ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность

Support for new RMAPS rank mapping component

Просмотр исходного кода 
This commit was SVN r17860.
Этот коммит содержится в:
Lenny Verkhovsky 2008-03-18 09:39:07 +00:00
родитель 14c32f87d5
Коммит 647bce6d3e
9 изменённых файлов: 433 добавлений и 215 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

397
ompi/runtime/ompi_mpi_init.c
Просмотреть файл

@ -34,6 +34,7 @@
#include "opal/mca/maffinity/base/base.h"
#include "opal/runtime/opal_progress.h"
#include "opal/threads/threads.h"
#include "opal/util/argv.h"
#include "opal/util/show_help.h"
#include "opal/util/stacktrace.h"
#include "opal/util/num_procs.h"
@ -88,7 +89,8 @@
#endif
#include "ompi/runtime/ompi_cr.h"
static int slot_list_to_cpu_set(char *slot_str);
#include "orte/runtime/orte_globals.h"
/*
* Global variables and symbols for the MPI layer
*/
@ -228,6 +230,7 @@ int ompi_mpi_init(int argc, char **argv, int requested, int *provided)
bool timing = false;
int param, value;
struct timeval ompistart, ompistop;
char *slot_list = NULL;
#if 0
/* see comment below about sched_yield */
int num_processors;
@ -275,6 +278,12 @@ int ompi_mpi_init(int argc, char **argv, int requested, int *provided)
}
/* Setup process affinity */
if ( NULL != ( slot_list = getenv("slot_list"))) {
if (ORTE_SUCCESS != (ret = slot_list_to_cpu_set(slot_list))){
error = "ompi_mpi_init: error slot_list assigning";
goto error;
}
}
if (ompi_mpi_paffinity_alone) {
bool set = false;
@ -282,7 +291,6 @@ int ompi_mpi_init(int argc, char **argv, int requested, int *provided)
if (param >= 0) {
if (OMPI_SUCCESS == mca_base_param_lookup_int(param, &value)) {
if (value >= 0) {
opal_paffinity_base_cpu_set_t mpi_cpumask;
OPAL_PAFFINITY_CPU_ZERO(mpi_cpumask);
OPAL_PAFFINITY_CPU_SET(value,mpi_cpumask);
@ -295,11 +303,11 @@ int ompi_mpi_init(int argc, char **argv, int requested, int *provided)
char *vpid;
orte_util_convert_vpid_to_string(&vpid, ORTE_PROC_MY_NAME->vpid);
opal_show_help("help-mpi-runtime",
"mpi_init:startup:paffinity-unavailable",
"mpi_init:startup:paffinity-unavailable",
true, vpid);
free(vpid);
}
/* If we were able to set processor affinity, try setting
up memory affinity */
@ -766,3 +774,384 @@ int ompi_mpi_init(int argc, char **argv, int requested, int *provided)
return MPI_SUCCESS;
}
/**
* This function receives a slot string ant translate it to
* cpu_set (long bitmap) using the PLPA module.
*/
static int socket_to_cpu_set(char **socket_list, int socket_cnt)
{
int i;
char **range;
int range_cnt;
int lower_range, upper_range;
int processor_id, num_processors;
int max_processor_id;
int rc;
opal_paffinity_base_cpu_set_t cpumask;
if (OPAL_SUCCESS != (rc = opal_paffinity_base_get_processor_info(&num_processors, &max_processor_id))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_ZERO(cpumask);
for (i=0; i<socket_cnt; i++) {
if (0 == strcmp("*", socket_list[i])) {
for ( processor_id=0; processor_id<=max_processor_id; processor_id++) {
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if (OPAL_SUCCESS != ( rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %d runon cpu #%d (any socket)",(long)ORTE_PROC_MY_NAME->vpid, processor_id);
}
}
continue;
}
range = opal_argv_split(socket_list[i],'-');
range_cnt = opal_argv_count(range);
switch (range_cnt) {
case 1:
processor_id = atoi(range[0]);
if (max_processor_id < processor_id) {
opal_output(0, "ERROR !!! max_processor_id (%d) < processor_id(%d), modify rankfile and run again\n",max_processor_id, processor_id);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if (OPAL_SUCCESS != ( rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %d runs on cpu #%d", (long)ORTE_PROC_MY_NAME->vpid, processor_id);
}
break;
case 2:
lower_range = atoi(range[0]);
upper_range = atoi(range[1]);
if (max_processor_id < upper_range || lower_range >= upper_range ) {
opal_output(0,"Error !!! Check your boundaries %d < %d(max_cpu) < %d , modify rankfile and run again\n",lower_range, max_processor_id, upper_range);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
