1
1
openmpi/orte/runtime/orte_cr.c

638 строки
17 KiB
C
Исходник Обычный вид История

/*
* Copyright (c) 2004-2007 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 Los Alamos National Security, LLC. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/** @file
*
* ORTE Layer Checkpoint/Restart Runtime functions
*
*/
#include "orte_config.h"
#include <errno.h>
#ifdef HAVE_UNISTD_H
#include <unistd.h>
#endif /* HAVE_UNISTD_H */
#ifdef HAVE_FCNTL_H
#include <fcntl.h>
#endif /* HAVE_FCNTL_H */
#ifdef HAVE_SYS_TYPES_H
#include <sys/types.h>
#endif /* HAVE_SYS_TYPES_H */
#ifdef HAVE_SYS_STAT_H
#include <sys/stat.h> /* for mkfifo */
#endif /* HAVE_SYS_STAT_H */
#include "opal/util/output.h"
#include "opal/util/opal_environ.h"
#include "opal/event/event.h"
#include "opal/mca/crs/crs.h"
#include "opal/mca/crs/base/base.h"
#include "opal/runtime/opal_cr.h"
#include "orte/orte_constants.h"
#include "orte/runtime/orte_cr.h"
#include "orte/runtime/runtime.h"
#include "orte/util/sys_info.h"
#include "orte/util/proc_info.h"
#include "orte/util/session_dir.h"
#include "orte/mca/gpr/gpr.h"
#include "orte/mca/gpr/base/base.h"
#include "orte/mca/gpr/gpr_types.h"
#include "orte/mca/pls/pls.h"
#include "orte/mca/pls/base/base.h"
#include "orte/mca/ns/ns.h"
#include "orte/mca/ns/base/base.h"
#include "orte/mca/ns/ns_types.h"
#include "orte/mca/sds/sds.h"
#include "orte/mca/sds/base/base.h"
#include "orte/mca/schema/schema.h"
#include "orte/mca/schema/base/base.h"
#include "orte/mca/rmgr/rmgr.h"
#include "orte/mca/rmgr/base/base.h"
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
#include "orte/mca/routed/base/base.h"
#include "orte/mca/routed/routed.h"
#include "orte/mca/rml/rml.h"
#include "orte/mca/rml/base/base.h"
#include "orte/mca/iof/iof.h"
#include "orte/mca/iof/base/base.h"
#include "orte/mca/odls/odls.h"
#include "orte/mca/odls/base/base.h"
#include "orte/mca/snapc/snapc.h"
#include "orte/mca/snapc/base/base.h"
#include "orte/mca/filem/filem.h"
#include "orte/mca/filem/base/base.h"
/*************
* Local functions
*************/
static int orte_cr_coord_pre_ckpt(void);
static int orte_cr_coord_pre_restart(void);
static int orte_cr_coord_pre_continue(void);
static int orte_cr_coord_post_ckpt(void);
static int orte_cr_coord_post_restart(void);
static int orte_cr_coord_post_continue(void);
/*************
* Local vars
*************/
static opal_cr_coord_callback_fn_t prev_coord_callback = NULL;
static int orte_cr_output = -1;
/*
* CR Init
*/
int orte_cr_init(void)
{
int ret, exit_status = ORTE_SUCCESS;
int val;
/*
* OPAL Frameworks
*/
if (OPAL_SUCCESS != (ret = opal_cr_init() ) ) {
exit_status = ret;
goto cleanup;
}
/*
* Register MCA Parameters
*/
mca_base_param_reg_int_name("orte_cr", "verbose",
"Verbose output for the ORTE Checkpoint/Restart functionality",
false, false,
0,
&val);
if(0 != val) {
orte_cr_output = opal_output_open(NULL);
opal_output_set_verbosity(orte_cr_output, val);
} else {
orte_cr_output = opal_cr_output;
}
opal_output_verbose(10, orte_cr_output,
"orte_cr: init: orte_cr_init()\n");
/* Init ORTE Entry Point Function */
if( ORTE_SUCCESS != (ret = orte_cr_entry_point_init()) ) {
exit_status = ret;
goto cleanup;
}
/* Register the ORTE interlevel coordination callback */
opal_cr_reg_coord_callback(orte_cr_coord, &prev_coord_callback);
cleanup:
return exit_status;
}
/*
* Finalize
*/
int orte_cr_finalize(void)
{
opal_output_verbose(10, orte_cr_output,
"orte_cr: finalize: orte_cr_finalize()");
orte_cr_entry_point_finalize();
/*
* OPAL Frameworks...
