1
1
openmpi/orte/mca/odls/bproc/odls_bproc.c

873 строки
29 KiB
C
Исходник Обычный вид 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.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file:
* Part of the bproc launcher.
* See odls_bproc.h for an overview of how it works.
*/
#include "orte_config.h"
#include <stdlib.h>
#include <unistd.h>
#include <sys/types.h>
#include <fcntl.h>
#include <pty.h>
#include <dirent.h>
#include "opal/mca/base/mca_base_param.h"
#include "opal/runtime/opal_progress.h"
#include "opal/threads/condition.h"
#include "opal/util/os_dirpath.h"
#include "opal/util/os_path.h"
#include "opal/util/output.h"
#include "orte/dss/dss.h"
#include "orte/util/sys_info.h"
#include "orte/orte_constants.h"
#include "orte/mca/errmgr/errmgr.h"
#include "orte/mca/gpr/gpr.h"
#include "orte/mca/iof/iof.h"
#include "orte/mca/iof/base/iof_base_setup.h"
#include "orte/mca/ns/base/base.h"
#include "orte/mca/oob/base/base.h"
#include "orte/mca/rml/rml.h"
#include "orte/util/session_dir.h"
#include "orte/util/univ_info.h"
#include "orte/mca/odls/base/odls_private.h"
#include "odls_bproc.h"
/**
* Initialization of the bproc_orted module with all the needed function pointers
*/
orte_odls_base_module_t orte_odls_bproc_module = {
orte_odls_bproc_get_add_procs_data,
orte_odls_bproc_launch_local_procs,
orte_odls_bproc_kill_local_procs,
orte_odls_bproc_signal_local_procs,
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
orte_odls_bproc_deliver_message,
orte_odls_bproc_get_local_proc_names
};
static int odls_bproc_make_dir(char *directory);
static char * odls_bproc_get_base_dir_name(int proc_rank, orte_jobid_t jobid,
orte_std_cntr_t app_context);
static void odls_bproc_delete_dir_tree(char * path);
static int odls_bproc_remove_dir(void);
static void odls_bproc_send_cb(int status, orte_process_name_t * peer,
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
orte_buffer_t* buffer, orte_rml_tag_t tag, void* cbdata);
static int odls_bproc_setup_stdio(orte_process_name_t *proc_name,
int proc_rank, orte_jobid_t jobid,
orte_std_cntr_t app_context, bool connect_stdin);
int orte_odls_bproc_get_add_procs_data(orte_gpr_notify_data_t **data, orte_job_map_t *map)
{
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
orte_gpr_notify_data_t *ndat;
orte_gpr_value_t **values, *value;
orte_std_cntr_t cnt;
opal_list_item_t *item, *m_item;
orte_mapped_node_t *node;
orte_mapped_proc_t *proc;
int rc;
/* set default answer */
*data = NULL;
ndat = OBJ_NEW(orte_gpr_notify_data_t);
if (NULL == ndat) {
ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
return ORTE_ERR_OUT_OF_RESOURCE;
}
/* construct a fake trigger name so that the we can extract the jobid from it later */
if (ORTE_SUCCESS != (rc = orte_schema.get_std_trigger_name(&(ndat->target), "bogus", map->job))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(ndat);
return rc;
}
/* our required info is in the mapped_node objects, so all we
* have to do is transfer it over
*/
for (m_item = opal_list_get_first(&map->nodes);
m_item != opal_list_get_end(&map->nodes);
m_item = opal_list_get_next(m_item)) {
node = (orte_mapped_node_t*)m_item;
for (item = opal_list_get_first(&node->procs);
item != opal_list_get_end(&node->procs);
item = opal_list_get_next(item)) {
proc = (orte_mapped_proc_t*)item;
/* must not have any tokens so that launch_procs can process it correctly */
if (ORTE_SUCCESS != (rc = orte_gpr.create_value(&value, 0, "bogus", 5, 0))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(ndat);
OBJ_RELEASE(value);
return rc;
}
if (ORTE_SUCCESS != (rc = orte_gpr.create_keyval(&(value->keyvals[0]),
ORTE_PROC_NAME_KEY,
ORTE_NAME, &proc->name))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(ndat);
OBJ_RELEASE(value);
return rc;
}
if (ORTE_SUCCESS != (rc = orte_gpr.create_keyval(&(value->keyvals[1]),
ORTE_PROC_APP_CONTEXT_KEY,
ORTE_STD_CNTR, &proc->app_idx))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(ndat);
OBJ_RELEASE(value);
return rc;
}
if (ORTE_SUCCESS != (rc = orte_gpr.create_keyval(&(value->keyvals[2]),
ORTE_NODE_NAME_KEY,
ORTE_STRING, node->nodename))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(ndat);
OBJ_RELEASE(value);
return rc;
}
if (ORTE_SUCCESS != (rc = orte_gpr.create_keyval(&(value->keyvals[3]),
ORTE_PROC_LOCAL_RANK_KEY,
ORTE_VPID, &proc->local_rank))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(ndat);
OBJ_RELEASE(value);
return rc;
}
if (ORTE_SUCCESS != (rc = orte_gpr.create_keyval(&(value->keyvals[4]),
ORTE_NODE_NUM_PROCS_KEY,
ORTE_STD_CNTR, &node->num_procs))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(ndat);
OBJ_RELEASE(value);
return rc;
}
if (ORTE_SUCCESS != (rc = orte_pointer_array_add(&cnt, ndat->values, value))) {
ORTE_ERROR_LOG(rc);
OBJ_RELEASE(ndat);
OBJ_RELEASE(values[0]);
return rc;
}
ndat->cnt += 1;
}
}
*data = ndat;
return ORTE_SUCCESS;
}
/**
* Creates the passed directory. If the directory already exists, it and its
* contents will be deleted then the directory will be created.