for (processor_id=lower_range; processor_id<=upper_range; processor_id++) {
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if (OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %d runs on cpu #%d (%d-%d)",
(long)ORTE_PROC_MY_NAME->vpid, processor_id, lower_range, upper_range);
}
}
break;
default:
opal_argv_free(range);
ORTE_ERROR_LOG(ORTE_ERROR);
return ORTE_ERROR;
}
opal_argv_free(range);
}
return ORTE_SUCCESS;
}
static int socket_core_to_cpu_set(char **socket_core_list, int socket_core_list_cnt)
{
int rc, i;
char **socket_core;
int socket_core_cnt;
char **range;
int range_cnt;
int lower_range, upper_range;
int socket, core, processor_id ;
int max_socket_num, max_core_num, max_processor_id;
int num_sockets, num_cores;
opal_paffinity_base_cpu_set_t cpumask;
socket_core = opal_argv_split (socket_core_list[0], ':');
socket_core_cnt = opal_argv_count(socket_core);
OPAL_PAFFINITY_CPU_ZERO(cpumask);
socket = atoi(socket_core[0]);
if ( OPAL_SUCCESS != ( rc = opal_paffinity_base_get_socket_info(&num_sockets, &max_socket_num))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if ( max_socket_num < socket) {
opal_output(0,"ERROR !!! socket(%d) > max_socket_num(%d), modify rankfile and run again", socket, max_socket_num);
return ORTE_ERROR;
}
if ( OPAL_SUCCESS != ( rc = opal_paffinity_base_get_core_info(socket, &num_cores, &max_core_num))) {
opal_output(0,"Error !!! Invalid socket number (%d) in rankfile, modify rankfile and run again\n", socket);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
if (0 == strcmp("*",socket_core[1])) {
for (core = 0; core <= max_core_num; core++) {
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_map_to_processor_id (socket, core, &processor_id))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if (OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %d runs on pair %d:%d (cpu #%d)",
(long)ORTE_PROC_MY_NAME->vpid, socket, core, processor_id);
}
}
} else {
range = opal_argv_split(socket_core[1], '-');
range_cnt = opal_argv_count(range);
switch (range_cnt) {
case 1:
core = atoi(range[0]);
if ( max_core_num < core ) {
opal_output(0,"Error !!! core(%d) > max_core (%d) on socket %d, modify rankfile and run again\n",
core, max_core_num, socket);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_map_to_processor_id (socket, core, &processor_id))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if (OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %d runs on pair %d:%d (cpu #%d)",
(long)ORTE_PROC_MY_NAME->vpid, socket, core, processor_id);
}
break;
case 2:
lower_range = atoi(range[0]);
upper_range = atoi(range[1]);
if ( 0 > lower_range || max_core_num < upper_range || lower_range >= upper_range ) {
opal_output(0,"Error !!! Check your boundaries %d < %d(max_core) < %d ,modify rankfile and run again\n",
lower_range, max_core_num, upper_range);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
for (core=lower_range; core<=upper_range; core++) {
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_map_to_processor_id (socket, core, &processor_id))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %d runs on pair %d:%d (cpu #%d)",
(long)ORTE_PROC_MY_NAME->vpid, socket, core, processor_id);
}
}
break;
default:
opal_argv_free(range);
opal_argv_free(socket_core);
ORTE_ERROR_LOG(ORTE_ERROR);
return ORTE_ERROR;
}
opal_argv_free(range);
opal_argv_free(socket_core);
}
for (i=1; i<socket_core_list_cnt; i++) {
socket_core = opal_argv_split (socket_core_list[i], ':');
socket_core_cnt = opal_argv_count(socket_core);
switch (socket_core_cnt) {
case 1:
range = opal_argv_split(socket_core[0], '-');
range_cnt = opal_argv_count(range);
switch (range_cnt) {
case 1:
core = atoi(range[0]);
/* use PLPA to construct the child->cpu_set */
if ( max_core_num < core ) {
opal_output(0,"Error !!! max_core(%d) < core(%d), modify rankfile and run again\n",max_core_num, core);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_map_to_processor_id (socket, core, &processor_id))) {
