1
1
openmpi/ompi/mca/coll/tuned/coll_tuned_dynamic_rules.c

421 строка
12 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
*
* $HEADER$
*/
#include "ompi_config.h"
#include "mpi.h"
#include "opal/mca/mca.h"
#include "ompi/constants.h"
#include "coll_tuned.h"
/* need to include our own topo prototypes so we can malloc data on the comm correctly */
#include "coll_tuned_topo.h"
/* also need the dynamic rule structures */
#include "coll_tuned_dynamic_rules.h"
#include <stdlib.h>
#include <stdio.h>
#include "coll_tuned_util.h"
ompi_coll_alg_rule_t* ompi_coll_tuned_mk_alg_rules (int n_alg)
{
int i;
ompi_coll_alg_rule_t* alg_rules;
alg_rules = (ompi_coll_alg_rule_t *) calloc (n_alg, sizeof (ompi_coll_alg_rule_t));
if (!alg_rules) return (alg_rules);
/* set all we can at this point */
for (i=0;i<n_alg;i++) {
alg_rules[i].alg_rule_id = i;
}
return (alg_rules);
}
ompi_coll_com_rule_t* ompi_coll_tuned_mk_com_rules (int n_com_rules, int alg_rule_id)
{
int i;
ompi_coll_com_rule_t * com_rules;
com_rules = (ompi_coll_com_rule_t *) calloc (n_com_rules, sizeof (ompi_coll_com_rule_t));
if (!com_rules) return (com_rules);
for (i=0;i<n_com_rules;i++) {
com_rules[i].mpi_comsize = 0; /* unknown */
com_rules[i].alg_rule_id = alg_rule_id;
com_rules[i].com_rule_id = i;
com_rules[i].n_msg_sizes = 0; /* unknown */
com_rules[i].msg_rules = (ompi_coll_msg_rule_t *) NULL;
}
return (com_rules);
}
ompi_coll_msg_rule_t* ompi_coll_tuned_mk_msg_rules (int n_msg_rules, int alg_rule_id, int com_rule_id, int mpi_comsize)
{
int i;
ompi_coll_msg_rule_t *msg_rules;
msg_rules = (ompi_coll_msg_rule_t *) calloc (n_msg_rules, sizeof (ompi_coll_msg_rule_t));
if (!msg_rules) return (msg_rules);
for (i=0;i<n_msg_rules;i++) {
msg_rules[i].mpi_comsize = mpi_comsize;
msg_rules[i].alg_rule_id = alg_rule_id;
msg_rules[i].com_rule_id = com_rule_id;
msg_rules[i].msg_rule_id = i;
msg_rules[i].msg_size = 0; /* unknown */
msg_rules[i].result_alg = 0; /* unknown */
msg_rules[i].result_topo_faninout = 0; /* unknown */
msg_rules[i].result_segsize = 0; /* unknown */
Adding flow control for leaf nodes in generalized reduce structure. This "feature" is disabled by default and it should not affect the current performance. In case when the message size is large and segment size is smaller than eager size for particular interface, the leaf nodes in generalized reduce function can overflood parent nodes by sending all segments without any synchronization. This can cause the parent to have HIGH number of unexpected messages (think 16MB message with 1KB segments for example). In case of binomial algorithm root node always has at least one child which is leaf, so this can potentially affect the root's performance significantly [Especially in large communicators where root may have quite a few children (binomial tree for example)]. When the segment size is bigger than the eager size, rendezvous protocol ensures that this does not happen so it is not necessary. Originally, the problem was exposed in "infinite" bucket allocator clean up time for "small" segment sizes (which may explain some "deadlocks" on Thunderbird tests). To prevent this, we allow user to specify mca parameter "--mca coll_tuned_reduce_algorithm_max_requests NUM" this limits number of outstanding messages from a leaf node in generalized reduce to the parent to NUM. Messages are sent as non-blocking synchrnous messages, so syncronization happens at "wait" time. The synchronization actually improved performance of pipeline and binomial algorithm for large message sizes with 1KB segments over MX, but I need to test it some more to make sure it is consistent. Since there is no easy way to find out what is "the eager" size for particular btl, I set the limit to 4000B. If message/individual segment size is greater than 4000B - we will not use this feature. This variable may or may not be exposed as mca parameter later... I did not have any problems running it and both "default" and "synchronous" tests passed Intel Reduce* tests up to 80 processes (over MX). This commit was SVN r14518.