* @param directory The directory to be created.
* @retval ORTE_SUCCESS
* @retval error
*/
static int
odls_bproc_make_dir(char *directory)
{
struct stat buf;
mode_t my_mode = S_IRWXU; /* at the least, I need to be able to do anything */
if (0 == stat(directory, &buf)) { /* exists - delete it and its contents */
odls_bproc_delete_dir_tree(directory);
}
/* try to create it with proper mode */
return(opal_os_dirpath_create(directory, my_mode));
}
/**
* Returns a path of the form:
* @code
* /tmp/openmpi-bproc-<user>/<universe>/<jobid>-<app_context>/<proc_rank>/
* @endcode
* which is used to put links to the pty/pipes in
* @param proc_rank the process's rank on the node
* @param jobid the jobid the proc belongs to
* @param app_context the application context number within the job
* @retval path
*/
static char *
odls_bproc_get_base_dir_name(int proc_rank, orte_jobid_t jobid,
orte_std_cntr_t app_context)
{
char *path = NULL, *user = NULL, *job = NULL;
int rc;
/* ensure that system info is set */
orte_sys_info();
if (NULL == orte_universe_info.name) { /* error condition */
ORTE_ERROR_LOG(ORTE_ERROR);
return NULL;
}
rc = orte_ns.convert_jobid_to_string(&job, jobid);
if(ORTE_SUCCESS != rc) {
ORTE_ERROR_LOG(rc);
return NULL;
}
/* get the username set by the bproc pls. We need to get it from here
* because on many bproc systems the method we use to get the username
* from the system on the backend fails and we only get the uid. */
rc = mca_base_param_register_string("pls", "bproc", "username", NULL,
orte_system_info.user);
mca_base_param_lookup_string(rc,&user);
if (0 > asprintf(&path, OPAL_PATH_SEP"tmp"OPAL_PATH_SEP"openmpi-bproc-%s"OPAL_PATH_SEP"%s"OPAL_PATH_SEP"%s-%d"OPAL_PATH_SEP"%d",
user, orte_universe_info.name,
job, (int) app_context, proc_rank)) {
ORTE_ERROR_LOG(ORTE_ERROR);
path = NULL;
}
if(0 < mca_odls_bproc_component.debug) {
opal_output(0, "odls bproc io setup. Path: %s\n", path);
}
free(user);
free(job);
return path;
}
/**
* deletes the passed directory tree recursively
* @param path the path to the base directory to delete
*/
static void
odls_bproc_delete_dir_tree(char * path)
{
DIR *dp;
struct dirent *ep;
char *filenm;
int ret;
struct stat buf;
dp = opendir(path);
if (NULL == dp) {
return;
}
while (NULL != (ep = readdir(dp)) ) {
/* skip: . and .. */
if ((0 != strcmp(ep->d_name, ".")) && (0 != strcmp(ep->d_name, ".."))) {
filenm = opal_os_path(false, path, ep->d_name, NULL);
ret = stat(filenm, &buf);
if (ret < 0 || S_ISDIR(buf.st_mode)) {
odls_bproc_delete_dir_tree(filenm);
free(filenm);
continue;
}
unlink(filenm);
free(filenm);
}
}
closedir(dp);
rmdir(path);
}
/**
* Removes the bproc directory
* @code /tmp/openmpi-bproc-<user>/ @endcode and all of its contents
* @retval ORTE_SUCCESS
* @retval error
*/
static int
odls_bproc_remove_dir()
{
char *frontend = NULL, *user = NULL, *filename = NULL;
int id;
/* get the username set by the bproc pls. We need to get it from here
* because on many bproc systems the method we use to get the username
* from the system on the backend fails and we only get the uid. */
id = mca_base_param_register_string("pls", "bproc", "username", NULL,
orte_system_info.user);
mca_base_param_lookup_string(id,&user);
asprintf(&filename, "openmpi-bproc-%s", user );
if( NULL == filename ) {
ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
return ORTE_ERROR;
}
frontend = opal_os_path(false, "tmp", filename, NULL );
free(filename); /* Always free the filename */
if (NULL == frontend) {
ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
return ORTE_ERROR;
}
/* we do our best to clean up the directory tree, but we ignore errors*/
odls_bproc_delete_dir_tree(frontend);
free(frontend);
