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

sm BTL initialization via modex.

Просмотр исходного кода 
This commit was SVN r26987.
Этот коммит содержится в:
Samuel Gutierrez 2012-08-10 19:51:41 +00:00
родитель e3e9b7345d
Коммит 2c80273246
8 изменённых файлов: 431 добавлений и 164 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

145
ompi/mca/btl/sm/btl_sm.c
Просмотреть файл

@ -46,6 +46,7 @@
#include "orte/util/proc_info.h"
#include "opal/datatype/opal_convertor.h"
#include "ompi/class/ompi_free_list.h"
#include "ompi/runtime/ompi_module_exchange.h"
#include "ompi/mca/btl/btl.h"
#include "ompi/mca/mpool/base/base.h"
#include "ompi/mca/mpool/sm/mpool_sm.h"
@ -111,7 +112,6 @@ mca_btl_sm_t mca_btl_sm = {
*/
#define OFFSET2ADDR(OFFSET, BASE) ((ptrdiff_t)(OFFSET) + (char*)(BASE))
static void *mpool_calloc(size_t nmemb, size_t size)
{
void *buf;
@ -127,17 +127,43 @@ static void *mpool_calloc(size_t nmemb, size_t size)
return buf;
}
static int sm_btl_first_time_init(mca_btl_sm_t *sm_btl, int n)
/*
* Returns a pointer to node rank zero. Returns NULL on error.
*/
static inline ompi_proc_t *
get_node_rank_zero_proc_ptr(ompi_proc_t **proc_world,
size_t proc_world_size)
{
size_t size, length, length_payload;
char *sm_ctl_file;
size_t num_local_procs = 0;
if (NULL == proc_world) {
return NULL;
}
/* sort the procs list and get a pointer to the lowest node rank */
if (OMPI_SUCCESS != mca_common_sm_local_proc_reorder(proc_world,
proc_world_size,
&num_local_procs)) {
opal_output(0, "mca_common_sm_local_proc_reorder failure! "
"Cannot continue.\n");
return NULL;
}
return proc_world[0];
}
static int
sm_btl_first_time_init(mca_btl_sm_t *sm_btl, int n)
{
size_t length, length_payload, proc_world_size = 0;
sm_fifo_t *my_fifos;
int my_mem_node, num_mem_nodes, i;
ompi_proc_t **procs;
size_t num_procs;
/* XXX SKG - malloc! */
mca_mpool_base_resources_t res;
mca_btl_sm_component_t* m = &mca_btl_sm_component;
ompi_proc_t **proc_world = NULL;
ompi_proc_t *proc_node_rank_zero = NULL;
mca_btl_sm_modex_t *modex = NULL;
size_t modex_size = 0;
/* Assume we don't have hwloc support and fill in dummy info */
mca_btl_sm_component.mem_node = my_mem_node = 0;
@ -191,8 +217,9 @@ static int sm_btl_first_time_init(mca_btl_sm_t *sm_btl, int n)
#endif
/* lookup shared memory pool */
mca_btl_sm_component.sm_mpools = (mca_mpool_base_module_t **) calloc(num_mem_nodes,
sizeof(mca_mpool_base_module_t*));
mca_btl_sm_component.sm_mpools =
(mca_mpool_base_module_t **)calloc(num_mem_nodes,
sizeof(mca_mpool_base_module_t *));
/* Disable memory binding, because each MPI process will claim
pages in the mpool for their local NUMA node */
@ -227,7 +254,7 @@ static int sm_btl_first_time_init(mca_btl_sm_t *sm_btl, int n)
}
res.size *= n;
/* now, create it */
/* SKG - mpool create - now, create it */
mca_btl_sm_component.sm_mpools[0] =
mca_mpool_base_module_create(mca_btl_sm_component.sm_mpool_name,
sm_btl, &res);
@ -248,34 +275,40 @@ static int sm_btl_first_time_init(mca_btl_sm_t *sm_btl, int n)
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* Allocate Shared Memory BTL process coordination
* data structure. This will reside in shared memory */
/* set file name */
if (asprintf(&sm_ctl_file, "%s"OPAL_PATH_SEP"shared_mem_btl_module.%s",
orte_process_info.job_session_dir,
orte_process_info.nodename) < 0) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* Pass in a data segment alignment of 0 to get no data
segment (only the shared control structure) */
size = sizeof(mca_common_sm_seg_header_t) +
n * (sizeof(sm_fifo_t*) + sizeof(char *) + sizeof(uint16_t)) + opal_cache_line_size;
procs = ompi_proc_world(&num_procs);
if (!(mca_btl_sm_component.sm_seg =
mca_common_sm_init(procs, num_procs, size, sm_ctl_file,
sizeof(mca_common_sm_seg_header_t),
opal_cache_line_size))) {
opal_output(0, "mca_btl_sm_add_procs: unable to create shared memory "
"BTL coordinating strucure :: size %lu \n",
(unsigned long)size);
free(procs);