2007-04-26 00:39:53 +04:00
msg_rules[i].result_max_requests = 0; /* unknown & default */
}
return (msg_rules);
}
/*
* Debug / IO routines
*
*/
int ompi_coll_tuned_dump_msg_rule (ompi_coll_msg_rule_t* msg_p)
{
if (!msg_p) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"Message rule was a NULL ptr?!\n"));
return (-1);
}
OPAL_OUTPUT((ompi_coll_tuned_stream,"alg_id %3d\tcom_id %3d\tcom_size %3d\tmsg_id %3d\t", msg_p->alg_rule_id,
msg_p->com_rule_id, msg_p->mpi_comsize, msg_p->msg_rule_id));
OPAL_OUTPUT((ompi_coll_tuned_stream,"msg_size %6d -> algorithm %2d\ttopo in/out %2d\tsegsize %5ld\tmax_requests %4d\n",
Adding flow control for leaf nodes in generalized reduce structure. This "feature" is disabled by default and it should not affect the current performance. In case when the message size is large and segment size is smaller than eager size for particular interface, the leaf nodes in generalized reduce function can overflood parent nodes by sending all segments without any synchronization. This can cause the parent to have HIGH number of unexpected messages (think 16MB message with 1KB segments for example). In case of binomial algorithm root node always has at least one child which is leaf, so this can potentially affect the root's performance significantly [Especially in large communicators where root may have quite a few children (binomial tree for example)]. When the segment size is bigger than the eager size, rendezvous protocol ensures that this does not happen so it is not necessary. Originally, the problem was exposed in "infinite" bucket allocator clean up time for "small" segment sizes (which may explain some "deadlocks" on Thunderbird tests). To prevent this, we allow user to specify mca parameter "--mca coll_tuned_reduce_algorithm_max_requests NUM" this limits number of outstanding messages from a leaf node in generalized reduce to the parent to NUM. Messages are sent as non-blocking synchrnous messages, so syncronization happens at "wait" time. The synchronization actually improved performance of pipeline and binomial algorithm for large message sizes with 1KB segments over MX, but I need to test it some more to make sure it is consistent. Since there is no easy way to find out what is "the eager" size for particular btl, I set the limit to 4000B. If message/individual segment size is greater than 4000B - we will not use this feature. This variable may or may not be exposed as mca parameter later... I did not have any problems running it and both "default" and "synchronous" tests passed Intel Reduce* tests up to 80 processes (over MX). This commit was SVN r14518.
2007-04-26 00:39:53 +04:00
msg_p->msg_size, msg_p->result_alg, msg_p->result_topo_faninout, msg_p->result_segsize,
msg_p->result_max_requests));
return (0);
}
int ompi_coll_tuned_dump_com_rule (ompi_coll_com_rule_t* com_p)
{
int i;
if (!com_p) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"Com rule was a NULL ptr?!\n"));
return (-1);
}
OPAL_OUTPUT((ompi_coll_tuned_stream, "alg_id %3d\tcom_id %3d\tcom_size %3d\t", com_p->alg_rule_id, com_p->com_rule_id, com_p->mpi_comsize));
if (!com_p->n_msg_sizes) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"no msgsizes defined\n"));
return (0);
}