return ORTE_SUCCESS;
}
/**
* Callback function for when we tell mpirun we are ready
* @param status
* @param peer
* @param buffer
* @param tag
* @param cbdata
*/
static void
odls_bproc_send_cb(int status, orte_process_name_t * peer,
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
orte_buffer_t* buffer,
orte_rml_tag_t tag, void* cbdata)
2005-09-16 20:59:53 +00:00
{
OBJ_RELEASE(buffer);
}
/**
* Create Standard I/O symlinks in the filesystem for a given proc
*
* Create Standard I/O symlinks in the filesystem for a given proc.
* The symlinks will be placed in:
* @code
* /tmp/openmpi-bproc-<user>/<universe>/<jobid>-<app_context>/<proc_rank>/
* @endcode
*
* The symlinks will be to FIFOs for stdin and stderr. stdout will either
* be to a FIFO or pty, depending on the configuration of Open MPI.
*
* @param proc_rank the process's rank on the node
* @param jobid the jobid the proc belongs to
* @param app_context the application context number within the job
* @param connect_stdin if true, stdin will be connected, otherwise it will be
* set to /dev/null
*
* @retval ORTE_SUCCESS
* @retval error
*/
static int
odls_bproc_setup_stdio(orte_process_name_t *proc_name, int proc_rank,
orte_jobid_t jobid,
orte_std_cntr_t app_context, bool connect_stdin)
{
char *path_prefix, *fd_link_path = NULL;
int rc = ORTE_SUCCESS, fd;
#if defined(HAVE_OPENPTY) && (OMPI_ENABLE_PTY_SUPPORT != 0)
int amaster, aslave;
char pty_name[256];
struct termios term_attrs;
#endif
path_prefix = odls_bproc_get_base_dir_name(proc_rank, jobid, (size_t)app_context);
if (NULL == path_prefix) {
rc = ORTE_ERROR;
ORTE_ERROR_LOG(rc);
goto cleanup;
}
/* check for existence and access, or create it */
if (ORTE_SUCCESS != (rc = odls_bproc_make_dir(path_prefix))) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
/* setup the stdin FIFO. Always use a fifo for the same reason we
always use a pipe in the iof_setup code -- don't want to flush
onto the floor during close */
fd_link_path = opal_os_path( false, path_prefix, "0", NULL );
if (NULL == fd_link_path) {
rc = ORTE_ERROR;
ORTE_ERROR_LOG(rc);
goto cleanup;
}
if (connect_stdin) {
if (0 != mkfifo(fd_link_path, S_IRWXU)) {
perror("odls_bproc mkfifo failed");
rc = ORTE_ERROR;
ORTE_ERROR_LOG(rc);
goto cleanup;
}
fd = open(fd_link_path, O_RDWR);
if (-1 == fd) {
perror("odls_bproc open failed");
rc = ORTE_ERROR;
ORTE_ERROR_LOG(rc);
goto cleanup;
}
orte_iof.iof_publish(proc_name, ORTE_IOF_SINK,
ORTE_IOF_STDIN, fd);
} else {
if(0 != symlink("/dev/null", fd_link_path)) {
perror("odls_bproc could not create symlink");
rc = ORTE_ERROR;
ORTE_ERROR_LOG(rc);
goto cleanup;
}
}
free(fd_link_path);
fd_link_path = NULL;
/* setup the stdout PTY / FIFO */
fd_link_path = opal_os_path( false, path_prefix, "1", NULL );
if (NULL == fd_link_path) {
rc = ORTE_ERROR;
ORTE_ERROR_LOG(rc);
goto cleanup;
}
#if defined(HAVE_OPENPTY) && (OMPI_ENABLE_PTY_SUPPORT != 0)
if (0 != openpty(&amaster, &aslave, pty_name, NULL, NULL)) {
opal_output(0, "odls_bproc: openpty failed, using pipes instead");
goto stdout_fifo_setup;
}
if (0 != symlink(pty_name, fd_link_path)) {
rc = ORTE_ERROR;
ORTE_ERROR_LOG(rc);
goto cleanup;
}
if (tcgetattr(aslave, &term_attrs) < 0) {
rc = ORTE_ERROR;
ORTE_ERROR_LOG(rc);
goto cleanup;
}
term_attrs.c_lflag &= ~ (ECHO | ECHOE | ECHOK |
ECHOCTL | ECHOKE | ECHONL);
term_attrs.c_iflag &= ~ (ICRNL | INLCR | ISTRIP | INPCK | IXON);
term_attrs.c_oflag &= ~ (OCRNL | ONLCR);
if (tcsetattr(aslave, TCSANOW, &term_attrs) == -1) {
rc = ORTE_ERROR;
ORTE_ERROR_LOG(rc);
goto cleanup;
}
orte_iof.iof_publish(proc_name, ORTE_IOF_SOURCE,
ORTE_IOF_STDOUT, amaster);
goto stderr_fifo_setup;
stdout_fifo_setup:
#endif
if (0 != mkfifo(fd_link_path, S_IRWXU)) {
