ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
openmpi/orte/mca/grpcomm/basic/grpcomm_basic_module.c
Ralph Castain 54b2cf747e These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC. 
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.

This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:

As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.

In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.

The incoming changes revamp these procedures in three ways:

1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.

The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.

Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.


2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.

The size of this data has been reduced in three ways:

(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.

To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.

(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.

(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.

While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.


3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.

It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.

Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.


There are a few minor additional changes in the commit that I'll just note in passing:

* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.

* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.

* cleanup of some stale header files

This commit was SVN r16364.
2007-10-05 19:48:23 +00:00

817 строки
30 KiB
C
Исходник Ответственный История

/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2005 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) 2007 Sun Microsystems, Inc. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "orte_config.h"
#include "orte/orte_constants.h"
#include <string.h>
#ifdef HAVE_SYS_TIME_H
#include <sys/time.h>
#endif /* HAVE_SYS_TIME_H */
#include "opal/threads/condition.h"
#include "opal/util/output.h"
#include "opal/util/bit_ops.h"
#include "orte/util/proc_info.h"
#include "orte/dss/dss.h"
#include "orte/mca/gpr/gpr.h"
#include "orte/mca/errmgr/errmgr.h"
#include "orte/mca/ns/ns.h"
#include "orte/mca/rmgr/rmgr.h"
#include "orte/mca/smr/smr.h"
#include "orte/mca/odls/odls_types.h"
#include "orte/mca/rml/rml.h"
#include "orte/runtime/params.h"
#include "orte/mca/grpcomm/base/base.h"
#include "grpcomm_basic.h"
/* Local functions */
static int xcast_binomial_tree(orte_jobid_t job,
orte_buffer_t *buffer,
orte_rml_tag_t tag);
static int xcast_linear(orte_jobid_t job,
orte_buffer_t *buffer,
orte_rml_tag_t tag);
static int xcast_direct(orte_jobid_t job,
orte_buffer_t *buffer,
orte_rml_tag_t tag);
/* define a callback function for use by the blocking version
* of xcast so we can "hold" the caller here until all non-blocking
* sends have completed
*/
static void xcast_send_cb(int status,
orte_process_name_t* peer,
orte_buffer_t* buffer,
orte_rml_tag_t tag,
void* cbdata)
{
OPAL_THREAD_LOCK(&orte_grpcomm_basic.mutex);
orte_grpcomm_basic.num_active--;
if (orte_grpcomm_basic.num_active <= 0) {
orte_grpcomm_basic.num_active = 0; /* just to be safe */
opal_condition_signal(&orte_grpcomm_basic.cond);
}
OPAL_THREAD_UNLOCK(&orte_grpcomm_basic.mutex);
return;
}
/**
* A "broadcast-like" function to a job's processes.
* @param jobid The job whose processes are to receive the message
* @param buffer The data to broadcast
*/
/* Non-blocking version */
static int xcast_nb(orte_jobid_t job,
orte_buffer_t *buffer,
orte_rml_tag_t tag)
{
int rc = ORTE_SUCCESS;
struct timeval start, stop;
orte_vpid_t num_daemons;
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