*/
opal_cr_finalize();
return ORTE_SUCCESS;
}
/*
* Interlayer coordination callback
*/
int orte_cr_coord(int state)
{
int ret, exit_status = ORTE_SUCCESS;
opal_output_verbose(10, orte_cr_output,
"orte_cr: coord: orte_cr_coord(%s)",
opal_crs_base_state_str((opal_crs_state_type_t)state));
/*
* Before calling the previous callback, we have the opportunity to
* take action given the state.
*/
if(OPAL_CRS_CHECKPOINT == state) {
/* Do Checkpoint Phase work */
orte_cr_coord_pre_ckpt();
}
else if (OPAL_CRS_CONTINUE == state ) {
/* Do Continue Phase work */
orte_cr_coord_pre_continue();
}
else if (OPAL_CRS_RESTART == state ) {
#if 0 /* JJH Do we need this ? */
if (ORTE_SUCCESS != (ret = orte_cr_init_quick_params())) {
opal_output(orte_cr_output,
"ERROR: orte_cr: coord: Unable to init quick parameters\n");
}
#endif
/* Do Restart Phase work */
orte_cr_coord_pre_restart();
}
else if (OPAL_CRS_TERM == state ) {
/* Do Continue Phase work in prep to terminate the application */
}
else {
/* We must have been in an error state from the checkpoint
* recreate everything, as in the Continue Phase
*/
}
/*
* Call the previous callback, which should be OPAL
*/
if(OPAL_SUCCESS != (ret = prev_coord_callback(state)) ) {
exit_status = ret;
goto cleanup;
}
/*
* After calling the previous callback, we have the opportunity to
* take action given the state to tidy up.
*/
if(OPAL_CRS_CHECKPOINT == state) {
/* Do Checkpoint Phase work */
orte_cr_coord_post_ckpt();
}
else if (OPAL_CRS_CONTINUE == state ) {
/* Do Continue Phase work */
orte_cr_coord_post_continue();
}
else if (OPAL_CRS_RESTART == state ) {
/* Do Restart Phase work */
orte_cr_coord_post_restart();
}
else if (OPAL_CRS_TERM == state ) {
/* Do Continue Phase work in prep to terminate the application */
}
else {
/* We must have been in an error state from the checkpoint
* recreate everything, as in the Continue Phase
*/
}
cleanup:
return exit_status;
}
/*************
* Pre Lower Layer
*************/
static int orte_cr_coord_pre_ckpt(void) {
int ret, exit_status = ORTE_SUCCESS;
/*
* All the checkpoint heavey lifting in here...
*/
opal_output_verbose(10, orte_cr_output,
"orte_cr: coord_pre_ckpt: orte_cr_coord_pre_ckpt()");
/*
* Notify IOF
*/
if( ORTE_SUCCESS != (ret = orte_iof.ft_event(OPAL_CRS_CHECKPOINT))) {
exit_status = ret;
goto cleanup;
}
/*
* Notify NS
*/
if( ORTE_SUCCESS != (ret = orte_ns.ft_event(OPAL_CRS_CHECKPOINT))) {
exit_status = ret;
goto cleanup;
}
/*
* Notify GPR
*/
if( ORTE_SUCCESS != (ret = orte_gpr.ft_event(OPAL_CRS_CHECKPOINT))) {
exit_status = ret;
goto cleanup;
}
/*
* Notify RML & OOB
*/
if( ORTE_SUCCESS != (ret = orte_rml.ft_event(OPAL_CRS_CHECKPOINT))) {
exit_status = ret;
goto cleanup;
}
cleanup:
return exit_status;
}
static int orte_cr_coord_pre_restart(void) {
/*
* Can not really do much until OPAL is up and running,
* so defer action until the post_restart function.
*/
opal_output_verbose(10, orte_cr_output,
"orte_cr: coord_pre_restart: orte_cr_coord_pre_restart()");
return ORTE_SUCCESS;
}
static int orte_cr_coord_pre_continue(void) {
/*
* Can not really do much until OPAL is up and running,
* so defer action until the post_continue function.
*/
opal_output_verbose(10, orte_cr_output,
"orte_cr: coord_pre_continue: orte_cr_coord_pre_continue()");
return ORTE_SUCCESS;
}
/*************
* Post Lower Layer
*************/
static int orte_cr_coord_post_ckpt(void) {
/*
* Now that OPAL is shutdown, we really can't do much
* so assume pre_ckpt took care of everything.