perror("odls_bproc mkfifo failed");
rc = ORTE_ERROR;
goto cleanup;
}
fd = open(fd_link_path, O_RDWR);
if (-1 == fd) {
perror("odls_bproc open failed");
rc = ORTE_ERROR;
goto cleanup;
}
orte_iof.iof_publish(proc_name, ORTE_IOF_SOURCE,
ORTE_IOF_STDOUT, fd);
#if defined(HAVE_OPENPTY) && (OMPI_ENABLE_PTY_SUPPORT != 0)
stderr_fifo_setup:
#endif
free(fd_link_path);
fd_link_path = NULL;
/* setup the stderr FIFO. Always a fifo */
fd_link_path = opal_os_path( false, path_prefix, "2", NULL );
if (NULL == fd_link_path) {
rc = ORTE_ERROR;
ORTE_ERROR_LOG(rc);
goto cleanup;
}
if (0 != mkfifo(fd_link_path, S_IRWXU)) {
perror("odls_bproc mkfifo failed");
rc = ORTE_ERROR;
goto cleanup;
}
fd = open(fd_link_path, O_RDWR);
if (-1 == fd) {
perror("odls_bproc open failed");
rc = ORTE_ERROR;
goto cleanup;
}
orte_iof.iof_publish(proc_name, ORTE_IOF_SOURCE,
ORTE_IOF_STDERR, fd);
cleanup:
if (NULL != path_prefix) {
free(path_prefix);
}
if (NULL != fd_link_path) {
free(fd_link_path);
}
return rc;
}
/**
* Setup io for the current node, then tell orterun we are ready for the actual
* processes.
* @retval ORTE_SUCCESS
* @retval error
*/
int
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
orte_odls_bproc_launch_local_procs(orte_gpr_notify_data_t *data)
{
odls_bproc_child_t *child;
opal_list_item_t* item;
orte_gpr_value_t *value, **values;
orte_gpr_keyval_t *kval;
char *node_name;
int rc;
orte_std_cntr_t i, j, kv, kv2, *sptr;
int src = 0;
2005-09-16 20:59:53 +00:00
orte_buffer_t *ack;
bool connect_stdin;
orte_jobid_t jobid;
int cycle = 0;
char *job_str=NULL, *vpid_str, *uri_file, *my_uri=NULL, *session_dir=NULL;
FILE *fp;
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
orte_vpid_t *vptr;
bool node_included;
/* first, retrieve the job number we are to launch from the
* returned data - we can extract the jobid directly from the
* subscription name we created
*/
if (ORTE_SUCCESS != (rc = orte_schema.extract_jobid_from_std_trigger_name(&jobid, data->target))) {
ORTE_ERROR_LOG(rc);
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
src = rc;
goto CALLHOME;
}
/**
Add ability to trap and propagate SIGUSR1/2 to remote processes. There are a number of small changes that hit a bunch of files: 1. Changed the RMGR and PLS APIs to add "signal_job" and "signal_proc" entry points. Only the "signal_job" entries are implemented - none of the components have implementations for "signal_proc" at this time. Thus, you can signal all of the procs in a job, but cannot currently signal only one specific proc. 2. Implemented those new API functions in all components except xgrid (Brian will do so very soon). Only the rsh/ssh and fork modules have been tested, however, and only under OS-X. 3. Added signal traps and callback functions for SIGUSR1/2 to orterun/mpirun that catch those signals and call the appropriate commands to propagate them out to all processes in the job. 4. Added a new test directory under the orte branch to (eventually) hold unit and system level tests for just the run-time. Since our test branch of the repository is under restricted access, people working on the RTE were continually developing their own system-level tests - thus making it hard to help diagnose problems. I have moved the more commonly-used functions here, and added one specifically for testing the SIGUSR1/2 functionality. I will be contacting people directly to seek help with testing the changes on more environments. Other than compile issues, you should see absolutely no change in behavior on any of your systems - this additional functionality is transparent to anyone who does not issue a SIGUSR1/2 to mpirun. Ralph This commit was SVN r10258.