free(sm_ctl_file);
/* now let's receive the modex info from node rank zero. */
if (NULL == (proc_world = ompi_proc_world(&proc_world_size))) {
opal_output(0, "ompi_proc_world failure! Cannot continue.\n");
return OMPI_ERROR;
}
free(procs);
free(sm_ctl_file);
if (NULL == (proc_node_rank_zero =
get_node_rank_zero_proc_ptr(proc_world, proc_world_size))) {
opal_output(0, "get_node_rank_zero_proc_ptr failure! "
"Cannot continue.\n");
free(proc_world);
return OMPI_ERROR;
}
if (OMPI_SUCCESS !=
ompi_modex_recv(&mca_btl_sm_component.super.btl_version,
proc_node_rank_zero,
(void **)&modex,
&modex_size)) {
opal_output(0, "sm_btl_first_time_init: ompi_modex_recv failure!\n");
free(proc_world);
return OMPI_ERROR;
}
if (NULL == (mca_btl_sm_component.sm_seg =
mca_common_sm_module_attach(&modex->sm_meta_buf,
sizeof(mca_common_sm_seg_header_t),
opal_cache_line_size))) {
/* don't have to detach here, because module_attach cleans up after
* itself on failure. */
opal_output(0, "sm_btl_first_time_init: "
"mca_common_sm_module_attach failure!\n");
free(proc_world);
return OMPI_ERROR;
}
free(proc_world);
free(modex);
/* check to make sure number of local procs is within the
* specified limits */
@ -374,6 +407,7 @@ static struct mca_btl_base_endpoint_t *
create_sm_endpoint(int local_proc, struct ompi_proc_t *proc)
{
struct mca_btl_base_endpoint_t *ep;
#if OMPI_ENABLE_PROGRESS_THREADS == 1
char path[PATH_MAX];
#endif
@ -401,22 +435,6 @@ create_sm_endpoint(int local_proc, struct ompi_proc_t *proc)
return ep;
}
static void calc_sm_max_procs(int n)
{
/* see if need to allocate space for extra procs */
if(0 > mca_btl_sm_component.sm_max_procs) {
/* no limit */
if(0 <= mca_btl_sm_component.sm_extra_procs) {
/* limit */
mca_btl_sm_component.sm_max_procs =
n + mca_btl_sm_component.sm_extra_procs;
} else {
/* no limit */
mca_btl_sm_component.sm_max_procs = 2 * n;
}
}
}
int mca_btl_sm_add_procs(
struct mca_btl_base_module_t* btl,
size_t nprocs,
@ -442,7 +460,7 @@ int mca_btl_sm_add_procs(
* and idetify procs that are on this host. Add procs on this
* host to shared memory reachbility list. Also, get number
* of local procs in the procs list. */
for(proc = 0; proc < (int32_t)nprocs; proc++) {
for (proc = 0; proc < (int32_t)nprocs; proc++) {
/* check to see if this proc can be reached via shmem (i.e.,
if they're on my local host and in my job) */
if (procs[proc]->proc_name.jobid != my_proc->proc_name.jobid ||
@ -477,18 +495,17 @@ int mca_btl_sm_add_procs(
goto CLEANUP;
/* make sure that my_smp_rank has been defined */
if(-1 == my_smp_rank) {
if (-1 == my_smp_rank) {
return_code = OMPI_ERROR;
goto CLEANUP;
}
calc_sm_max_procs(n_local_procs);
if (!sm_btl->btl_inited) {
return_code =
sm_btl_first_time_init(sm_btl, mca_btl_sm_component.sm_max_procs);
if(return_code != OMPI_SUCCESS)
if (return_code != OMPI_SUCCESS) {
goto CLEANUP;
}
}
/* set local proc's smp rank in the peers structure for
@ -524,13 +541,25 @@ int mca_btl_sm_add_procs(
/* Sync with other local procs. Force the FIFO initialization to always
* happens before the readers access it.
*/
opal_atomic_add_32( &mca_btl_sm_component.sm_seg->module_seg->seg_inited, 1);
opal_atomic_add_32(&mca_btl_sm_component.sm_seg->module_seg->seg_inited, 1);
while( n_local_procs >
mca_btl_sm_component.sm_seg->module_seg->seg_inited) {
opal_progress();
opal_atomic_rmb();
}
/* it is now safe to unlink the shared memory segment. only one process
* needs to do this, so just let smp rank zero take care of it. */
if (0 == my_smp_rank) {
if (OMPI_SUCCESS !=
mca_common_sm_module_unlink(mca_btl_sm_component.sm_seg)) {
/* it is "okay" if this fails at this point. we have gone this far,
* so just warn about the failure and continue. this is probably
* only triggered by a programming error. */
opal_output(0, "WARNING: common_sm_module_unlink failed.\n");
}
}
/* coordinate with other processes */
for(j = mca_btl_sm_component.num_smp_procs;
j < mca_btl_sm_component.num_smp_procs + n_local_procs; j++) {