"%s xcast_nb sent to job %ld tag %ld",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(long)job, (long)tag));
/* if there is no message to send, then just return ok */
if (NULL == buffer) {
return ORTE_SUCCESS;
}
if (orte_timing) {
gettimeofday(&start, NULL);
}
/* get the number of daemons currently in the system so we can
* select the "optimal" algorithm
*/
if (ORTE_SUCCESS != (rc = orte_ns.get_vpid_range(0, &num_daemons))) {
ORTE_ERROR_LOG(rc);
return rc;
}
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
"%s xcast_nb: num_daemons %ld linear xover: %ld binomial xover: %ld",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(long)num_daemons, (long)orte_grpcomm_basic.xcast_linear_xover,
(long)orte_grpcomm_basic.xcast_binomial_xover));
if (num_daemons < 2) {
/* if there is only one daemon in the system, then we must
* use the direct mode - there is no other option. Note that
* since the HNP is the one that typically does xcast sends,
* only one daemon means that the HNP itself is sending to
* itself. This is required in singletons - where the
* singleton acts as the HNP - and as an HNP starts
* itself up
*
* NOTE: although we allow users to alter crossover points
* for selecting specific xcast modes, this required
* use-case behavior MUST always be retained or else
* singletons and HNP startup will fail!
*/
rc = xcast_direct(job, buffer, tag);
goto DONE;
}
/* now use the crossover points to select the proper transmission
* mode. We have built-in default crossover points for this
* decision tree, but the user is free to alter them as
* they wish via MCA params
*/
if (num_daemons < orte_grpcomm_basic.xcast_linear_xover) {
rc = xcast_direct(job, buffer, tag);
} else if (num_daemons < orte_grpcomm_basic.xcast_binomial_xover) {
rc = xcast_linear(job, buffer, tag);
} else {
rc = xcast_binomial_tree(job, buffer, tag);
}
DONE:
if (orte_timing) {
gettimeofday(&stop, NULL);
opal_output(0, "xcast_nb %s: time %ld usec", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(long int)((stop.tv_sec - start.tv_sec)*1000000 +
(stop.tv_usec - start.tv_usec)));
}
return rc;
}
/* Blocking version */
static int xcast(orte_jobid_t job,
orte_buffer_t *buffer,
orte_rml_tag_t tag)
{
int rc = ORTE_SUCCESS;
struct timeval start, stop;
orte_vpid_t num_daemons;
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
"%s xcast sent to job %ld tag %ld",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(long)job, (long)tag));
/* if there is no message to send, then just return ok */
if (NULL == buffer) {
return ORTE_SUCCESS;
}
if (orte_timing) {
gettimeofday(&start, NULL);
}
/* get the number of daemons currently in the system so we can
* select the "optimal" algorithm
*/
if (ORTE_SUCCESS != (rc = orte_ns.get_vpid_range(0, &num_daemons))) {
ORTE_ERROR_LOG(rc);
return rc;
}
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
"%s xcast: num_daemons %ld linear xover: %ld binomial xover: %ld",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(long)num_daemons, (long)orte_grpcomm_basic.xcast_linear_xover,
(long)orte_grpcomm_basic.xcast_binomial_xover));
if (num_daemons < 2) {
/* if there is only one daemon in the system, then we must
* use the direct mode - there is no other option. Note that
* since the HNP is the one that typically does xcast sends,
* only one daemon means that the HNP itself is sending to
* itself. This is required in singletons - where the
* singleton acts as the HNP - and as an HNP starts
* itself up
*
* NOTE: although we allow users to alter crossover points
* for selecting specific xcast modes, this required
* use-case behavior MUST always be retained or else