2006-06-08 18:27:17 +00:00
* hack for bproc4, change process group so that we do not receive signals
* from the parent/front-end process, as bproc4 does not currently allow the
* process to intercept the signal
*/
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
setpgid(0,0);
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
/* set the flag indicating this node is not included in the launch data */
node_included = false;
/* loop through the returned data to find the global info and
* the info for processes going onto this node
*/
values = (orte_gpr_value_t**)(data->values)->addr;
for (j=0, i=0; i < data->cnt && j < (data->values)->size; j++) { /* loop through all returned values */
if (NULL != values[j]) {
i++;
value = values[j];
/* this must have come from one of the process containers, so it must
* contain data for a proc structure - see if it belongs to this node
*/
for (kv=0; kv < value->cnt; kv++) {
kval = value->keyvals[kv];
if (strcmp(kval->key, ORTE_NODE_NAME_KEY) == 0) {
/* Most C-compilers will bark if we try to directly compare the string in the
* kval data area against a regular string, so we need to "get" the data
* so we can access it */
if (ORTE_SUCCESS != (rc = orte_dss.get((void**)&node_name, kval->value, ORTE_STRING))) {
ORTE_ERROR_LOG(rc);
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
src = rc;
goto CALLHOME;
}
/* if this is our node...must also protect against a zero-length string */
if (NULL != node_name && 0 == strcmp(node_name, orte_system_info.nodename)) {
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
/* indicate that there is something for us to do */
node_included = true;
/* setup and populate the child object */
child = OBJ_NEW(odls_bproc_child_t);
for (kv2 = 0; kv2 < value->cnt; kv2++) {
kval = value->keyvals[kv2];
if(strcmp(kval->key, ORTE_PROC_NAME_KEY) == 0) {
/* copy the name into the child object */
if (ORTE_SUCCESS != (rc = orte_dss.copy((void**)&(child->name), kval->value->data, ORTE_NAME))) {
ORTE_ERROR_LOG(rc);
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
src = rc;
goto CALLHOME;
}
continue;
}
if(strcmp(kval->key, ORTE_PROC_APP_CONTEXT_KEY) == 0) {
if (ORTE_SUCCESS != (rc = orte_dss.get((void**)&sptr, kval->value, ORTE_STD_CNTR))) {
ORTE_ERROR_LOG(rc);
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
src = rc;
goto CALLHOME;
}
child->app_idx = *sptr; /* save the index into the app_context objects */
continue;
}
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
if(strcmp(kval->key, ORTE_PROC_LOCAL_RANK_KEY) == 0) {
if (ORTE_SUCCESS != (rc = orte_dss.get((void**)&vptr, kval->value, ORTE_VPID))) {
ORTE_ERROR_LOG(rc);
src = rc;
goto CALLHOME;
}
child->local_rank = *vptr; /* save the local_rank */
continue;
}
if(strcmp(kval->key, ORTE_NODE_NUM_PROCS_KEY) == 0) {
if (ORTE_SUCCESS != (rc = orte_dss.get((void**)&sptr, kval->value, ORTE_STD_CNTR))) {
ORTE_ERROR_LOG(rc);
src = rc;
goto CALLHOME;
}
child->num_procs = *sptr; /* save the number of procs from this job on this node */
continue;
}
} /* kv2 */
/* protect operation on the global list of children */
OPAL_THREAD_LOCK(&mca_odls_bproc_component.lock);
opal_list_append(&mca_odls_bproc_component.children, &child->super);
opal_condition_signal(&mca_odls_bproc_component.cond);
OPAL_THREAD_UNLOCK(&mca_odls_bproc_component.lock);
}
}
} /* for kv */
} /* for j */
}
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
/* if there is nothing for us to do, we still have to report back
* before we just return
*/
if (!node_included) {
rc = ORTE_SUCCESS;
goto CALLHOME;
}
/* setup some values we'll need to drop my uri for each child */
orte_ns.convert_jobid_to_string(&job_str, jobid);
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
my_uri = orte_rml.get_contact_info();
/* set up the io files for our children */
for(item = opal_list_get_first(&mca_odls_bproc_component.children);
item != opal_list_get_end(&mca_odls_bproc_component.children);
item = opal_list_get_next(item)) {
child = (odls_bproc_child_t *) item;
if(0 < mca_odls_bproc_component.debug) {
opal_output(0, "orte_odls_bproc_launch: setting up io for "