13
ompi/mca/btl/sm/btl_sm.h
Просмотреть файл

@ -11,7 +11,7 @@
* All rights reserved.
* Copyright (c) 2006-2007 Voltaire. All rights reserved.
* Copyright (c) 2009-2010 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2010 Los Alamos National Security, LLC.
* Copyright (c) 2010-2012 Los Alamos National Security, LLC.
* All rights reserved.
* Copyright (c) 2010-2012 IBM Corporation. All rights reserved.
* $COPYRIGHT$
@ -42,6 +42,8 @@
#include "opal/util/bit_ops.h"
#include "opal/class/opal_free_list.h"
#include "opal/mca/shmem/shmem.h"
#include "ompi/mca/btl/btl.h"
#include "ompi/mca/common/sm/common_sm.h"
@ -121,6 +123,15 @@ typedef struct mca_btl_sm_mem_node_t {
mca_mpool_base_module_t* sm_mpool; /**< shared memory pool */
} mca_btl_sm_mem_node_t;
/**
* Shared Memory (SM) BTL modex.
*/
struct mca_btl_sm_modex_t {
opal_shmem_ds_t sm_meta_buf;
opal_shmem_ds_t sm_mpool_meta_buf;
};
typedef struct mca_btl_sm_modex_t mca_btl_sm_modex_t;
/**
* Shared Memory (SM) BTL module.
*/