* singletons and HNP startup will fail!
*/
rc = xcast_direct(job, buffer, tag);
goto DONE;
}
/* now use the crossover points to select the proper transmission
* mode. We have built-in default crossover points for this
* decision tree, but the user is free to alter them as
* they wish via MCA params
*/
if (num_daemons < orte_grpcomm_basic.xcast_linear_xover) {
rc = xcast_direct(job, buffer, tag);
} else if (num_daemons < orte_grpcomm_basic.xcast_binomial_xover) {
rc = xcast_linear(job, buffer, tag);
} else {
rc = xcast_binomial_tree(job, buffer, tag);
}
DONE:
/* now go to sleep until woken up */
OPAL_THREAD_LOCK(&orte_grpcomm_basic.mutex);
if (orte_grpcomm_basic.num_active > 0) {
opal_condition_wait(&orte_grpcomm_basic.cond, &orte_grpcomm_basic.mutex);
}
OPAL_THREAD_UNLOCK(&orte_grpcomm_basic.mutex);
if (orte_timing) {
gettimeofday(&stop, NULL);
opal_output(0, "xcast %s: time %ld usec", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(long int)((stop.tv_sec - start.tv_sec)*1000000 +
(stop.tv_usec - start.tv_usec)));
}
return rc;
}
static int xcast_binomial_tree(orte_jobid_t job,
orte_buffer_t *buffer,
orte_rml_tag_t tag)
{
orte_daemon_cmd_flag_t command, mode;
int rc;
orte_process_name_t target;
orte_buffer_t *buf;
orte_vpid_t nd;
orte_std_cntr_t num_daemons;
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
"%s xcast_binomial",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* this is the HNP end, so it starts the procedure. Since the HNP is always the
* vpid=0 at this time, we take advantage of that fact to figure out who we
* should send this to on the first step
*/
/* need to pack the msg for sending - be sure to include routing info so it
* can properly be sent through the daemons
*/
buf = OBJ_NEW(orte_buffer_t);
/* tell the daemon the routing algorithm so it can figure
* out how to forward the message down the tree, if at all
*/
mode = ORTE_DAEMON_ROUTE_BINOMIAL;
if (ORTE_SUCCESS != (rc = orte_dss.pack(buf, &mode, 1, ORTE_DAEMON_CMD))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* get the number of daemons currently in the system and tell the daemon so
* it can properly route
*/
if (ORTE_SUCCESS != (rc = orte_ns.get_vpid_range(0, &nd))) {
ORTE_ERROR_LOG(rc);
goto CLEANUP;
}
num_daemons = (orte_std_cntr_t)nd;
if (ORTE_SUCCESS != (rc = orte_dss.pack(buf, &num_daemons, 1, ORTE_STD_CNTR))) {
ORTE_ERROR_LOG(rc);
goto CLEANUP;
}
/* if this isn't intended for the daemon command tag, then we better
* tell the daemon to deliver it to the procs, and what job is supposed
* to get it - this occurs when a caller just wants to send something
* to all the procs in a job. In that use-case, the caller doesn't know
* anything about inserting daemon commands or what routing algo might
* be used, so we have to help them out a little. Functions that are
* sending commands to the daemons themselves are smart enough to know
* what they need to do.
*/
if (ORTE_RML_TAG_DAEMON != tag) {
command = ORTE_DAEMON_MESSAGE_LOCAL_PROCS;
if (ORTE_SUCCESS != (rc = orte_dss.pack(buf, &command, 1, ORTE_DAEMON_CMD))) {
ORTE_ERROR_LOG(rc);
goto CLEANUP;
}
if (ORTE_SUCCESS != (rc = orte_dss.pack(buf, &job, 1, ORTE_JOBID))) {
ORTE_ERROR_LOG(rc);
goto CLEANUP;
}
if (ORTE_SUCCESS != (rc = orte_dss.pack(buf, &tag, 1, ORTE_RML_TAG))) {
ORTE_ERROR_LOG(rc);
goto CLEANUP;
}
}
/* copy the payload into the new buffer - this is non-destructive, so our
* caller is still responsible for releasing any memory in the buffer they
* gave to us
*/
if (ORTE_SUCCESS != (rc = orte_dss.copy_payload(buf, buffer))) {
ORTE_ERROR_LOG(rc);
goto CLEANUP;
}
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
