSTRUCT(&buf, opal_buffer_t);
|
2014-09-06 19:19:44 +00:00
|
|
|
if (NULL != msg->data) {
|
|
|
|
opal_dss.load(&buf, msg->data, msg->hdr.nbytes);
|
|
|
|
}
|
Per the PMIx RFC:
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
2014-08-21 18:56:47 +00:00
|
|
|
msg->data = NULL; // protect the data region
|
|
|
|
if (NULL != rcv->cbfunc) {
|
|
|
|
rcv->cbfunc(&buf, rcv->cbdata);
|
|
|
|
}
|
|
|
|
OBJ_DESTRUCT(&buf); // free's the msg data
|
|
|
|
/* also done with the recv */
|
|
|
|
opal_list_remove_item(&mca_pmix_native_component.posted_recvs, &rcv->super);
|
|
|
|
OBJ_RELEASE(rcv);
|
|
|
|
OBJ_RELEASE(msg);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* we get here if no matching recv was found - this is an error */
|
|
|
|
opal_output(0, "%s UNEXPECTED MESSAGE",
|
|
|
|
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
|
|
|
|
OBJ_RELEASE(msg);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Try connecting to a peer
|
|
|
|
*/
|
|
|
|
static void pmix_usock_try_connect(int fd, short args, void *cbdata)
|
|
|
|
{
|
Fat SMPs (i.e., systems with nodes containing large numbers of cpus) were failing to start due to connection failures of the opal/pmix support. Root cause was that (a) we were setting the client socket to non-blocking before calling connect, and (b) the server was using the event library to harvest the accepts, and also did the handshake while in that event. So the server would backup beyond the connection backlog limit, and we would fail.
Changing the client to leave its socket as blocking during the connect doesn't solve the problem by itself - you also have to introduce a sleep delay once the backlog is hit to avoid simply machine-gunning your way thru retries. This gets somewhat difficult to adjust as you don't want to unnecessarily prolong startup time.
We've solved this before by adding a listening thread that simply reaps accepts and shoves them into the event library for subsequent processing. This would resolve the problem, but meant yet another daemon-level thread. So I centralized the listening thread support and let multiple elements register listeners on it. Thus, each daemon now has a single listening thread that reaps accepts from multiple sources - for now, the orte/pmix server and the oob/usock support are using it. I'll add in the oob/tcp component later.
This still didn't fully resolve the SMP problem, especially on coprocessor cards (e.g., KNC). Removing the shared memory dstore support helped further improve the behavior - it looks like there is some kind of memory paging issue there that needs further understanding. Given that the shared memory support was about to be lost when I bring over the PMIx integration (until it is restored in that library), it seemed like a reasonable thing to just remove it at this point.
2015-05-29 14:28:26 -07:00
|
|
|
int rc, flags;
|
Per the PMIx RFC:
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
2014-08-21 18:56:47 +00:00
|
|
|
opal_socklen_t addrlen = 0;
|
|
|
|
|
|
|
|
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
|
|
|
|
"%s usock_peer_try_connect: attempting to connect to server",
|
|
|
|
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
|
|
|
|
|
|
|
|
addrlen = sizeof(struct sockaddr_un);
|
Fat SMPs (i.e., systems with nodes containing large numbers of cpus) were failing to start due to connection failures of the opal/pmix support. Root cause was that (a) we were setting the client socket to non-blocking before calling connect, and (b) the server was using the event library to harvest the accepts, and also did the handshake while in that event. So the server would backup beyond the connection backlog limit, and we would fail.
Changing the client to leave its socket as blocking during the connect doesn't solve the problem by itself - you also have to introduce a sleep delay once the backlog is hit to avoid simply machine-gunning your way thru retries. This gets somewhat difficult to adjust as you don't want to unnecessarily prolong startup time.
We've solved this before by adding a listening thread that simply reaps accepts and shoves them into the event library for subsequent processing. This would resolve the problem, but meant yet another daemon-level thread. So I centralized the listening thread support and let multiple elements register listeners on it. Thus, each daemon now has a single listening thread that reaps accepts from multiple sources - for now, the orte/pmix server and the oob/usock support are using it. I'll add in the oob/tcp component later.