198
ompi/mca/btl/sm/btl_sm_component.c
Просмотреть файл

@ -11,7 +11,7 @@
* All rights reserved.
* Copyright (c) 2006-2007 Voltaire. All rights reserved.
* Copyright (c) 2009-2010 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2010-2011 Los Alamos National Security, LLC.
* Copyright (c) 2010-2012 Los Alamos National Security, LLC.
* All rights reserved.
* Copyright (c) 2011 NVIDIA Corporation. All rights reserved.
* Copyright (c) 2010-2012 IBM Corporation. All rights reserved.
@ -42,15 +42,20 @@
#include <sys/stat.h> /* for mkfifo */
#endif /* HAVE_SYS_STAT_H */
#include "ompi/constants.h"
#include "opal/mca/event/event.h"
#include "opal/mca/base/mca_base_param.h"
#include "opal/mca/shmem/base/base.h"
#include "opal/mca/shmem/shmem.h"
#include "opal/util/bit_ops.h"
#include "opal/util/output.h"
#include "orte/util/proc_info.h"
#include "orte/util/show_help.h"
#include "orte/runtime/orte_globals.h"
#include "orte/util/proc_info.h"
#include "opal/mca/base/mca_base_param.h"
#include "ompi/constants.h"
#include "ompi/runtime/ompi_module_exchange.h"
#include "ompi/mca/mpool/base/base.h"
#include "ompi/mca/common/sm/common_sm.h"
#include "ompi/mca/btl/base/btl_base_error.h"
@ -351,52 +356,197 @@ CLEANUP:
return return_value;
}
/*
* Returns the number of processes on the node.
*/
static inline int
get_num_local_procs(void)
{
/* num_local_peers does not include us in
* its calculation, so adjust for that */
return (int)(1 + orte_process_info.num_local_peers);
}
static void
calc_sm_max_procs(int n)
{
/* see if need to allocate space for extra procs */
if (0 > mca_btl_sm_component.sm_max_procs) {
/* no limit */
if (0 <= mca_btl_sm_component.sm_extra_procs) {
/* limit */
mca_btl_sm_component.sm_max_procs =
n + mca_btl_sm_component.sm_extra_procs;
} else {
/* no limit */
mca_btl_sm_component.sm_max_procs = 2 * n;
}
}
}
/*
* Creates the shared-memory segments required for this BTL. One for the sm
* mpool and another for the shared memory store and populates *modex_buf_ptr.
*/
static int
populate_modex_bufp(mca_btl_sm_component_t *comp_ptr,
mca_btl_sm_modex_t *modex_buf_ptr)
{
int rc = OMPI_SUCCESS;
size_t size = 0;
char *sm_mpool_ctl_file = NULL;
char *sm_ctl_file = NULL;
/* first generate some unique paths for the shared-memory segments */
if (asprintf(&sm_mpool_ctl_file,
"%s"OPAL_PATH_SEP"shared_mem_btl_module.%s",
orte_process_info.job_session_dir,
orte_process_info.nodename) < 0) {
rc = OMPI_ERR_OUT_OF_RESOURCE;
goto out;
}
if (asprintf(&sm_ctl_file,
"%s"OPAL_PATH_SEP"shared_mem_btl_module.%s",
orte_process_info.job_session_dir,
orte_process_info.nodename) < 0) {
rc = OMPI_ERR_OUT_OF_RESOURCE;
goto out;
}
/* calculate the segment size */
size = sizeof(mca_common_sm_seg_header_t) +
comp_ptr->sm_max_procs * (sizeof(sm_fifo_t *) +
sizeof(char *) + sizeof(uint16_t)) + opal_cache_line_size;
/* create the things */
if (NULL == (comp_ptr->sm_seg =
mca_common_sm_module_create(size, sm_ctl_file,
sizeof(mca_common_sm_seg_header_t),
opal_cache_line_size))) {
opal_output(0, "mca_btl_sm_add_procs: unable to create shared memory "
"BTL coordinating strucure :: size %lu \n",
(unsigned long)size);
rc = OMPI_ERROR;
goto out;
}
/* now extract and store the shmem_ds info from the returned module */
if (OPAL_SUCCESS != opal_shmem_ds_copy(&(comp_ptr->sm_seg->shmem_ds),
&(modex_buf_ptr->sm_meta_buf))) {
rc = OMPI_ERROR;
goto out;
}
out:
if (NULL != sm_mpool_ctl_file) {
free(sm_mpool_ctl_file);
}
if (NULL != sm_ctl_file) {
free(sm_ctl_file);
}
return rc;
}
/*
* Creates information required for the sm modex and modex sends it.
*/
static int
send_modex(mca_btl_sm_component_t *comp_ptr)
{
int rc = OMPI_SUCCESS;
mca_btl_sm_modex_t *sm_modex = NULL;
if (NULL == (sm_modex = calloc(1, sizeof(*sm_modex)))) {
/* out of resources, so just bail. */
return OMPI_ERR_OUT_OF_RESOURCE;
}
if (OMPI_SUCCESS != (rc = populate_modex_bufp(comp_ptr, sm_modex))) {
opal_output(0, "send_modex: populate_modex_bufp failure!\n");
/* rc is set */
goto out;
}
/* send the modex */
rc = ompi_modex_send(&comp_ptr->super.btl_version, sm_modex,
sizeof(*sm_modex));
out:
if (NULL != sm_modex) {
free(sm_modex);
}
return rc;
}
/*
* SM component initialization
*/
static mca_btl_base_module_t** mca_btl_sm_component_init(
int *num_btls,
bool enable_progress_threads,
bool enable_mpi_threads)
static mca_btl_base_module_t **
mca_btl_sm_component_init(int *num_btls,
bool enable_progress_threads,
bool enable_mpi_threads)
{
mca_btl_base_module_t **btls = NULL;
orte_node_rank_t my_node_rank = ORTE_NODE_RANK_INVALID;
#if OMPI_BTL_SM_HAVE_KNEM
int rc;
#endif
*num_btls = 0;
/* if no session directory was created, then we cannot be used */
if (!orte_create_session_dirs) {
return NULL;
}
/* lookup/create shared memory pool only when used */
mca_btl_sm_component.sm_mpool = NULL;
mca_btl_sm_component.sm_mpool_base = NULL;
#if OMPI_ENABLE_PROGRESS_THREADS == 1
/* create a named pipe to receive events */
sprintf( mca_btl_sm_component.sm_fifo_path,
"%s"OPAL_PATH_SEP"sm_fifo.%lu", orte_process_info.job_session_dir,
(unsigned long)ORTE_PROC_MY_NAME->vpid );
if(mkfifo(mca_btl_sm_component.sm_fifo_path, 0660) < 0) {
opal_output(0, "mca_btl_sm_component_init: mkfifo failed with errno=%d\n",errno);
/* if no session directory was created, then we cannot be used */
/* SKG - this isn't true anymore. Some backing facilities don't require a
* file-backed store. Extend shmem to provide this info one day. */
if (!orte_create_session_dirs) {
return NULL;
}
mca_btl_sm_component.sm_fifo_fd = open(mca_btl_sm_component.sm_fifo_path, O_RDWR);
/* if we don't have locality information, then we cannot be used */
if (ORTE_NODE_RANK_INVALID ==
(my_node_rank = orte_process_info.my_node_rank)) {
orte_show_help("help-mpi-btl-sm.txt", "no locality", true);
return NULL;
}
/* calculate max procs so we can figure out how large to make the
* shared-memory segment. this routine sets component sm_max_procs. */
calc_sm_max_procs(get_num_local_procs());
/* let local rank 0 create the shared-memory segments and send shmem info */
if (0 == my_node_rank) {
if (OMPI_SUCCESS != send_modex(&mca_btl_sm_component)) {
return NULL;
}
}
#if OMPI_ENABLE_PROGRESS_THREADS == 1
/* create a named pipe to receive events */
sprintf(mca_btl_sm_component.sm_fifo_path,
"%s"OPAL_PATH_SEP"sm_fifo.%lu",
orte_process_info.job_session_dir,
(unsigned long)ORTE_PROC_MY_NAME->vpid);
if (mkfifo(mca_btl_sm_component.sm_fifo_path, 0660) < 0) {
opal_output(0, "mca_btl_sm_component_init: "
"mkfifo failed with errno=%d\n",errno);
return NULL;
}
mca_btl_sm_component.sm_fifo_fd = open(mca_btl_sm_component.sm_fifo_path,
O_RDWR);
if(mca_btl_sm_component.sm_fifo_fd < 0) {
opal_output(0, "mca_btl_sm_component_init: open(%s) failed with errno=%d\n",
opal_output(0, "mca_btl_sm_component_init: "
"open(%s) failed with errno=%d\n",
mca_btl_sm_component.sm_fifo_path, errno);
return NULL;
}
OBJ_CONSTRUCT(&mca_btl_sm_component.sm_fifo_thread, opal_thread_t);
mca_btl_sm_component.sm_fifo_thread.t_run = (opal_thread_fn_t) mca_btl_sm_component_event_thread;
mca_btl_sm_component.sm_fifo_thread.t_run =
(opal_thread_fn_t)mca_btl_sm_component_event_thread;
opal_thread_start(&mca_btl_sm_component.sm_fifo_thread);
#endif
mca_btl_sm_component.sm_btls = (mca_btl_sm_t **) malloc( mca_btl_sm_component.sm_max_btls * sizeof (mca_btl_sm_t *));
mca_btl_sm_component.sm_btls =
(mca_btl_sm_t **)malloc(mca_btl_sm_component.sm_max_btls *
sizeof(mca_btl_sm_t *));
if (NULL == mca_btl_sm_component.sm_btls) {
return NULL;
}