"%s xcast_binomial: buffer size %ld",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(long)buf->bytes_used));
/* all we need to do is send this to ourselves - our relay logic
* will ensure everyone else gets it!
*/
target.jobid = 0;
target.vpid = 0;
++orte_grpcomm_basic.num_active;
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
"xcast_binomial: num_active now %ld sending %s => %s",
(long)orte_grpcomm_basic.num_active,
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&target)));
if (0 > (rc = orte_rml.send_buffer_nb(&target, buf, ORTE_RML_TAG_ORTED_ROUTED,
0, xcast_send_cb, NULL))) {
ORTE_ERROR_LOG(ORTE_ERR_COMM_FAILURE);
rc = ORTE_ERR_COMM_FAILURE;
OPAL_THREAD_LOCK(&orte_grpcomm_basic.mutex);
--orte_grpcomm_basic.num_active;
OPAL_THREAD_UNLOCK(&orte_grpcomm_basic.mutex);
goto CLEANUP;
}
CLEANUP:
OBJ_RELEASE(buf);
return rc;
}
static int xcast_linear(orte_jobid_t job,
orte_buffer_t *buffer,
orte_rml_tag_t tag)
{
int rc;
orte_buffer_t *buf;
orte_daemon_cmd_flag_t command, mode=ORTE_DAEMON_ROUTE_NONE;
orte_vpid_t i, range;
orte_process_name_t dummy;
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
"%s xcast_linear",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* since we have to pack some additional info into the buffer to be
* sent to the daemons, we create a new buffer into which we will
* put the intermediate payload - i.e., the info that goes to the
* daemon. This buffer will contain all the info needed by the
* daemon, plus the payload intended for the processes themselves
*/
buf = OBJ_NEW(orte_buffer_t);
/* tell the daemon that no further routing required */
if (ORTE_SUCCESS != (rc = orte_dss.pack(buf, &mode, 1, ORTE_DAEMON_CMD))) {
ORTE_ERROR_LOG(rc);
goto CLEANUP;
}
/* if this isn't intended for the daemon command tag, then we better
* tell the daemon to deliver it to the procs, and what job is supposed
* to get it - this occurs when a caller just wants to send something
* to all the procs in a job. In that use-case, the caller doesn't know
* anything about inserting daemon commands or what routing algo might
* be used, so we have to help them out a little. Functions that are
* sending commands to the daemons themselves are smart enough to know
* what they need to do.
*/
if (ORTE_RML_TAG_DAEMON != tag) {
command = ORTE_DAEMON_MESSAGE_LOCAL_PROCS;
if (ORTE_SUCCESS != (rc = orte_dss.pack(buf, &command, 1, ORTE_DAEMON_CMD))) {
ORTE_ERROR_LOG(rc);
goto CLEANUP;
}
if (ORTE_SUCCESS != (rc = orte_dss.pack(buf, &job, 1, ORTE_JOBID))) {
ORTE_ERROR_LOG(rc);
goto CLEANUP;
}
if (ORTE_SUCCESS != (rc = orte_dss.pack(buf, &tag, 1, ORTE_RML_TAG))) {
ORTE_ERROR_LOG(rc);
goto CLEANUP;
}
}
/* copy the payload into the new buffer - this is non-destructive, so our
* caller is still responsible for releasing any memory in the buffer they
* gave to us
*/
if (ORTE_SUCCESS != (rc = orte_dss.copy_payload(buf, buffer))) {
ORTE_ERROR_LOG(rc);
goto CLEANUP;
}
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
"%s xcast_linear: buffer size %ld",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(long)buf->bytes_used));
/* get the number of daemons out there */
orte_ns.get_vpid_range(0, &range);
/* we have to account for all of the messages we are about to send
* because the non-blocking send can come back almost immediately - before
* we would get the chance to increment the num_active. This causes us
* to not correctly wakeup and reset the xcast_in_progress flag
*/
OPAL_THREAD_LOCK(&orte_grpcomm_basic.mutex);
orte_grpcomm_basic.num_active += range;
OPAL_THREAD_UNLOCK(&orte_grpcomm_basic.mutex);
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
"%s xcast_linear: num_active now %ld",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(long)orte_grpcomm_basic.num_active));