This still didn't fully resolve the SMP problem, especially on coprocessor cards (e.g., KNC). Removing the shared memory dstore support helped further improve the behavior - it looks like there is some kind of memory paging issue there that needs further understanding. Given that the shared memory support was about to be lost when I bring over the PMIx integration (until it is restored in that library), it seemed like a reasonable thing to just remove it at this point.
2015-05-29 14:28:26 -07:00
|
|
|
|
|
|
|
while (mca_pmix_native_component.retries < mca_pmix_native_component.max_retries) {
|
|
|
|
mca_pmix_native_component.retries++;
|
|
|
|
/* Create the new socket */
|
|
|
|
mca_pmix_native_component.sd = socket(PF_UNIX, SOCK_STREAM, 0);
|
|
|
|
if (mca_pmix_native_component.sd < 0) {
|
|
|
|
opal_output(0, "pmix:create_socket: socket() failed: %s (%d)\n",
|
|
|
|
strerror(opal_socket_errno),
|
|
|
|
opal_socket_errno);
|
|
|
|
continue;
|
Per the PMIx RFC:
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
2014-08-21 18:56:47 +00:00
|
|
|
}
|
Fat SMPs (i.e., systems with nodes containing large numbers of cpus) were failing to start due to connection failures of the opal/pmix support. Root cause was that (a) we were setting the client socket to non-blocking before calling connect, and (b) the server was using the event library to harvest the accepts, and also did the handshake while in that event. So the server would backup beyond the connection backlog limit, and we would fail.
Changing the client to leave its socket as blocking during the connect doesn't solve the problem by itself - you also have to introduce a sleep delay once the backlog is hit to avoid simply machine-gunning your way thru retries. This gets somewhat difficult to adjust as you don't want to unnecessarily prolong startup time.
We've solved this before by adding a listening thread that simply reaps accepts and shoves them into the event library for subsequent processing. This would resolve the problem, but meant yet another daemon-level thread. So I centralized the listening thread support and let multiple elements register listeners on it. Thus, each daemon now has a single listening thread that reaps accepts from multiple sources - for now, the orte/pmix server and the oob/usock support are using it. I'll add in the oob/tcp component later.
This still didn't fully resolve the SMP problem, especially on coprocessor cards (e.g., KNC). Removing the shared memory dstore support helped further improve the behavior - it looks like there is some kind of memory paging issue there that needs further understanding. Given that the shared memory support was about to be lost when I bring over the PMIx integration (until it is restored in that library), it seemed like a reasonable thing to just remove it at this point.
2015-05-29 14:28:26 -07:00
|
|
|
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
|
|
|
|
"usock_peer_try_connect: attempting to connect to server on socket %d",
|
|
|
|
mca_pmix_native_component.sd);
|
|
|
|
/* try to connect */
|
|
|
|
if (connect(mca_pmix_native_component.sd, &mca_pmix_native_component.address, addrlen) < 0) {
|
|
|
|
if (opal_socket_errno == ETIMEDOUT) {
|
|
|
|
/* The server may be too busy to accept new connections */
|
|
|
|
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
|
|
|
|
"timeout connecting to server");
|
|
|
|
CLOSE_THE_SOCKET(mca_pmix_native_component.sd);
|
|
|
|
continue;
|
|
|
|
}
|
Per the PMIx RFC:
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
2014-08-21 18:56:47 +00:00
|
|
|
|
Fat SMPs (i.e., systems with nodes containing large numbers of cpus) were failing to start due to connection failures of the opal/pmix support. Root cause was that (a) we were setting the client socket to non-blocking before calling connect, and (b) the server was using the event library to harvest the accepts, and also did the handshake while in that event. So the server would backup beyond the connection backlog limit, and we would fail.