5
ompi/mca/btl/sm/help-mpi-btl-sm.txt
Просмотреть файл

@ -4,6 +4,8 @@
# of Tennessee Research Foundation. All rights
# reserved.
# Copyright (c) 2006-2010 Cisco Systems, Inc. All rights reserved.
# Copyright (c) 2012 Los Alamos National Security, LLC.
# All rights reserved.
# $COPYRIGHT$
#
# Additional copyrights may follow
@ -12,6 +14,9 @@
#
# This is the US/English help file for Open MPI's shared memory support.
#
[no locality]
WARNING: Missing locality info XXX FIXME SKG
#
[knem requested but not supported]
WARNING: Linux kernel knem support was requested for the shared memory
(sm) BTL, but it is not supported. Deactivating the shared memory

85
ompi/mca/common/sm/common_sm.c
Просмотреть файл

@ -42,6 +42,7 @@
#include "opal/align.h"
#include "opal/util/argv.h"
#include "opal/mca/shmem/shmem.h"
#if OPAL_ENABLE_FT_CR == 1
#include "opal/runtime/opal_cr.h"
#endif
@ -170,7 +171,7 @@ mca_common_sm_module_create(size_t size,
mca_common_sm_module_t *map = NULL;
opal_shmem_ds_t *seg_meta = NULL;
if (NULL == (seg_meta = (opal_shmem_ds_t *) malloc(sizeof(*seg_meta)))) {
if (NULL == (seg_meta = (opal_shmem_ds_t *)malloc(sizeof(*seg_meta)))) {
/* out of resources */
return NULL;
}
@ -197,33 +198,40 @@ mca_common_sm_module_attach(opal_shmem_ds_t *seg_meta,
size_t size_ctl_structure,
size_t data_seg_alignment)
{
mca_common_sm_module_t *map = NULL;
/* notice that size is 0 here. it really doesn't matter because size WILL
* NOT be used because this is an attach (first_call is false). */
map = attach_and_init(seg_meta, 0, size_ctl_structure,
data_seg_alignment, false);
return map;
return attach_and_init(seg_meta, 0, size_ctl_structure,
data_seg_alignment, false);
}
/* ////////////////////////////////////////////////////////////////////////// */
mca_common_sm_module_t *
mca_common_sm_init(ompi_proc_t **procs,
size_t num_procs,
size_t size,
char *file_name,
size_t size_ctl_structure,
size_t data_seg_alignment)
int
mca_common_sm_module_unlink(mca_common_sm_module_t *modp)
{
/* indicates whether or not i'm the lowest named process */
bool lowest_local_proc = false;
mca_common_sm_module_t *map = NULL;
ompi_proc_t *temp_proc = NULL;
bool found_lowest = false;
size_t num_local_procs = 0, p = 0;
opal_shmem_ds_t *seg_meta = NULL;
if (NULL == modp) {
return OMPI_ERROR;
}
if (OPAL_SUCCESS != opal_shmem_unlink(&modp->shmem_ds)) {
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}
/* ////////////////////////////////////////////////////////////////////////// */
int
mca_common_sm_local_proc_reorder(ompi_proc_t **procs,
size_t num_procs,
size_t *out_num_local_procs)
{
size_t num_local_procs = 0;
bool found_lowest = false;
ompi_proc_t *temp_proc = NULL;
size_t p;
if (NULL == out_num_local_procs || NULL == procs) {
return OMPI_ERR_BAD_PARAM;
}
/* o reorder procs array to have all the local procs at the beginning.
* o look for the local proc with the lowest name.
* o determine the number of local procs.
@ -240,8 +248,7 @@ mca_common_sm_init(ompi_proc_t **procs,
/* save this proc */
procs[num_local_procs] = procs[p];
/* if we have a new lowest, swap it with position 0
* so that procs[0] is always the lowest named proc
*/
* so that procs[0] is always the lowest named proc */
if (OPAL_VALUE2_GREATER ==
orte_util_compare_name_fields(ORTE_NS_CMP_ALL,
&(procs[p]->proc_name),
@ -257,6 +264,31 @@ mca_common_sm_init(ompi_proc_t **procs,
++num_local_procs;
}
}
*out_num_local_procs = num_local_procs;
return OMPI_SUCCESS;
}
/* ////////////////////////////////////////////////////////////////////////// */
mca_common_sm_module_t *
mca_common_sm_init(ompi_proc_t **procs,
size_t num_procs,
size_t size,
char *file_name,
size_t size_ctl_structure,
size_t data_seg_alignment)
{
/* indicates whether or not i'm the lowest named process */
bool lowest_local_proc = false;
mca_common_sm_module_t *map = NULL;
size_t num_local_procs = 0;
opal_shmem_ds_t *seg_meta = NULL;
if (OMPI_SUCCESS != mca_common_sm_local_proc_reorder(procs,
num_procs,
&num_local_procs)) {
return NULL;
}
/* if there is less than 2 local processes, there's nothing to do. */
if (num_local_procs < 2) {
@ -270,9 +302,9 @@ mca_common_sm_init(ompi_proc_t **procs,
/* determine whether or not i am the lowest local process */
lowest_local_proc =
(0 == orte_util_compare_name_fields(ORTE_NS_CMP_ALL,
ORTE_PROC_MY_NAME,
&(procs[0]->proc_name)));
(OPAL_EQUAL == orte_util_compare_name_fields(ORTE_NS_CMP_ALL,
ORTE_PROC_MY_NAME,
&(procs[0]->proc_name)));
/* figure out if i am the lowest rank in the group.
* if so, i will create the shared memory backing store
@ -434,4 +466,3 @@ mca_common_sm_fini(mca_common_sm_module_t *mca_common_sm_module)
}
return rc;
}