opal_output(0,"Error !!! Invalid socket : core pair ( #%d : %d), modify rankfile and run again\n",socket, core);
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %d runs on pair %d:%d (cpu #%d)",
(long)ORTE_PROC_MY_NAME->vpid, socket, core, processor_id);
}
break;
case 2:
lower_range = atoi(range[0]);
upper_range = atoi(range[1]);
if ( 0 > lower_range || max_core_num < upper_range || lower_range >= upper_range) {
opal_output(0,"Error !!! Check your boundaries %d < %d(max_core) < %d, modify rankfile and run again\n",
lower_range, max_core_num, upper_range);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
for (core=lower_range; core<=upper_range; core++) {
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_map_to_processor_id (socket, core, &processor_id))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %d runs on pair %d:%d (cpu #%d)",
(long)ORTE_PROC_MY_NAME->vpid, socket, core, processor_id);
}
}
break;
default:
opal_argv_free(range);
opal_argv_free(socket_core);
ORTE_ERROR_LOG(ORTE_ERROR);
return ORTE_ERROR;
}
opal_argv_free(range);
break;
case 2:
socket = atoi(socket_core[0]);
if (0 == strcmp("*",socket_core[1])) {
for (core=0; core<=max_core_num; core++) {
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_map_to_processor_id ( socket, core, &processor_id))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %d runs on pair %d:%d (cpu #%d)",
(long)ORTE_PROC_MY_NAME->vpid, socket, core, processor_id);
}
}
} else {
range = opal_argv_split(socket_core[1], '-');
range_cnt = opal_argv_count(range);
socket = atoi(socket_core[0]);
switch (range_cnt) {
case 1:
core = atoi(range[0]);
if ( max_core_num < core ) {
opal_output(0,"Error !!! max_core(%d) < core(%d), modify rankfile and run again\n", max_core_num, core);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_map_to_processor_id (socket, core, &processor_id))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %d runs on pair %d:%d (cpu #%d)",
(long)ORTE_PROC_MY_NAME->vpid, socket, core, processor_id);
}
break;
case 2:
lower_range = atoi(range[0]);
upper_range = atoi(range[1]);
if ( 0 > lower_range || max_core_num < upper_range || lower_range > upper_range) {
opal_output(0,"Error !!! Check your boundaries %d < %d(max_core) < %d, modify rankfile and run again\n",
lower_range, max_core_num, upper_range);
ORTE_ERROR_LOG(OPAL_ERR_BAD_PARAM);
return ORTE_ERROR;
}
for ( core = lower_range; core <= upper_range; core++) {
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_map_to_processor_id (socket, core, &processor_id))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
OPAL_PAFFINITY_CPU_SET(processor_id, cpumask);
if ( OPAL_SUCCESS != (rc = opal_paffinity_base_set(cpumask))) {
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
if (rmaps_rank_file_debug) {
opal_output(0,"rank %d runs on pair %d:%d (cpu #%d)",
(long)ORTE_PROC_MY_NAME->vpid, socket, core, processor_id);
}
}
break;
default:
opal_argv_free(range);
opal_argv_free(socket_core);
ORTE_ERROR_LOG(ORTE_ERROR);
return ORTE_ERROR;
}
opal_argv_free(range);
}
break;
default:
opal_argv_free(socket_core);
ORTE_ERROR_LOG(ORTE_ERROR);
return ORTE_ERROR;
}
opal_argv_free(socket_core);
}
return ORTE_SUCCESS;
}
static int slot_list_to_cpu_set(char *slot_str)
{
char **item;
char **socket_core;
orte_std_cntr_t item_cnt, socket_core_cnt;
int rc;
item = opal_argv_split (slot_str, ',');
item_cnt = opal_argv_count (item);
socket_core = opal_argv_split (item[0], ':');
socket_core_cnt = opal_argv_count(socket_core);
opal_argv_free(socket_core);
switch (socket_core_cnt) {
case 1:
if (ORTE_SUCCESS != (rc = socket_to_cpu_set(item, item_cnt))) {
opal_argv_free(item);
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
break;
case 2:
if (ORTE_SUCCESS != (rc = socket_core_to_cpu_set(item, item_cnt))) {
opal_argv_free(item);
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
break;
default:
opal_argv_free(item);
return ORTE_ERROR;
}
opal_argv_free(item);
return ORTE_SUCCESS;
}