/* send the message to each daemon as fast as we can */
dummy.jobid = 0;
for (i=0; i < range; i++) {
dummy.vpid = i;
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
"xcast_linear: %s => %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&dummy)));
if (0 > (rc = orte_rml.send_buffer_nb(&dummy, buf, ORTE_RML_TAG_ORTED_ROUTED,
0, xcast_send_cb, NULL))) {
if (ORTE_ERR_ADDRESSEE_UNKNOWN != rc) {
ORTE_ERROR_LOG(ORTE_ERR_COMM_FAILURE);
rc = ORTE_ERR_COMM_FAILURE;
OPAL_THREAD_LOCK(&orte_grpcomm_basic.mutex);
orte_grpcomm_basic.num_active -= (range-i);
OPAL_THREAD_UNLOCK(&orte_grpcomm_basic.mutex);
goto CLEANUP;
}
/* decrement the number we are waiting to see */
OPAL_THREAD_LOCK(&orte_grpcomm_basic.mutex);
orte_grpcomm_basic.num_active--;
OPAL_THREAD_UNLOCK(&orte_grpcomm_basic.mutex);
}
}
rc = ORTE_SUCCESS;
/* cleanup */
CLEANUP:
OBJ_RELEASE(buf);
return rc;
}
static int xcast_direct(orte_jobid_t job,
orte_buffer_t *buffer,
orte_rml_tag_t tag)
{
orte_std_cntr_t i;
int rc;
orte_process_name_t *peers=NULL;
orte_std_cntr_t n;
opal_list_t attrs;
opal_list_item_t *item;
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
"%s xcast_direct",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* need to get the job peers so we know who to send the message to */
OBJ_CONSTRUCT(&attrs, opal_list_t);
orte_rmgr.add_attribute(&attrs, ORTE_NS_USE_JOBID, ORTE_JOBID, &job, ORTE_RMGR_ATTR_OVERRIDE);
if (ORTE_SUCCESS != (rc = orte_ns.get_peers(&peers, &n, &attrs))) {
ORTE_ERROR_LOG(rc);
OBJ_DESTRUCT(&attrs);
return rc;
}
item = opal_list_remove_first(&attrs);
OBJ_RELEASE(item);
OBJ_DESTRUCT(&attrs);
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
"%s xcast_direct: buffer size %ld",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(long)buffer->bytes_used));
/* we have to account for all of the messages we are about to send
* because the non-blocking send can come back almost immediately - before
* we would get the chance to increment the num_active. This causes us
* to not correctly wakeup and reset the xcast_in_progress flag
*/
OPAL_THREAD_LOCK(&orte_grpcomm_basic.mutex);
orte_grpcomm_basic.num_active += n;
OPAL_THREAD_UNLOCK(&orte_grpcomm_basic.mutex);
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
"%s xcast_direct: num_active now %ld",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(long)orte_grpcomm_basic.num_active));
for(i=0; i<n; i++) {
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
"xcast_direct: %s => %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(peers+i)));
if (0 > (rc = orte_rml.send_buffer_nb(peers+i, buffer, tag, 0, xcast_send_cb, NULL))) {
if (ORTE_ERR_ADDRESSEE_UNKNOWN != rc) {
ORTE_ERROR_LOG(ORTE_ERR_COMM_FAILURE);
rc = ORTE_ERR_COMM_FAILURE;
OPAL_THREAD_LOCK(&orte_grpcomm_basic.mutex);
orte_grpcomm_basic.num_active -= (n-i);
OPAL_THREAD_UNLOCK(&orte_grpcomm_basic.mutex);
goto CLEANUP;
}
/* decrement the number we are waiting to see */
OPAL_THREAD_LOCK(&orte_grpcomm_basic.mutex);
orte_grpcomm_basic.num_active--;
OPAL_THREAD_UNLOCK(&orte_grpcomm_basic.mutex);
}
}
rc = ORTE_SUCCESS;
CLEANUP:
free(peers);
return rc;
}
static int allgather(orte_buffer_t *sbuf, orte_buffer_t *rbuf)
{
orte_process_name_t name;
int rc;
orte_std_cntr_t i;
orte_buffer_t tmpbuf;
/* everything happens within my jobid */
name.jobid = ORTE_PROC_MY_NAME->jobid;
if (0 != ORTE_PROC_MY_NAME->vpid) {
/* everyone but rank=0 sends data */
name.vpid = 0;
if (0 > orte_rml.send_buffer(&name, sbuf, ORTE_RML_TAG_ALLGATHER, 0)) {
ORTE_ERROR_LOG(ORTE_ERR_COMM_FAILURE);
return ORTE_ERR_COMM_FAILURE;
}
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