Changing the client to leave its socket as blocking during the connect doesn't solve the problem by itself - you also have to introduce a sleep delay once the backlog is hit to avoid simply machine-gunning your way thru retries. This gets somewhat difficult to adjust as you don't want to unnecessarily prolong startup time.
We've solved this before by adding a listening thread that simply reaps accepts and shoves them into the event library for subsequent processing. This would resolve the problem, but meant yet another daemon-level thread. So I centralized the listening thread support and let multiple elements register listeners on it. Thus, each daemon now has a single listening thread that reaps accepts from multiple sources - for now, the orte/pmix server and the oob/usock support are using it. I'll add in the oob/tcp component later.
This still didn't fully resolve the SMP problem, especially on coprocessor cards (e.g., KNC). Removing the shared memory dstore support helped further improve the behavior - it looks like there is some kind of memory paging issue there that needs further understanding. Given that the shared memory support was about to be lost when I bring over the PMIx integration (until it is restored in that library), it seemed like a reasonable thing to just remove it at this point.
2015-05-29 14:28:26 -07:00
|
|
|
/* Some kernels (Linux 2.6) will automatically software
|
|
|
|
abort a connection that was ECONNREFUSED on the last
|
|
|
|
attempt, without even trying to establish the
|
|
|
|
connection. Handle that case in a semi-rational
|
|
|
|
way by trying twice before giving up */
|
|
|
|
if (ECONNABORTED == opal_socket_errno) {
|
Per the PMIx RFC:
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
2014-08-21 18:56:47 +00:00
|
|
|
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
|
Fat SMPs (i.e., systems with nodes containing large numbers of cpus) were failing to start due to connection failures of the opal/pmix support. Root cause was that (a) we were setting the client socket to non-blocking before calling connect, and (b) the server was using the event library to harvest the accepts, and also did the handshake while in that event. So the server would backup beyond the connection backlog limit, and we would fail.
Changing the client to leave its socket as blocking during the connect doesn't solve the problem by itself - you also have to introduce a sleep delay once the backlog is hit to avoid simply machine-gunning your way thru retries. This gets somewhat difficult to adjust as you don't want to unnecessarily prolong startup time.
We've solved this before by adding a listening thread that simply reaps accepts and shoves them into the event library for subsequent processing. This would resolve the problem, but meant yet another daemon-level thread. So I centralized the listening thread support and let multiple elements register listeners on it. Thus, each daemon now has a single listening thread that reaps accepts from multiple sources - for now, the orte/pmix server and the oob/usock support are using it. I'll add in the oob/tcp component later.
This still didn't fully resolve the SMP problem, especially on coprocessor cards (e.g., KNC). Removing the shared memory dstore support helped further improve the behavior - it looks like there is some kind of memory paging issue there that needs further understanding. Given that the shared memory support was about to be lost when I bring over the PMIx integration (until it is restored in that library), it seemed like a reasonable thing to just remove it at this point.
2015-05-29 14:28:26 -07:00
|
|
|
"connection to server aborted by OS - retrying");
|
Per the PMIx RFC:
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
2014-08-21 18:56:47 +00:00
|
|
|
CLOSE_THE_SOCKET(mca_pmix_native_component.sd);
|
Fat SMPs (i.e., systems with nodes containing large numbers of cpus) were failing to start due to connection failures of the opal/pmix support. Root cause was that (a) we were setting the client socket to non-blocking before calling connect, and (b) the server was using the event library to harvest the accepts, and also did the handshake while in that event. So the server would backup beyond the connection backlog limit, and we would fail.
Changing the client to leave its socket as blocking during the connect doesn't solve the problem by itself - you also have to introduce a sleep delay once the backlog is hit to avoid simply machine-gunning your way thru retries. This gets somewhat difficult to adjust as you don't want to unnecessarily prolong startup time.
We've solved this before by adding a listening thread that simply reaps accepts and shoves them into the event library for subsequent processing. This would resolve the problem, but meant yet another daemon-level thread. So I centralized the listening thread support and let multiple elements register listeners on it. Thus, each daemon now has a single listening thread that reaps accepts from multiple sources - for now, the orte/pmix server and the oob/usock support are using it. I'll add in the oob/tcp component later.