19
ompi/runtime/ompi_mpi_params.c
Просмотреть файл

@ -49,6 +49,7 @@ bool ompi_debug_no_free_handles = false;
bool ompi_mpi_show_mca_params = false;
char *ompi_mpi_show_mca_params_file = NULL;
bool ompi_mpi_paffinity_alone = false;
bool rmaps_rank_file_debug = false;
bool ompi_mpi_abort_print_stack = false;
int ompi_mpi_abort_delay = 0;
bool ompi_mpi_keep_peer_hostnames = true;
@ -152,12 +153,28 @@ int ompi_mpi_register_params(void)
false, false,
(int) ompi_mpi_paffinity_alone, &value);
ompi_mpi_paffinity_alone = OPAL_INT_TO_BOOL(value);
if ( ompi_mpi_paffinity_alone ){
char *rank_file_path;
mca_base_param_reg_string_name("rmaps","rank_file_path",
"The path to the rank mapping file",
false, false, NULL, &rank_file_path);
if (NULL != rank_file_path) {
opal_output(0, "WARNING: Rankfile component can't be set with paffinity_alone, paffinity_alone set to 0");
ompi_mpi_paffinity_alone = 0;
}
}
mca_base_param_reg_int_name("mpi", "paffinity_processor",
"If set, pin this process to the processor number indicated by the value",
true, false,
-1, NULL);
mca_base_param_reg_int_name("rmaps", "rank_file_debug",
"If nonzero, prints binding to processors ",
false, false,
(int) rmaps_rank_file_debug, &value);
rmaps_rank_file_debug = OPAL_INT_TO_BOOL(value);
/* Do we want to save hostnames for debugging messages? This can
eat quite a bit of memory... */

5
ompi/runtime/params.h
Просмотреть файл

@ -100,6 +100,11 @@ OMPI_DECLSPEC extern char * ompi_mpi_show_mca_params_file;
*/
OMPI_DECLSPEC extern bool ompi_mpi_paffinity_alone;
/**
* If this value is true, we can check process binding to CPU
*/
OMPI_DECLSPEC extern bool rmaps_rank_file_debug;
/**
* Whether we should keep the string hostnames of all the MPI
* process peers around or not (eats up a good bit of memory).