33
ompi/mca/common/sm/common_sm.h
Просмотреть файл

@ -10,7 +10,7 @@
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2009-2010 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2010-2011 Los Alamos National Security, LLC.
* Copyright (c) 2010-2012 Los Alamos National Security, LLC.
* All rights reserved.
* $COPYRIGHT$
*
@ -72,6 +72,19 @@ typedef struct mca_common_sm_module_t {
OBJ_CLASS_DECLARATION(mca_common_sm_module_t);
/**
* This routine reorders procs array to have all the local procs at the
* beginning and returns the number of local procs through out_num_local_procs.
* The proc with the lowest name is at the beginning of the reordered procs
* array.
*
* @returnvalue OMPI_SUCCESS on success, something else, otherwise.
*/
OMPI_DECLSPEC extern int
mca_common_sm_local_proc_reorder(ompi_proc_t **procs,
size_t num_procs,
size_t *out_num_local_procs);
/**
* This routine is used to create a shared memory segment (whether
* it's an mmaped file or a SYSV IPC segment). It is assumed that
@ -80,7 +93,7 @@ OBJ_CLASS_DECLARATION(mca_common_sm_module_t);
* @returnvalue pointer to control structure at head of shared memory segment.
* Returns NULL if an error occurred.
*/
mca_common_sm_module_t *
OMPI_DECLSPEC extern mca_common_sm_module_t *
mca_common_sm_module_create(size_t size,
char *file_name,
size_t size_ctl_structure,
@ -96,11 +109,22 @@ mca_common_sm_module_create(size_t size,
* @returnvalue pointer to control structure at head of shared memory segment.
* Returns NULL if an error occurred.
*/
mca_common_sm_module_t *
OMPI_DECLSPEC extern mca_common_sm_module_t *
mca_common_sm_module_attach(opal_shmem_ds_t *seg_meta,
size_t size_ctl_structure,
size_t data_seg_alignment);
/**
* A thin wrapper around opal_shmem_unlink.
*
* @ modp points to an initialized mca_common_sm_module_t.
*
* @returnvalue OMPI_SUCCESS if the operation completed successfully,
* OMPI_ERROR otherwise.
*/
OMPI_DECLSPEC extern int
mca_common_sm_module_unlink(mca_common_sm_module_t *modp);
/**
* This routine is used to set up a shared memory segment (whether
* it's an mmaped file or a SYSV IPC segment). It is assumed that
@ -164,7 +188,7 @@ mca_common_sm_init_group(ompi_group_t *group,
*/
OMPI_DECLSPEC extern void *
mca_common_sm_seg_alloc(struct mca_mpool_base_module_t *mpool,
size_t* size,
size_t *size,
mca_mpool_base_registration_t **registration);
/**
@ -189,4 +213,3 @@ OMPI_DECLSPEC extern mca_common_sm_module_t *mca_common_sm_module;
END_C_DECLS
#endif /* _COMMON_SM_H_ */