This still didn't fully resolve the SMP problem, especially on coprocessor cards (e.g., KNC). Removing the shared memory dstore support helped further improve the behavior - it looks like there is some kind of memory paging issue there that needs further understanding. Given that the shared memory support was about to be lost when I bring over the PMIx integration (until it is restored in that library), it seemed like a reasonable thing to just remove it at this point.
2015-05-29 14:28:26 -07:00
|
|
|
continue;
|
Per the PMIx RFC:
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
2014-08-21 18:56:47 +00:00
|
|
|
}
|
|
|
|
}
|
Fat SMPs (i.e., systems with nodes containing large numbers of cpus) were failing to start due to connection failures of the opal/pmix support. Root cause was that (a) we were setting the client socket to non-blocking before calling connect, and (b) the server was using the event library to harvest the accepts, and also did the handshake while in that event. So the server would backup beyond the connection backlog limit, and we would fail.
Changing the client to leave its socket as blocking during the connect doesn't solve the problem by itself - you also have to introduce a sleep delay once the backlog is hit to avoid simply machine-gunning your way thru retries. This gets somewhat difficult to adjust as you don't want to unnecessarily prolong startup time.
We've solved this before by adding a listening thread that simply reaps accepts and shoves them into the event library for subsequent processing. This would resolve the problem, but meant yet another daemon-level thread. So I centralized the listening thread support and let multiple elements register listeners on it. Thus, each daemon now has a single listening thread that reaps accepts from multiple sources - for now, the orte/pmix server and the oob/usock support are using it. I'll add in the oob/tcp component later.
This still didn't fully resolve the SMP problem, especially on coprocessor cards (e.g., KNC). Removing the shared memory dstore support helped further improve the behavior - it looks like there is some kind of memory paging issue there that needs further understanding. Given that the shared memory support was about to be lost when I bring over the PMIx integration (until it is restored in that library), it seemed like a reasonable thing to just remove it at this point.
2015-05-29 14:28:26 -07:00
|
|
|
/* otherwise, the connect succeeded - so break out of the loop */
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (mca_pmix_native_component.retries == mca_pmix_native_component.max_retries ||
|
|
|
|
mca_pmix_native_component.sd < 0){
|
|
|
|
/* We were unsuccessful in establishing this connection, and are
|
|
|
|
* not likely to suddenly become successful */
|
|
|
|
opal_output(0, "pmix:create_socket: connection to server failed");
|
|
|
|
if (0 <= mca_pmix_native_component.sd) {
|
|
|
|
CLOSE_THE_SOCKET(mca_pmix_native_component.sd);
|
|
|
|
}
|
|
|
|
return;
|
Per the PMIx RFC:
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
2014-08-21 18:56:47 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/* connection succeeded */
|
|
|
|
mca_pmix_native_component.retries = 0;
|
|
|
|
|
|
|
|
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
|
|
|
|
"%s sock_peer_try_connect: Connection across to server succeeded",
|
|
|
|
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
|
|
|
|
|
Fat SMPs (i.e., systems with nodes containing large numbers of cpus) were failing to start due to connection failures of the opal/pmix support. Root cause was that (a) we were setting the client socket to non-blocking before calling connect, and (b) the server was using the event library to harvest the accepts, and also did the handshake while in that event. So the server would backup beyond the connection backlog limit, and we would fail.
Changing the client to leave its socket as blocking during the connect doesn't solve the problem by itself - you also have to introduce a sleep delay once the backlog is hit to avoid simply machine-gunning your way thru retries. This gets somewhat difficult to adjust as you don't want to unnecessarily prolong startup time.
We've solved this before by adding a listening thread that simply reaps accepts and shoves them into the event library for subsequent processing. This would resolve the problem, but meant yet another daemon-level thread. So I centralized the listening thread support and let multiple elements register listeners on it. Thus, each daemon now has a single listening thread that reaps accepts from multiple sources - for now, the orte/pmix server and the oob/usock support are using it. I'll add in the oob/tcp component later.