217
orte/mca/odls/base/odls_base_default_fns.c
Просмотреть файл

@ -65,209 +65,6 @@
#include "orte/mca/odls/base/odls_private.h"
/**
* This function receives a slot string ant translate it to
* cpu_set (long bitmap) using the PLPA module.
* currently this function is doing nothig because PLPA is not
* part of the oprn MPI project.
*/
static int socket_to_cpu_set(char **socket_list, int socket_cnt, orte_odls_child_t *child)
{
int i, j;
char **range;
int range_cnt;
int lower_range, upper_range, socket_num;
for (i=0; i < socket_cnt; i++) {
if (0 == strcmp("*", socket_list[i])) {
/* use PLPA to construct the child->cpu_set */
opal_output(orte_odls_globals.output,"rank %d can run on any socket", child->name->vpid);
continue;
}
range = opal_argv_split(socket_list[i], '-');
range_cnt = opal_argv_count(range);
switch (range_cnt) {
case 1:
socket_num = atoi(range[0]);
/* use PLPA to construct the child->cpu_set */
opal_output(orte_odls_globals.output,"rank %d runs on socket #%d", child->name->vpid, socket_num);
break;
case 2:
lower_range = atoi(range[0]);
upper_range = atoi(range[1]);
for (j=lower_range; j<= upper_range; j++) {
socket_num = j;
/* use PLPA to construct the child->cpu_set */
opal_output(orte_odls_globals.output,"rank %d runs on socket #%d", child->name->vpid, socket_num);
}
break;
default:
opal_argv_free(range);
ORTE_ERROR_LOG(ORTE_ERROR);
return ORTE_ERROR;
}
opal_argv_free(range);
}
return ORTE_SUCCESS;
}
static int socket_core_to_cpu_set(char **socket_core_list, int socket_core_list_cnt, orte_odls_child_t *child)
{
int i, j;
char **socket_core;
int socket_core_cnt;
char **range;
int range_cnt;
int lower_range, upper_range;
int socket, core;
socket_core = opal_argv_split (socket_core_list[0], ':');
socket_core_cnt = opal_argv_count(socket_core);
socket = atoi(socket_core[0]);
if (0 == strcmp("*",socket_core[1])) {
/* use PLPA to construct the child->cpu_set */
opal_output(orte_odls_globals.output,"rank %d runs on socket #%d any core", child->name->vpid, socket);
}
else {
range = opal_argv_split(socket_core[1], '-');
range_cnt = opal_argv_count(range);
switch (range_cnt) {
case 1:
core = atoi(range[0]);
/* use PLPA to construct the child->cpu_set */
opal_output(orte_odls_globals.output,"rank %d runs on socket #%d core #%d", child->name->vpid, socket, core);
break;
case 2:
lower_range = atoi(range[0]);
upper_range = atoi(range[1]);
for (j=lower_range; j<= upper_range; j++) {
core = j;
/* use PLPA to construct the child->cpu_set */
opal_output(orte_odls_globals.output,"rank %d runs on socket #%d core #%d", child->name->vpid, socket, core);
}
break;
default:
opal_argv_free(range);
opal_argv_free(socket_core);
ORTE_ERROR_LOG(ORTE_ERROR);
return ORTE_ERROR;
}
opal_argv_free(range);
opal_argv_free(socket_core);
}
for (i=1; i < socket_core_list_cnt; i++) {
socket_core = opal_argv_split (socket_core_list[i], ':');
socket_core_cnt = opal_argv_count(socket_core);
switch (socket_core_cnt) {
case 1:
range = opal_argv_split(socket_core[0], '-');
range_cnt = opal_argv_count(range);
switch (range_cnt) {
case 1:
core = atoi(range[0]);
/* use PLPA to construct the child->cpu_set */
opal_output(orte_odls_globals.output,"and on core #%d", core);
break;
case 2:
lower_range = atoi(range[0]);
upper_range = atoi(range[1]);
for (j=lower_range; j<= upper_range; j++) {
core = j;
/* use PLPA to construct the child->cpu_set */
opal_output(orte_odls_globals.output,"and on core #%d", core);
}
break;
default:
opal_argv_free(range);
opal_argv_free(socket_core);
ORTE_ERROR_LOG(ORTE_ERROR);
return ORTE_ERROR;
}
opal_argv_free(range);
break;
case 2:
socket = atoi(socket_core[0]);
if (0 == strcmp("*",socket_core[1])) {
/* use PLPA to construct the child->cpu_set */
opal_output(orte_odls_globals.output,"and on socket #%d any core", socket);
}
else {
range = opal_argv_split(socket_core[1], '-');
range_cnt = opal_argv_count(range);
switch (range_cnt) {
case 1:
core = atoi(range[0]);
/* use PLPA to construct the child->cpu_set */
opal_output(orte_odls_globals.output,"and on socket #%d core #%d", socket, core);
break;
case 2:
lower_range = atoi(range[0]);
upper_range = atoi(range[1]);
for (j=lower_range; j<= upper_range; j++) {
core = j;
/* use PLPA to construct the child->cpu_set */
opal_output(orte_odls_globals.output,"and on socket #%d core #%d", socket, core);
}
break;
default:
opal_argv_free(range);
opal_argv_free(socket_core);
ORTE_ERROR_LOG(ORTE_ERROR);
return ORTE_ERROR;
}
opal_argv_free(range);
}
break;
default:
opal_argv_free(socket_core);
ORTE_ERROR_LOG(ORTE_ERROR);
return ORTE_ERROR;
}
opal_argv_free(socket_core);
}
return ORTE_SUCCESS;
}
static int slot_list_to_cpu_set(char *slot_str, orte_odls_child_t *child)
{
char **item;
char **socket_core;
int item_cnt, socket_core_cnt;
int rc;
item = opal_argv_split (slot_str, ',');
item_cnt = opal_argv_count (item);
socket_core = opal_argv_split (item[0], ':');
socket_core_cnt = opal_argv_count(socket_core);
opal_argv_free(socket_core);
switch (socket_core_cnt) {
case 1:
if (ORTE_SUCCESS != (rc = socket_to_cpu_set(item, item_cnt, child))) {
opal_argv_free(item);
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
break;
case 2:
if (ORTE_SUCCESS != (rc = socket_core_to_cpu_set(item, item_cnt, child))) {
opal_argv_free(item);
ORTE_ERROR_LOG(rc);
return ORTE_ERROR;
}
break;
default:
opal_argv_free(item);
return ORTE_ERROR;
}
opal_argv_free(item);
return ORTE_SUCCESS;
}
/* IT IS CRITICAL THAT ANY CHANGE IN THE ORDER OF THE INFO PACKED IN
* THIS FUNCTION BE REFLECTED IN THE CONSTRUCT_CHILD_LIST PARSER BELOW
*/
@ -580,11 +377,7 @@ int orte_odls_base_default_construct_child_list(opal_buffer_t *data,
child->app_idx = app_idx; /* save the index into the app_context objects */
child->local_rank = local_rank; /* save the local_rank */
if (NULL != slot_str) {
if (ORTE_SUCCESS != (rc = slot_list_to_cpu_set(slot_str, child))){
ORTE_ERROR_LOG(rc);
free(slot_str);
return rc;
}
child->slot_list = strdup(slot_str);
free(slot_str);
}
/* protect operation on the global list of children */
@ -1128,7 +921,6 @@ int orte_odls_base_default_launch_local(orte_jobid_t job,
*/
opal_setenv("OMPI_COMM_WORLD_LOCAL_RANK", value, true, &app->env);
free(value);
if (want_processor) {
param = mca_base_param_environ_variable("mpi", NULL,
"paffinity_processor");
@ -1142,7 +934,12 @@ int orte_odls_base_default_launch_local(orte_jobid_t job,
opal_unsetenv(param, &app->env);
free(param);
}
if ( NULL != child->slot_list ) {
opal_setenv("slot_list", child->slot_list, true, &app->env);
}else{
opal_unsetenv("slot_list", &app->env);
}
/* must unlock prior to fork to keep things clean in the
* event library
*/