"%s allgather buffer sent",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* now receive the final result from rank=0 */
if (0 > orte_rml.recv_buffer(ORTE_NAME_WILDCARD, rbuf, ORTE_RML_TAG_ALLGATHER, 0)) {
ORTE_ERROR_LOG(ORTE_ERR_COMM_FAILURE);
return ORTE_ERR_COMM_FAILURE;
}
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
"%s allgather buffer received",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
return ORTE_SUCCESS;
}
/* seed the outgoing buffer with the num_procs so it can be unpacked */
if (ORTE_SUCCESS != (rc = orte_dss.pack(rbuf, &orte_process_info.num_procs, 1, ORTE_STD_CNTR))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* put my own information into the outgoing buffer */
if (ORTE_SUCCESS != (rc = orte_dss.copy_payload(rbuf, sbuf))) {
ORTE_ERROR_LOG(rc);
return rc;
}
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
"%s allgather collecting buffers",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* rank=0 receives everyone else's data */
for (i=1; i < orte_process_info.num_procs; i++) {
name.vpid = (orte_vpid_t)i;
OBJ_CONSTRUCT(&tmpbuf, orte_buffer_t);
if (0 > orte_rml.recv_buffer(&name, &tmpbuf, ORTE_RML_TAG_ALLGATHER, 0)) {
ORTE_ERROR_LOG(ORTE_ERR_COMM_FAILURE);
return ORTE_ERR_COMM_FAILURE;
}
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
"%s allgather buffer %ld received",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), (long)i));
/* append this data to the rbuf */
if (ORTE_SUCCESS != (rc = orte_dss.copy_payload(rbuf, &tmpbuf))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* clear out the tmpbuf */
OBJ_DESTRUCT(&tmpbuf);
}
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
"%s allgather xcasting collected data",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* xcast the results */
orte_grpcomm.xcast(ORTE_PROC_MY_NAME->jobid, rbuf, ORTE_RML_TAG_ALLGATHER);
/* xcast automatically ensures that the sender -always- gets a copy
* of the message. This is required to ensure proper operation of the
* launch system as the HNP -must- get a copy itself. So we have to
* post our own receive here so that we don't leave a message rattling
* around in our RML
*/
OBJ_CONSTRUCT(&tmpbuf, orte_buffer_t);
if (0 > (rc = orte_rml.recv_buffer(ORTE_NAME_WILDCARD, &tmpbuf, ORTE_RML_TAG_ALLGATHER, 0))) {
ORTE_ERROR_LOG(rc);
return rc;
}
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
"%s allgather buffer received",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* don't need the received buffer - we already have what we need */
OBJ_DESTRUCT(&tmpbuf);
return ORTE_SUCCESS;
}
static int allgather_list(opal_list_t *names, orte_buffer_t *sbuf, orte_buffer_t *rbuf)
{
opal_list_item_t *item;
orte_namelist_t *peer, *root;
orte_std_cntr_t i, num_peers;
orte_buffer_t tmpbuf;
int rc;
/* the first entry on the list is the "root" that collects
* all the data - everyone else just sends and gets back
* the results
*/
root = (orte_namelist_t*)opal_list_get_first(names);
if (ORTE_EQUAL != orte_dss.compare(root->name, ORTE_PROC_MY_NAME, ORTE_NAME)) {
/* everyone but root sends data */
if (0 > orte_rml.send_buffer(root->name, sbuf, ORTE_RML_TAG_ALLGATHER_LIST, 0)) {
ORTE_ERROR_LOG(ORTE_ERR_COMM_FAILURE);
return ORTE_ERR_COMM_FAILURE;
}
/* now receive the final result */
if (0 > orte_rml.recv_buffer(root->name, rbuf, ORTE_RML_TAG_ALLGATHER_LIST, 0)) {
ORTE_ERROR_LOG(ORTE_ERR_COMM_FAILURE);
return ORTE_ERR_COMM_FAILURE;
}
return ORTE_SUCCESS;
}
/* count how many peers are participating, including myself */
num_peers = (orte_std_cntr_t)opal_list_get_size(names);
/* seed the outgoing buffer with the num_procs so it can be unpacked */