This still didn't fully resolve the SMP problem, especially on coprocessor cards (e.g., KNC). Removing the shared memory dstore support helped further improve the behavior - it looks like there is some kind of memory paging issue there that needs further understanding. Given that the shared memory support was about to be lost when I bring over the PMIx integration (until it is restored in that library), it seemed like a reasonable thing to just remove it at this point.
2015-05-29 14:28:26 -07:00
|
|
|
/* setup event callbacks */
|
|
|
|
opal_event_set(mca_pmix_native_component.evbase,
|
|
|
|
&mca_pmix_native_component.recv_event,
|
|
|
|
mca_pmix_native_component.sd,
|
|
|
|
OPAL_EV_READ|OPAL_EV_PERSIST,
|
|
|
|
pmix_usock_recv_handler, NULL);
|
|
|
|
opal_event_set_priority(&mca_pmix_native_component.recv_event, OPAL_EV_MSG_LO_PRI);
|
|
|
|
mca_pmix_native_component.recv_ev_active = false;
|
|
|
|
|
|
|
|
opal_event_set(mca_pmix_native_component.evbase,
|
|
|
|
&mca_pmix_native_component.send_event,
|
|
|
|
mca_pmix_native_component.sd,
|
|
|
|
OPAL_EV_WRITE|OPAL_EV_PERSIST,
|
|
|
|
pmix_usock_send_handler, NULL);
|
|
|
|
opal_event_set_priority(&mca_pmix_native_component.send_event, OPAL_EV_MSG_LO_PRI);
|
|
|
|
mca_pmix_native_component.send_ev_active = false;
|
|
|
|
|
|
|
|
/* setup the socket as non-blocking */
|
|
|
|
if ((flags = fcntl(mca_pmix_native_component.sd, F_GETFL, 0)) < 0) {
|
|
|
|
opal_output(0, "usock_peer_connect: fcntl(F_GETFL) failed: %s (%d)\n",
|
|
|
|
strerror(opal_socket_errno),
|
|
|
|
opal_socket_errno);
|
|
|
|
} else {
|
|
|
|
flags |= O_NONBLOCK;
|
|
|
|
if (fcntl(mca_pmix_native_component.sd, F_SETFL, flags) < 0)
|
|
|
|
opal_output(0, "usock_peer_connect: fcntl(F_SETFL) failed: %s (%d)\n",
|
|
|
|
strerror(opal_socket_errno),
|
|
|
|
opal_socket_errno);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* setup our recv to catch the return ack call */
|
Per the PMIx RFC:
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
2014-08-21 18:56:47 +00:00
|
|
|
if (!mca_pmix_native_component.recv_ev_active) {
|
|
|
|
opal_event_add(&mca_pmix_native_component.recv_event, 0);
|
|
|
|
mca_pmix_native_component.recv_ev_active = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* send our globally unique process identifier to the server */
|
|
|
|
if (OPAL_SUCCESS == (rc = usock_send_connect_ack())) {
|
|
|
|
mca_pmix_native_component.state = PMIX_USOCK_CONNECT_ACK;
|
|
|
|
} else {
|
|
|
|
opal_output(0,
|
|
|
|
"%s usock_peer_try_connect: "
|
|
|
|
"usock_send_connect_ack to server failed: %s (%d)",
|
|
|
|
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
|
|
|
|
opal_strerror(rc), rc);
|
|
|
|
mca_pmix_native_component.state = PMIX_USOCK_FAILED;
|
|
|
|
CLOSE_THE_SOCKET(mca_pmix_native_component.sd);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
int usock_send_connect_ack(void)
|
|
|
|
{
|
|
|
|
char *msg;
|
|
|
|
pmix_usock_hdr_t hdr;
|
|
|
|
int rc;
|
|
|
|
size_t sdsize;
|
2015-03-28 20:34:26 -07:00
|
|
|
char *cred;
|
|
|
|
size_t credsize;
|
|
|
|
|
Per the PMIx RFC:
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