*/
opal_output_verbose(10, orte_cr_output,
"orte_cr: coord_post_ckpt: orte_cr_coord_post_ckpt()");
return ORTE_SUCCESS;
}
static int orte_cr_coord_post_restart(void) {
int ret, exit_status = ORTE_SUCCESS;
char * procid_str = NULL;
char * jobid_str = NULL;
int id;
opal_output_verbose(10, orte_cr_output,
"orte_cr: coord_post_restart: orte_cr_coord_post_restart()");
/* Need to invalidate these so we can grab the new values from the environment
* Don't call orte_proc_info_finalize() since we want to preserve some values
* such as orte_process_info.gpr_replica
*/
if (NULL != orte_process_info.ns_replica_uri) {
free(orte_process_info.ns_replica_uri);
orte_process_info.ns_replica_uri = NULL;
}
if (NULL != orte_process_info.gpr_replica_uri) {
free(orte_process_info.gpr_replica_uri);
orte_process_info.gpr_replica_uri = NULL;
}
if (NULL != orte_process_info.tmpdir_base) {
free(orte_process_info.tmpdir_base);
orte_process_info.tmpdir_base = NULL;
}
if (NULL != orte_process_info.top_session_dir) {
free(orte_process_info.top_session_dir);
orte_process_info.top_session_dir = NULL;
}
if (NULL != orte_process_info.universe_session_dir) {
free(orte_process_info.universe_session_dir);
orte_process_info.universe_session_dir = NULL;
}
if (NULL != orte_process_info.job_session_dir) {
free(orte_process_info.job_session_dir);
orte_process_info.job_session_dir = NULL;
}
if (NULL != orte_process_info.proc_session_dir) {
free(orte_process_info.proc_session_dir);
orte_process_info.proc_session_dir = NULL;
}
if (NULL != orte_process_info.sock_stdin) {
free(orte_process_info.sock_stdin);
orte_process_info.sock_stdin = NULL;
}
if (NULL != orte_process_info.sock_stdout) {
free(orte_process_info.sock_stdout);
orte_process_info.sock_stdout = NULL;
}
if (NULL != orte_process_info.sock_stderr) {
free(orte_process_info.sock_stderr);
orte_process_info.sock_stderr = NULL;
}
if( NULL != orte_system_info.nodename ) {
free(orte_system_info.nodename);
orte_system_info.nodename = NULL;
}
/* We want these to be read out of the HNP contact info file ? */
id = mca_base_param_find("gpr", "replica", "uri");
mca_base_param_unset(id);
id = mca_base_param_find("ns", "replica", "uri");
mca_base_param_unset(id);
/*
* Refresh System information
*/
orte_system_info.init = false;
if( ORTE_SUCCESS != (ret = orte_sys_info()) ) {
exit_status = ret;
}
if( ORTE_SUCCESS != (ret = orte_proc_info()) ) {
exit_status = ret;
}
if (NULL != orte_process_info.my_name) {
free(orte_process_info.my_name);
orte_process_info.my_name = NULL;
}
/*
* Re-attach to session directory
*/
orte_universe_info.state = ORTE_UNIVERSE_STATE_PRE_INIT;
if( ORTE_SUCCESS != (ret = orte_univ_info()) ) {
exit_status = ret;
}
/*
* Notify RML & OOB
*/
if( ORTE_SUCCESS != (ret = orte_rml.ft_event(OPAL_CRS_RESTART))) {
exit_status = ret;
goto cleanup;
}
/*
* Startup Discovery Service:
* - Connect to the universe
* Structure Elements Refreshed:
* orte_universe_info.name
* orte_universe_info.host
* orte_universe_info.uid
* orte_universe_info.persistence
* orte_universe_info.scope
* orte_universe_info.seed_uri
* orte_universe_info.console_connected
* orte_universe_info.scriptfile
*
* orte_process_info.ns_replica_uri
* orte_process_info.gpr_replica_uri
*/
if (ORTE_SUCCESS != (ret = orte_sds_base_open())) {
exit_status = ret;
}
if (ORTE_SUCCESS != (ret = orte_sds_base_select())) {
exit_status = ret;
}
if( ORTE_SUCCESS != (ret = orte_sds_base_contact_universe() ) ) {
exit_status = ret;
}
/*
* - Reset Contact information