OPAL_OUTPUT((ompi_coll_tuned_stream,"number of message sizes %3d\n", com_p->n_msg_sizes));
for (i=0;i<com_p->n_msg_sizes;i++) {
ompi_coll_tuned_dump_msg_rule (&(com_p->msg_rules[i]));
}
return (0);
}
int ompi_coll_tuned_dump_alg_rule (ompi_coll_alg_rule_t* alg_p)
{
int i;
if (!alg_p) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"Algorithm rule was a NULL ptr?!\n"));
return (-1);
}
OPAL_OUTPUT((ompi_coll_tuned_stream,"alg_id %3d\t", alg_p->alg_rule_id));
if (!alg_p->n_com_sizes) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"no coms defined\n"));
return (0);
}
OPAL_OUTPUT((ompi_coll_tuned_stream,"number of com sizes %3d\n", alg_p->n_com_sizes));
for (i=0;i<alg_p->n_com_sizes;i++) {
ompi_coll_tuned_dump_com_rule (&(alg_p->com_rules[i]));
}
return (0);
}
int ompi_coll_tuned_dump_all_rules (ompi_coll_alg_rule_t* alg_p, int n_rules)
{
int i;
if (!alg_p) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"Algorithm rule was a NULL ptr?!\n"));
return (-1);
}
OPAL_OUTPUT((ompi_coll_tuned_stream,"Number of algorithm rules %3d\n", n_rules));
for (i=0;i<n_rules;i++) {
ompi_coll_tuned_dump_alg_rule (&(alg_p[i]));
}
return (0);
}
/*
* Memory free routines
*
*/
int ompi_coll_tuned_free_msg_rules_in_com_rule (ompi_coll_com_rule_t* com_p)
{
int rc=0;
ompi_coll_msg_rule_t* msg_p;
if (!com_p) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"attempt to free NULL com_rule ptr\n"));
return (-1);
}
if (com_p->n_msg_sizes) {
msg_p = com_p->msg_rules;
if (!msg_p) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"attempt to free NULL msg_rules when msg count was %d\n", com_p->n_msg_sizes));
rc = -1; /* some error */
}
else {
/* ok, memory exists for the msg rules so free that first */
free (com_p->msg_rules);
com_p->msg_rules = (ompi_coll_msg_rule_t*) NULL;
}
} /* if we have msg rules to free as well */
return (rc);
}
int ompi_coll_tuned_free_coms_in_alg_rule (ompi_coll_alg_rule_t* alg_p)
{
int rc=0;
int i;
ompi_coll_com_rule_t* com_p;
if (!alg_p) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"attempt to free NULL alg_rule ptr\n"));
return (-1);
}
if (alg_p->n_com_sizes) {
com_p = alg_p->com_rules;
if (!com_p) {
OPAL_OUTPUT((ompi_coll_tuned_stream,"attempt to free NULL com_rules when com count was %d\n", alg_p->n_com_sizes));
}
else {
/* ok, memory exists for the com rules so free their message rules first */
for (i=0;i<alg_p->n_com_sizes;i++) {
com_p = &(alg_p->com_rules[i]);
ompi_coll_tuned_free_msg_rules_in_com_rule (com_p);
}
/* we are now free to free the com rules themselives */
free (alg_p->com_rules);
alg_p->com_rules = (ompi_coll_com_rule_t*) NULL;
}
} /* if we have msg rules to free as well */
return (rc);
}
int ompi_coll_tuned_free_all_rules (ompi_coll_alg_rule_t* alg_p, int n_algs)
{
int i;
int rc = 0;
for(i=0;i<n_algs;i++) {
rc += ompi_coll_tuned_free_coms_in_alg_rule (&(alg_p[i]));
}
free (alg_p);
return (rc);
}
/*
* query functions
* i.e. the functions that get me the algorithm, topo fanin/out and segment size fast
* and also get the rules that are needed by each communicator as needed
*
*/
/*
* This function is used to get the pointer to the nearest (less than or equal)
* com rule for this MPI collective (alg_id) for a given
* MPI communicator size. The complete rule base must be presented.