2014-08-21 18:56:47 +00:00
|
|
|
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
|
|
|
|
"%s SEND CONNECT ACK",
|
|
|
|
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
|
|
|
|
|
|
|
|
/* setup the header */
|
|
|
|
hdr.id = OPAL_PROC_MY_NAME;
|
|
|
|
hdr.tag = UINT32_MAX;
|
|
|
|
hdr.type = PMIX_USOCK_IDENT;
|
|
|
|
|
|
|
|
/* get our security credential */
|
2015-03-28 20:34:26 -07:00
|
|
|
if (OPAL_SUCCESS != (rc = opal_sec.get_my_credential(NULL, opal_dstore_internal, &OPAL_PROC_MY_NAME, &cred, &credsize))) {
|
Per the PMIx RFC:
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
2014-08-21 18:56:47 +00:00
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* set the number of bytes to be read beyond the header */
|
2015-03-28 20:34:26 -07:00
|
|
|
hdr.nbytes = strlen(opal_version_string) + 1 + credsize;
|
Per the PMIx RFC:
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
2014-08-21 18:56:47 +00:00
|
|
|
|
|
|
|
/* create a space for our message */
|
2015-03-28 20:34:26 -07:00
|
|
|
sdsize = (sizeof(hdr) + strlen(opal_version_string) + 1 + credsize);
|
Per the PMIx RFC:
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
2014-08-21 18:56:47 +00:00
|
|
|
if (NULL == (msg = (char*)malloc(sdsize))) {
|
|
|
|
return OPAL_ERR_OUT_OF_RESOURCE;
|
|
|
|
}
|
|
|
|
memset(msg, 0, sdsize);
|
|
|
|
|
|
|
|
/* load the message */
|
|
|
|
memcpy(msg, &hdr, sizeof(hdr));
|
|
|
|
memcpy(msg+sizeof(hdr), opal_version_string, strlen(opal_version_string));
|
2015-03-28 20:34:26 -07:00
|
|
|
memcpy(msg+sizeof(hdr)+strlen(opal_version_string)+1, cred, credsize);
|
|
|
|
if (NULL != cred) {
|
|
|
|
free(cred);
|
|
|
|
}
|
Per the PMIx RFC:
WHAT: Merge the PMIx branch into the devel repo, creating a new
OPAL “lmix” framework to abstract PMI support for all RTEs.
Replace the ORTE daemon-level collectives with a new PMIx
server and update the ORTE grpcomm framework to support
server-to-server collectives
WHY: We’ve had problems dealing with variations in PMI implementations,
and need to extend the existing PMI definitions to meet exascale
requirements.
WHEN: Mon, Aug 25
WHERE: https://github.com/rhc54/ompi-svn-mirror.git
Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding.
All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level.
Accordingly, we have:
* created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations.
* Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported.
* Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint
* removed the prior OMPI/OPAL modex code
* added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform.
* retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand
This commit was SVN r32570.
2014-08-21 18:56:47 +00:00
|
|
|
|
|
|
|
if (OPAL_SUCCESS != usock_send_blocking(msg, sdsize)) {
|
|
|
|
free(msg);
|
|
|
|
return OPAL_ERR_UNREACH;
|
|
|
|
}
|
|
|
|
free(msg);
|
|
|
|
return OPAL_SUCCESS;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* A blocking send on a non-blocking socket. Used to send the small amount of connection
|
|
|
|
* information that identifies the peers endpoint.