*/
if(NULL != orte_process_info.ns_replica_uri) {
orte_rml.set_contact_info(orte_process_info.ns_replica_uri);
}
if(NULL != orte_process_info.gpr_replica_uri) {
orte_rml.set_contact_info(orte_process_info.gpr_replica_uri);
}
if( ORTE_SUCCESS != (ret = orte_sds_base_set_name() ) ) {
exit_status = ret;
}
orte_sds_base_close();
/* Session directory stuff:
* orte_process_info.top_session_dir
* orte_process_info.universe_session_dir
* orte_process_info.job_session_dir
* orte_process_info.proc_session_dir
*/
if (ORTE_SUCCESS != (ret = orte_ns.get_jobid_string(&jobid_str, orte_process_info.my_name))) {
exit_status = ret;
}
if (ORTE_SUCCESS != (ret = orte_ns.get_vpid_string(&procid_str, orte_process_info.my_name))) {
exit_status = ret;
}
if (ORTE_SUCCESS != (ret = orte_session_dir(true,
orte_process_info.tmpdir_base,
orte_system_info.user,
orte_system_info.nodename,
NULL, /* Batch ID -- Not used */
orte_universe_info.name,
jobid_str,
procid_str))) {
exit_status = ret;
}
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
/*
* Re-enable communication through the RML
*/
if (ORTE_SUCCESS != (ret = orte_rml.enable_comm())) {
exit_status = ret;
goto cleanup;
}
/*
* Notify NS
*/
if( ORTE_SUCCESS != (ret = orte_ns.ft_event(OPAL_CRS_RESTART))) {
exit_status = ret;
goto cleanup;
}
/*
* Notify GPR
*/
if( ORTE_SUCCESS != (ret = orte_gpr.ft_event(OPAL_CRS_RESTART))) {
exit_status = ret;
goto cleanup;
}
/*
* Notify IOF
*/
if( ORTE_SUCCESS != (ret = orte_iof.ft_event(OPAL_CRS_RESTART))) {
exit_status = ret;
goto cleanup;
}
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
/*
* Re-exchange the routes
*/
if (ORTE_SUCCESS != (ret = orte_routed.init_routes(ORTE_PROC_MY_NAME->jobid, NULL))) {
exit_status = ret;
goto cleanup;
}
/*
* Send new PID to GPR
* The checkpointer could have used a proxy program to boot us
* so the pid that the orted got from fork() may not be the
* PID of this application.
* - Note: BLCR does this because it tries to preseve the PID
* of the program across checkpointes
*/
if( ORTE_SUCCESS != (ret = orte_rmgr.set_process_info(orte_process_info.my_name, getpid(), orte_system_info.nodename) ) ) {
exit_status = ret;
goto cleanup;
}
cleanup:
if (NULL != jobid_str) {
free(jobid_str);
jobid_str = NULL;
}
if (NULL != procid_str) {
free(procid_str);
procid_str = NULL;
}
return exit_status;
}
static int orte_cr_coord_post_continue(void) {
int ret, exit_status = ORTE_SUCCESS;
opal_output_verbose(10, orte_cr_output,
"orte_cr: coord_post_continue: orte_cr_coord_post_continue()\n");
/*
* Notify RML & OOB
*/
if( ORTE_SUCCESS != (ret = orte_rml.ft_event(OPAL_CRS_CONTINUE))) {
exit_status = ret;
goto cleanup;
}
/*
* Notify NS
*/
if( ORTE_SUCCESS != (ret = orte_gpr.ft_event(OPAL_CRS_CONTINUE))) {
exit_status = ret;
goto cleanup;
}
/*
* Notify GPR
*/
if( ORTE_SUCCESS != (ret = orte_gpr.ft_event(OPAL_CRS_CONTINUE))) {
exit_status = ret;
goto cleanup;
}
/*
* Notify IOF
*/
if( ORTE_SUCCESS != (ret = orte_iof.ft_event(OPAL_CRS_CONTINUE))) {
exit_status = ret;
goto cleanup;
}
cleanup:
return exit_status;
}
/*************************************************
* ORTE Entry Point functionality
*************************************************/
int orte_cr_entry_point_init(void)
{
#if 0
/* JJH XXX
* Make sure to finalize the OPAL Entry Point function if it is active.
*/
opal_cr_entry_point_finalize();
#endif
return ORTE_SUCCESS;
}
int orte_cr_entry_point_finalize(void)
{
/* Nothing to do here... */
return ORTE_SUCCESS;
}