"%s proc rank %ld\n",
ORTE_NAME_PRINT((child->name)),
(long)child->name->vpid);
}
/* only setup to forward stdin if it is rank 0, otherwise connect
* to /dev/null */
if(0 == child->name->vpid) {
connect_stdin = true;
} else {
connect_stdin = false;
}
rc = odls_bproc_setup_stdio(child->name, cycle,
jobid, child->app_idx,
connect_stdin);
if (ORTE_SUCCESS != rc) {
ORTE_ERROR_LOG(rc);
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
src = rc;
goto CALLHOME;
}
/* record my uri in a file within the session directory so the child can contact me */
/* get the session dir for this proc */
orte_ns.convert_vpid_to_string(&vpid_str, child->name->vpid);
if (ORTE_SUCCESS != (rc = orte_session_dir(true, NULL, NULL, NULL,
NULL, NULL, job_str, vpid_str))) {
ORTE_ERROR_LOG(rc);
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
src = rc;
goto CALLHOME;
}
/* get the session dir name so we can put the file there */
if (ORTE_SUCCESS != (rc = orte_session_dir_get_name(&session_dir, NULL, NULL, NULL,
NULL, NULL, NULL, job_str, vpid_str))) {
ORTE_ERROR_LOG(rc);
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
src = rc;
goto CALLHOME;
}
free(vpid_str);
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
/* create the file and put my uri, this child's local rank, and the
* number of local procs into it */
uri_file = opal_os_path(false, session_dir, "orted-uri.txt", NULL);
fp = fopen(uri_file, "w");
if (NULL == fp) {
ORTE_ERROR_LOG(ORTE_ERR_FILE_OPEN_FAILURE);
rc = ORTE_ERR_FILE_OPEN_FAILURE;
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
src = rc;
goto CALLHOME;
}
fprintf(fp, "%s\n", my_uri);
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
fprintf(fp, "%ld\n", (long)child->local_rank);
fprintf(fp, "%ld\n", (long)child->num_procs);
fclose(fp);
free(uri_file);
cycle++;
}
/* release the jobid string and uri */
free(job_str);
free(my_uri);
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
CALLHOME:
2005-09-16 20:59:53 +00:00
/* message to indicate that we are ready */
ack = OBJ_NEW(orte_buffer_t);
rc = orte_dss.pack(ack, &src, 1, ORTE_INT);
if(ORTE_SUCCESS != rc) {
ORTE_ERROR_LOG(rc);
}
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
rc = orte_rml.send_buffer_nb(ORTE_PROC_MY_HNP, ack, ORTE_RML_TAG_BPROC, 0,
odls_bproc_send_cb, NULL);
if (0 > rc) {
ORTE_ERROR_LOG(rc);
goto cleanup;
}
rc = ORTE_SUCCESS;
cleanup:
return rc;
}
/**
* Function to terminate a job. Since this component only runs on remote nodes
* and doesn't actually launch any processes, this function is not needed
* so is a noop.
Add ability to trap and propagate SIGUSR1/2 to remote processes. There are a number of small changes that hit a bunch of files: 1. Changed the RMGR and PLS APIs to add "signal_job" and "signal_proc" entry points. Only the "signal_job" entries are implemented - none of the components have implementations for "signal_proc" at this time. Thus, you can signal all of the procs in a job, but cannot currently signal only one specific proc. 2. Implemented those new API functions in all components except xgrid (Brian will do so very soon). Only the rsh/ssh and fork modules have been tested, however, and only under OS-X. 3. Added signal traps and callback functions for SIGUSR1/2 to orterun/mpirun that catch those signals and call the appropriate commands to propagate them out to all processes in the job. 4. Added a new test directory under the orte branch to (eventually) hold unit and system level tests for just the run-time. Since our test branch of the repository is under restricted access, people working on the RTE were continually developing their own system-level tests - thus making it hard to help diagnose problems. I have moved the more commonly-used functions here, and added one specifically for testing the SIGUSR1/2 functionality. I will be contacting people directly to seek help with testing the changes on more environments. Other than compile issues, you should see absolutely no change in behavior on any of your systems - this additional functionality is transparent to anyone who does not issue a SIGUSR1/2 to mpirun. Ralph This commit was SVN r10258.