if (ORTE_SUCCESS != (rc = orte_dss.pack(rbuf, &num_peers, 1, ORTE_STD_CNTR))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* put my own information into the outgoing buffer */
if (ORTE_SUCCESS != (rc = orte_dss.copy_payload(rbuf, sbuf))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* root receives everyone else's data */
for (i=1; i < num_peers; i++) {
/* receive the buffer from this process */
OBJ_CONSTRUCT(&tmpbuf, orte_buffer_t);
if (0 > orte_rml.recv_buffer(ORTE_NAME_WILDCARD, &tmpbuf, ORTE_RML_TAG_ALLGATHER_LIST, 0)) {
ORTE_ERROR_LOG(ORTE_ERR_COMM_FAILURE);
return ORTE_ERR_COMM_FAILURE;
}
/* append this data to the rbuf */
if (ORTE_SUCCESS != (rc = orte_dss.copy_payload(rbuf, &tmpbuf))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* clear out the tmpbuf */
OBJ_DESTRUCT(&tmpbuf);
}
/* broadcast the results */
for (item = opal_list_get_first(names);
item != opal_list_get_end(names);
item = opal_list_get_next(item)) {
peer = (orte_namelist_t*)item;
/* skip myself */
if (ORTE_EQUAL == orte_dss.compare(root->name, peer->name, ORTE_NAME)) {
continue;
}
/* transmit the buffer to this process */
if (0 > orte_rml.send_buffer(peer->name, rbuf, ORTE_RML_TAG_ALLGATHER_LIST, 0)) {
ORTE_ERROR_LOG(ORTE_ERR_COMM_FAILURE);
return ORTE_ERR_COMM_FAILURE;
}
}
return ORTE_SUCCESS;
}
static int barrier(void)
{
orte_process_name_t name;
orte_std_cntr_t i;
orte_buffer_t buf;
int rc;
/* everything happens within the same jobid */
name.jobid = ORTE_PROC_MY_NAME->jobid;
/* All non-root send & receive zero-length message. */
if (0 != ORTE_PROC_MY_NAME->vpid) {
name.vpid = 0;
OBJ_CONSTRUCT(&buf, orte_buffer_t);
i=0;
orte_dss.pack(&buf, &i, 1, ORTE_STD_CNTR); /* put something meaningless here */
rc = orte_rml.send_buffer(&name,&buf,ORTE_RML_TAG_BARRIER,0);
if (rc < 0) {
ORTE_ERROR_LOG(rc);
return rc;
}
OBJ_DESTRUCT(&buf);
/* get the release from rank=0 */
OBJ_CONSTRUCT(&buf, orte_buffer_t);
rc = orte_rml.recv_buffer(ORTE_NAME_WILDCARD,&buf,ORTE_RML_TAG_BARRIER,0);
if (rc < 0) {
ORTE_ERROR_LOG(rc);
return rc;
}
OBJ_DESTRUCT(&buf);
return ORTE_SUCCESS;
}
for (i = 1; i < orte_process_info.num_procs; i++) {
name.vpid = (orte_vpid_t)i;
OBJ_CONSTRUCT(&buf, orte_buffer_t);
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
"%s barrier %ld received",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), (long)i));
rc = orte_rml.recv_buffer(&name,&buf,ORTE_RML_TAG_BARRIER,0);
if (rc < 0) {
ORTE_ERROR_LOG(rc);
return rc;
}
OBJ_DESTRUCT(&buf);
}
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
"%s barrier xcasting release",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* xcast the release */
OBJ_CONSTRUCT(&buf, orte_buffer_t);
orte_dss.pack(&buf, &i, 1, ORTE_STD_CNTR); /* put something meaningless here */
orte_grpcomm.xcast(ORTE_PROC_MY_NAME->jobid, &buf, ORTE_RML_TAG_BARRIER);
OBJ_DESTRUCT(&buf);
/* xcast automatically ensures that the sender -always- gets a copy
* of the message. This is required to ensure proper operation of the
* launch system as the HNP -must- get a copy itself. So we have to
* post our own receive here so that we don't leave a message rattling
* around in our RML
*/
OBJ_CONSTRUCT(&buf, orte_buffer_t);
if (0 > (rc = orte_rml.recv_buffer(ORTE_NAME_WILDCARD, &buf, ORTE_RML_TAG_BARRIER, 0))) {
ORTE_ERROR_LOG(rc);
return rc;
}
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
"%s barrier release received",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
OBJ_DESTRUCT(&buf);
return ORTE_SUCCESS;
}
orte_grpcomm_base_module_t orte_grpcomm_basic_module = {
xcast,
xcast_nb,
allgather,
allgather_list,
barrier
};
Показать git blame Копировать постоянную ссылку