2
orte/mca/odls/base/odls_base_open.c
Просмотреть файл

@ -67,12 +67,14 @@ static void orte_odls_child_constructor(orte_odls_child_t *ptr)
ptr->exit_code = 0;
ptr->cpu_set = 0xffffffff;
ptr->rml_uri = NULL;
ptr->slot_list = NULL;
}
static void orte_odls_child_destructor(orte_odls_child_t *ptr)
{
if (NULL != ptr->name) free(ptr->name);
if (NULL != ptr->rml_uri) free(ptr->rml_uri);
if (NULL != ptr->slot_list) free(ptr->slot_list);
}
OBJ_CLASS_INSTANCE(orte_odls_child_t,
opal_list_item_t,

1
orte/mca/odls/base/odls_private.h
Просмотреть файл

@ -61,6 +61,7 @@ typedef struct orte_odls_child_t {
orte_exit_code_t exit_code; /* process exit code */
unsigned long cpu_set;
char *rml_uri; /* contact info for this child */
char *slot_list; /* list of slots for this child */
} orte_odls_child_t;
ORTE_DECLSPEC OBJ_CLASS_DECLARATION(orte_odls_child_t);

5
orte/mca/rmaps/base/rmaps_base_support_fns.c
Просмотреть файл

@ -299,6 +299,11 @@ int orte_rmaps_base_claim_slot(orte_job_t *jdata,
proc->name.vpid = vpid;
proc->app_idx = app_idx;
OBJ_RETAIN(current_node); /* maintain accounting on object */
if ( NULL != current_node->slot_list) {
proc->slot_list = strdup(current_node->slot_list);
}
current_node->slot_list = NULL;
proc->node = current_node;
if (NULL != current_node->name) {
proc->nodename = strdup(current_node->name);

1
orte/runtime/orte_globals.h
Просмотреть файл

@ -156,6 +156,7 @@ typedef struct {
orte_std_cntr_t slots_max;
/** Username on this node, if specified */
char *username;
char *slot_list;
} orte_node_t;
ORTE_DECLSPEC OBJ_CLASS_DECLARATION(orte_node_t);

1
orte/runtime/orte_globals_class_instances.h
Просмотреть файл

@ -201,6 +201,7 @@ static void orte_node_construct(orte_node_t* node)
node->slots_alloc = 0;
node->slots_max = 0;
node->username = NULL;
node->slot_list = NULL;
}
static void orte_node_destruct(orte_node_t* node)