2006-06-08 18:27:17 +00:00
*/
int orte_odls_bproc_kill_local_procs(orte_jobid_t job, bool set_state)
Add ability to trap and propagate SIGUSR1/2 to remote processes. There are a number of small changes that hit a bunch of files: 1. Changed the RMGR and PLS APIs to add "signal_job" and "signal_proc" entry points. Only the "signal_job" entries are implemented - none of the components have implementations for "signal_proc" at this time. Thus, you can signal all of the procs in a job, but cannot currently signal only one specific proc. 2. Implemented those new API functions in all components except xgrid (Brian will do so very soon). Only the rsh/ssh and fork modules have been tested, however, and only under OS-X. 3. Added signal traps and callback functions for SIGUSR1/2 to orterun/mpirun that catch those signals and call the appropriate commands to propagate them out to all processes in the job. 4. Added a new test directory under the orte branch to (eventually) hold unit and system level tests for just the run-time. Since our test branch of the repository is under restricted access, people working on the RTE were continually developing their own system-level tests - thus making it hard to help diagnose problems. I have moved the more commonly-used functions here, and added one specifically for testing the SIGUSR1/2 functionality. I will be contacting people directly to seek help with testing the changes on more environments. Other than compile issues, you should see absolutely no change in behavior on any of your systems - this additional functionality is transparent to anyone who does not issue a SIGUSR1/2 to mpirun. Ralph This commit was SVN r10258.
2006-06-08 18:27:17 +00:00
{
orte_iof.iof_flush();
Add ability to trap and propagate SIGUSR1/2 to remote processes. There are a number of small changes that hit a bunch of files: 1. Changed the RMGR and PLS APIs to add "signal_job" and "signal_proc" entry points. Only the "signal_job" entries are implemented - none of the components have implementations for "signal_proc" at this time. Thus, you can signal all of the procs in a job, but cannot currently signal only one specific proc. 2. Implemented those new API functions in all components except xgrid (Brian will do so very soon). Only the rsh/ssh and fork modules have been tested, however, and only under OS-X. 3. Added signal traps and callback functions for SIGUSR1/2 to orterun/mpirun that catch those signals and call the appropriate commands to propagate them out to all processes in the job. 4. Added a new test directory under the orte branch to (eventually) hold unit and system level tests for just the run-time. Since our test branch of the repository is under restricted access, people working on the RTE were continually developing their own system-level tests - thus making it hard to help diagnose problems. I have moved the more commonly-used functions here, and added one specifically for testing the SIGUSR1/2 functionality. I will be contacting people directly to seek help with testing the changes on more environments. Other than compile issues, you should see absolutely no change in behavior on any of your systems - this additional functionality is transparent to anyone who does not issue a SIGUSR1/2 to mpirun. Ralph This commit was SVN r10258.
2006-06-08 18:27:17 +00:00
return ORTE_SUCCESS;
}
/**
* Function to signal a process. Since this component only runs on remote nodes
* and doesn't actually launch any processes, this function is not needed
* so is a noop.
* @param proc the process's name
* @param signal The signal to send
* @retval ORTE_SUCCESS
*/
int orte_odls_bproc_signal_local_procs(const orte_process_name_t* proc, int32_t signal)
Add ability to trap and propagate SIGUSR1/2 to remote processes. There are a number of small changes that hit a bunch of files: 1. Changed the RMGR and PLS APIs to add "signal_job" and "signal_proc" entry points. Only the "signal_job" entries are implemented - none of the components have implementations for "signal_proc" at this time. Thus, you can signal all of the procs in a job, but cannot currently signal only one specific proc. 2. Implemented those new API functions in all components except xgrid (Brian will do so very soon). Only the rsh/ssh and fork modules have been tested, however, and only under OS-X. 3. Added signal traps and callback functions for SIGUSR1/2 to orterun/mpirun that catch those signals and call the appropriate commands to propagate them out to all processes in the job. 4. Added a new test directory under the orte branch to (eventually) hold unit and system level tests for just the run-time. Since our test branch of the repository is under restricted access, people working on the RTE were continually developing their own system-level tests - thus making it hard to help diagnose problems. I have moved the more commonly-used functions here, and added one specifically for testing the SIGUSR1/2 functionality. I will be contacting people directly to seek help with testing the changes on more environments. Other than compile issues, you should see absolutely no change in behavior on any of your systems - this additional functionality is transparent to anyone who does not issue a SIGUSR1/2 to mpirun. Ralph This commit was SVN r10258.