*
* If no rule exits returns NULL, else the com rule ptr
* (which can be used in the coll_tuned_get_target_method_params() call)
*
*/
ompi_coll_com_rule_t* ompi_coll_tuned_get_com_rule_ptr (ompi_coll_alg_rule_t* rules, int alg_id, int mpi_comsize)
{
ompi_coll_alg_rule_t* alg_p = (ompi_coll_alg_rule_t*) NULL;
ompi_coll_com_rule_t* com_p = (ompi_coll_com_rule_t*) NULL;
ompi_coll_com_rule_t* best_com_p = (ompi_coll_com_rule_t*) NULL;
int i, best;
if (!rules) { /* no rule base no resulting com rule */
return ((ompi_coll_com_rule_t*)NULL);
}
alg_p = &(rules[alg_id]); /* get the algorithm rule pointer */
if (!alg_p->n_com_sizes) { /* check for count of communicator sizes */
return ((ompi_coll_com_rule_t*)NULL); /* no com sizes so no rule */
}
/* ok have some com sizes, now to find the one closest to my mpi_comsize */
/* make a copy of the first com rule */
best_com_p = com_p = alg_p->com_rules;
i = best = 0;
while (i<alg_p->n_com_sizes) {
/* OPAL_OUTPUT((ompi_coll_tuned_stream,"checking comsize %d against alg_id %d com_id %d index %d com_size %d", */
/* mpi_comsize, com_p->alg_rule_id, com_p->com_rule_id, i, com_p->mpi_comsize)); */
if (com_p->mpi_comsize <= mpi_comsize) {
best = i;
best_com_p = com_p;
/* OPAL_OUTPUT((ompi_coll_tuned_stream(":ok\n")); */
}
else {
/* OPAL_OUTPUT((ompi_coll_tuned_stream(":nop\n")); */
break;
}
/* go to the next entry */
com_p++;
i++;
}
OPAL_OUTPUT((ompi_coll_tuned_stream,"Selected the following com rule id %d\n", best_com_p->com_rule_id));
ompi_coll_tuned_dump_com_rule (best_com_p);
return (best_com_p);
}
/*
* This function takes a com_rule ptr (from the communicators coll tuned data structure)
* (Which is chosen for a particular MPI collective)
* and a (total_)msg_size and it returns (0) and a algorithm to use and a recommended topo faninout and segment size
* all based on the user supplied rules
*
* Just like the above functions it uses a less than or equal msg size
* (hense config file must have a default defined for '0' if we reach this point)
* else if no rules match we return '0' + '0,0' or used fixed decision table with no topo chand and no segmentation
* of users data.. shame.
*
* On error return 0 so we default to fixed rules anyway :)
*
*/
int ompi_coll_tuned_get_target_method_params (ompi_coll_com_rule_t* base_com_rule, int mpi_msgsize, int *result_topo_faninout,
int* result_segsize, int* max_requests)
{
ompi_coll_msg_rule_t* msg_p = (ompi_coll_msg_rule_t*) NULL;
ompi_coll_msg_rule_t* best_msg_p = (ompi_coll_msg_rule_t*) NULL;
int i, best;
if (!base_com_rule) {
return (0);
}
if (!result_topo_faninout) {
return (0);
}
if (!result_segsize) {
return (0);
}
if (!max_requests) {
return (0);
}
if (!base_com_rule->n_msg_sizes) { /* check for count of message sizes */
return (0); /* no msg sizes so no rule */
}
/* ok have some msg sizes, now to find the one closest to my mpi_msgsize */
/* make a copy of the first msg rule */
best_msg_p = msg_p = base_com_rule->msg_rules;
i = best = 0;
while (i<base_com_rule->n_msg_sizes) {
/* OPAL_OUTPUT((ompi_coll_tuned_stream,"checking mpi_msgsize %d against com_id %d msg_id %d index %d msg_size %d", */
/* mpi_msgsize, msg_p->com_rule_id, msg_p->msg_rule_id, i, msg_p->msg_size)); */
if (msg_p->msg_size <= mpi_msgsize) {
best = i;
best_msg_p = msg_p;
/* OPAL_OUTPUT((ompi_coll_tuned_stream(":ok\n")); */
}
else {
/* OPAL_OUTPUT((ompi_coll_tuned_stream(":nop\n")); */
break;
}
/* go to the next entry */
msg_p++;
i++;
}
OPAL_OUTPUT((ompi_coll_tuned_stream,"Selected the following msg rule id %d\n", best_msg_p->msg_rule_id));
ompi_coll_tuned_dump_msg_rule (best_msg_p);
/* return the segment size */
*result_topo_faninout = best_msg_p->result_topo_faninout;
/* return the segment size */
*result_segsize = best_msg_p->result_segsize;
/* return the maximum requests */
*max_requests = best_msg_p->result_max_requests;
/* return the algorithm/method to use */
return (best_msg_p->result_alg);
}