17
ompi/mca/mpool/sm/mpool_sm.h
Просмотреть файл

@ -11,7 +11,7 @@
* All rights reserved.
* Copyright (c) 2007 Sun Microsystems, Inc. All rights reserved.
* Copyright (c) 2009 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2010 Los Alamos National Security, LLC.
* Copyright (c) 2010-2012 Los Alamos National Security, LLC.
* All rights reserved.
* $COPYRIGHT$
*
@ -28,6 +28,7 @@
#include "ompi_config.h"
#include "opal/mca/event/event.h"
#include "opal/mca/shmem/shmem.h"
#include "ompi/mca/common/sm/common_sm.h"
#include "ompi/mca/mpool/mpool.h"
@ -36,17 +37,19 @@
BEGIN_C_DECLS
struct mca_mpool_sm_component_t {
mca_mpool_base_component_t super;
/* mca_allocator_base_module_t* sm_allocator; */
char* sm_allocator_name;
int verbose;
/* struct mca_mpool_sm_mmap_t *sm_mmap; */
mca_mpool_base_component_t super;
/* mca_allocator_base_module_t* sm_allocator; */
char *sm_allocator_name;
int verbose;
/* struct mca_mpool_sm_mmap_t *sm_mmap; */
};
typedef struct mca_mpool_sm_component_t mca_mpool_sm_component_t;
typedef struct mca_mpool_base_resources_t {
size_t size;
int32_t mem_node;
/* backing store metadata */
opal_shmem_ds_t bs_meta_buf;
} mca_mpool_base_resources_t;
OMPI_MODULE_DECLSPEC extern mca_mpool_sm_component_t mca_mpool_sm_component;
@ -54,7 +57,7 @@ OMPI_MODULE_DECLSPEC extern mca_mpool_sm_component_t mca_mpool_sm_component;
typedef struct mca_mpool_sm_module_t {
mca_mpool_base_module_t super;
long sm_size;
mca_allocator_base_module_t * sm_allocator;
mca_allocator_base_module_t *sm_allocator;
struct mca_mpool_sm_mmap_t *sm_mmap;
mca_common_sm_module_t *sm_common_module;
int32_t mem_node;