2006-06-08 18:27:17 +00:00
{
orte_iof.iof_flush();
Add ability to trap and propagate SIGUSR1/2 to remote processes. There are a number of small changes that hit a bunch of files: 1. Changed the RMGR and PLS APIs to add "signal_job" and "signal_proc" entry points. Only the "signal_job" entries are implemented - none of the components have implementations for "signal_proc" at this time. Thus, you can signal all of the procs in a job, but cannot currently signal only one specific proc. 2. Implemented those new API functions in all components except xgrid (Brian will do so very soon). Only the rsh/ssh and fork modules have been tested, however, and only under OS-X. 3. Added signal traps and callback functions for SIGUSR1/2 to orterun/mpirun that catch those signals and call the appropriate commands to propagate them out to all processes in the job. 4. Added a new test directory under the orte branch to (eventually) hold unit and system level tests for just the run-time. Since our test branch of the repository is under restricted access, people working on the RTE were continually developing their own system-level tests - thus making it hard to help diagnose problems. I have moved the more commonly-used functions here, and added one specifically for testing the SIGUSR1/2 functionality. I will be contacting people directly to seek help with testing the changes on more environments. Other than compile issues, you should see absolutely no change in behavior on any of your systems - this additional functionality is transparent to anyone who does not issue a SIGUSR1/2 to mpirun. Ralph This commit was SVN r10258.
2006-06-08 18:27:17 +00:00
return ORTE_SUCCESS;
}
int orte_odls_bproc_deliver_message(orte_jobid_t job, orte_buffer_t *buffer, orte_rml_tag_t tag)
{
int rc;
opal_list_item_t *item;
orte_odls_child_t *child;
/* protect operations involving the global list of children */
OPAL_THREAD_LOCK(&mca_odls_bproc_component.lock);
for (item = opal_list_get_first(&mca_odls_bproc_component.children);
item != opal_list_get_end(&mca_odls_bproc_component.children);
item = opal_list_get_next(item)) {
child = (orte_odls_child_t*)item;
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief: 1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names. 2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used. 3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying. Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed. This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems. This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
/* do we have a child from the specified job. Because the
* job could be given as a WILDCARD value, we must use
* the dss.compare function to check for equality.
*/
if (ORTE_EQUAL != orte_dss.compare(&job, &(child->name->jobid), ORTE_JOBID)) {
continue;
}
/* if so, send the message */
rc = orte_rml.send_buffer(child->name, buffer, tag, 0);
if (rc < 0) {
ORTE_ERROR_LOG(rc);
}
}
opal_condition_signal(&mca_odls_bproc_component.cond);
OPAL_THREAD_UNLOCK(&mca_odls_bproc_component.lock);
return ORTE_SUCCESS;
}
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
int orte_odls_bproc_get_local_proc_names(opal_list_t *names, orte_jobid_t job)
{
opal_list_item_t *item;
orte_odls_child_t *child;
orte_namelist_t *nitem;
/* protect operations involving the global list of children */
OPAL_THREAD_LOCK(&mca_odls_bproc_component.lock);
for (item = opal_list_get_first(&mca_odls_bproc_component.children);
item != opal_list_get_end(&mca_odls_bproc_component.children);
item = opal_list_get_next(item)) {
child = (orte_odls_child_t*)item;
/* do we have a child from the specified job. Because the
* job could be given as a WILDCARD value, we must use
* the dss.compare function to check for equality.
*/
if (ORTE_EQUAL != orte_dss.compare(&job, &(child->name->jobid), ORTE_JOBID)) {
continue;
}
/* add this name to the list */
nitem = OBJ_NEW(orte_namelist_t);
orte_dss.copy((void**)&nitem->name, child->name, ORTE_NAME);
opal_list_append(names, &nitem->item);
}
opal_condition_signal(&mca_odls_bproc_component.cond);
OPAL_THREAD_UNLOCK(&mca_odls_bproc_component.lock);
return ORTE_SUCCESS;
}
/**
* Finalizes the bproc module. Cleanup tmp directory/files
2005-09-16 20:59:53 +00:00
* used for I/O forwarding.
* @retval ORTE_SUCCESS
*/
int orte_odls_bproc_finalize(void)
2005-09-16 20:59:53 +00:00
{
orte_iof.iof_flush();
odls_bproc_remove_dir();
orte_session_dir_finalize(orte_process_info.my_name);
return ORTE_SUCCESS;
}