|
|
|
|
*/
|
|
|
|
static int usock_send_blocking(char *ptr, size_t size)
|
|
|
|
{
|
|
|
|
size_t cnt = 0;
|
|
|
|
int retval;
|
|
|
|
|
|
|
|
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
|
|
|
|
"%s send blocking of %"PRIsize_t" bytes to socket %d",
|
|
|
|
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
|
|
|
|
size, mca_pmix_native_component.sd);
|
|
|
|
|
|
|
|
while (cnt < size) {
|
|
|
|
retval = send(mca_pmix_native_component.sd, (char*)ptr+cnt, size-cnt, 0);
|
|
|
|
if (retval < 0) {
|
|
|
|
if (opal_socket_errno != EINTR && opal_socket_errno != EAGAIN && opal_socket_errno != EWOULDBLOCK) {
|
|
|
|
opal_output(0, "%s usock_peer_send_blocking: send() to socket %d failed: %s (%d)\n",
|
|
|
|
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
|
|
|
|
mca_pmix_native_component.sd,
|
|
|
|
strerror(opal_socket_errno),
|
|
|
|
opal_socket_errno);
|
|
|
|
mca_pmix_native_component.state = PMIX_USOCK_FAILED;
|
|
|
|
CLOSE_THE_SOCKET(mca_pmix_native_component.sd);
|
|
|
|
return OPAL_ERR_UNREACH;
|
|
|
|
}
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
cnt += retval;
|
|
|
|
}
|
|
|
|
|
|
|
|
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
|
|
|
|
"%s blocking send complete to socket %d",
|
|
|
|
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
|
|
|
|
mca_pmix_native_component.sd);
|
|
|
|
|
|
|
|
return OPAL_SUCCESS;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Routine for debugging to print the connection state and socket options
|
|
|
|
*/
|
|
|
|
void pmix_usock_dump(const char* msg)
|
|
|
|
{
|
|
|
|
char buff[255];
|
|
|
|
int nodelay,flags;
|
|
|
|
|
|
|
|
if ((flags = fcntl(mca_pmix_native_component.sd, F_GETFL, 0)) < 0) {
|
|
|
|
opal_output(0, "%s usock_peer_dump: fcntl(F_GETFL) failed: %s (%d)\n",
|
|
|
|
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
|
|
|
|
strerror(opal_socket_errno),
|
|
|
|
opal_socket_errno);
|
|
|
|
}
|
|
|
|
|
|
|
|
#if defined(USOCK_NODELAY)
|
|
|
|
optlen = sizeof(nodelay);
|
|
|
|
if (getsockopt(mca_pmix_native_component.sd, IPPROTO_USOCK, USOCK_NODELAY, (char *)&nodelay, &optlen) < 0) {
|
|
|
|
opal_output(0, "%s usock_peer_dump: USOCK_NODELAY option: %s (%d)\n",
|
|
|
|
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
|
|
|
|
strerror(opal_socket_errno),
|
|
|
|
opal_socket_errno);
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
nodelay = 0;
|
|
|
|
#endif
|
|
|
|
|
|
|
|
snprintf(buff, sizeof(buff), "%s %s: nodelay %d flags %08x\n",
|
|
|
|
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
|
|
|
|
msg, nodelay, flags);
|
|
|
|
opal_output(0, "%s", buff);
|
|
|
|
}
|
|
|
|
|
|
|
|
char* pmix_usock_state_print(pmix_usock_state_t state)
|
|
|
|
{
|
|
|
|
switch (state) {
|
|
|
|
case PMIX_USOCK_UNCONNECTED:
|
|
|
|
return "UNCONNECTED";
|
|
|
|
case PMIX_USOCK_CLOSED:
|
|
|
|
return "CLOSED";
|
|
|
|
case PMIX_USOCK_RESOLVE:
|
|
|
|
return "RESOLVE";
|
|
|
|
case PMIX_USOCK_CONNECTING:
|
|
|
|
return "CONNECTING";
|
|
|
|
case PMIX_USOCK_CONNECT_ACK:
|
|
|
|
return "ACK";
|
|
|
|
case PMIX_USOCK_CONNECTED:
|
|
|
|
return "CONNECTED";
|
|
|
|
case PMIX_USOCK_FAILED:
|
|
|
|
return "FAILED";
|
|
|
|
default:
|
|
|
|
return "UNKNOWN";
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|