95
ompi/mca/mpool/sm/mpool_sm_component.c
Просмотреть файл

@ -11,7 +11,7 @@
* All rights reserved.
* Copyright (c) 2007-2009 Sun Microsystems, Inc. All rights reserved.
* Copyright (c) 2008-2009 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2010 Los Alamos National Security, LLC.
* Copyright (c) 2010-2012 Los Alamos National Security, LLC.
* All rights reserved.
* $COPYRIGHT$
*
@ -45,10 +45,14 @@
/*
* Local functions
*/
static int mca_mpool_sm_open(void);
static int mca_mpool_sm_close( void );
static mca_mpool_base_module_t* mca_mpool_sm_init(
struct mca_mpool_base_resources_t* resources);
static int
mca_mpool_sm_open(void);
static int
mca_mpool_sm_close(void);
static mca_mpool_base_module_t *
mca_mpool_sm_init(struct mca_mpool_base_resources_t* resources);
mca_mpool_sm_component_t mca_mpool_sm_component = {
{
@ -90,8 +94,8 @@ static int mca_mpool_sm_open(void)
/* register SM component parameters */
mca_base_param_reg_string(&mca_mpool_sm_component.super.mpool_version,
"allocator",
"Name of allocator component to use with sm mpool",
false, false,
"Name of allocator component "
"to use with sm mpool", false, false,
"bucket",
&mca_mpool_sm_component.sm_allocator_name);
@ -100,18 +104,18 @@ static int mca_mpool_sm_open(void)
* to be set up to 2GB-1 for 32 bit and much greater for 64 bit. */
asprintf(&size_str, "%ld", default_min);
mca_base_param_reg_string(&mca_mpool_sm_component.super.mpool_version,
"min_size",
"Minimum size of the sm mpool shared memory file",
false, false, size_str, &min_size_param);
"min_size",
"Minimum size of the sm mpool shared memory file",
false, false, size_str, &min_size_param);
free(size_str);
mca_base_param_reg_int(&mca_mpool_sm_component.super.mpool_version,
"verbose",
"Enable verbose output for mpool sm component",
false, false, 0, &value);
"verbose",
"Enable verbose output for mpool sm component",
false, false, 0, &value);
if (value != 0) {
mca_mpool_sm_component.verbose = opal_output_open(NULL);
mca_mpool_sm_component.verbose = opal_output_open(NULL);
} else {
mca_mpool_sm_component.verbose = -1;
mca_mpool_sm_component.verbose = -1;
}
return OMPI_SUCCESS;
@ -128,41 +132,46 @@ static int mca_mpool_sm_close( void )
return OMPI_SUCCESS;
}
static mca_mpool_base_module_t* mca_mpool_sm_init(
struct mca_mpool_base_resources_t* resources)
static mca_mpool_base_module_t *
mca_mpool_sm_init(struct mca_mpool_base_resources_t *resources)
{
char *file_name;
int len;
mca_mpool_sm_module_t* mpool_module;
mca_mpool_sm_module_t *mpool_module;
mca_allocator_base_component_t* allocator_component;
long min_size;
ompi_proc_t **procs;
size_t num_all_procs, i, num_local_procs = 0;
/* README: this needs to change if procs in different jobs (even
spawned ones) are to talk using shared memory */
procs = ompi_proc_world(&num_all_procs);
* spawned ones) are to talk using shared memory */
if (NULL == (procs = ompi_proc_world(&num_all_procs))) {
/* out of resources, so just bail */
return NULL;
}
for (i = 0 ; i < num_all_procs ; ++i) {
if (OPAL_PROC_ON_LOCAL_NODE(procs[i]->proc_flags)) {
num_local_procs++;
}
}
/* parse the min size and validate it */
/* if other parameters are added, absolutely necessary to reset errno each time */
/* if other parameters are added, absolutely
* necessary to reset errno each time */
errno = 0;
min_size = strtol(min_size_param, (char **)NULL, 10);
if (errno == ERANGE) {
opal_output(0, "mca_mpool_sm_init: min_size overflows! set to default (%ld)", default_min);
opal_output(0, "mca_mpool_sm_init: min_size overflows! "
"set to default (%ld)", default_min);
min_size = default_min;
} else if (errno == EINVAL) {
opal_output(0, "mca_mpool_sm_init: invalid min_size entered. set it to (%ld)", default_min);
opal_output(0, "mca_mpool_sm_init: invalid min_size entered. "
"set it to (%ld)", default_min);
min_size = default_min;
}
/* Make a new mpool module */
mpool_module =
(mca_mpool_sm_module_t*)malloc(sizeof(mca_mpool_sm_module_t));
(mca_mpool_sm_module_t *)malloc(sizeof(mca_mpool_sm_module_t));
mca_mpool_sm_module_init(mpool_module);
/* set sm_size */
@ -180,16 +189,22 @@ static mca_mpool_base_module_t* mca_mpool_sm_init(
mca_mpool_sm_component.sm_allocator_name);
/* if specified allocator cannot be loaded - look for an alternative */
if(NULL == allocator_component) {
if(opal_list_get_size(&mca_allocator_base_components) == 0) {
mca_base_component_list_item_t* item = (mca_base_component_list_item_t*)
if (NULL == allocator_component) {
if (opal_list_get_size(&mca_allocator_base_components) == 0) {
mca_base_component_list_item_t *item =
(mca_base_component_list_item_t *)
opal_list_get_first(&mca_allocator_base_components);
allocator_component = (mca_allocator_base_component_t*)item->cli_component;
opal_output(0, "mca_mpool_sm_init: unable to locate allocator: %s - using %s\n",
mca_mpool_sm_component.sm_allocator_name, allocator_component->allocator_version.mca_component_name);
allocator_component =
(mca_allocator_base_component_t *)item->cli_component;
opal_output(
0, "mca_mpool_sm_init: "
"unable to locate allocator: %s - using %s\n",
mca_mpool_sm_component.sm_allocator_name,
allocator_component->allocator_version.mca_component_name);
} else {
opal_output(0, "mca_mpool_sm_init: unable to locate allocator: %s\n",
mca_mpool_sm_component.sm_allocator_name);
opal_output(0, "mca_mpool_sm_init: "
"unable to locate allocator: %s\n",
mca_mpool_sm_component.sm_allocator_name);
free(procs);
return NULL;
}
@ -198,10 +213,10 @@ static mca_mpool_base_module_t* mca_mpool_sm_init(
mpool_module->mem_node = resources->mem_node;
/* create initial shared memory mapping */
len = asprintf( &file_name, "%s"OPAL_PATH_SEP"shared_mem_pool.%s",
orte_process_info.job_session_dir,
orte_process_info.nodename );
if ( 0 > len ) {
len = asprintf(&file_name, "%s"OPAL_PATH_SEP"shared_mem_pool.%s",
orte_process_info.job_session_dir,
orte_process_info.nodename);
if (0 > len) {
free(mpool_module);
free(procs);
return NULL;
@ -216,8 +231,8 @@ static mca_mpool_base_module_t* mca_mpool_sm_init(
mpool_module->sm_size,
file_name,
sizeof(mca_common_sm_module_t), 8))) {
opal_output(mca_mpool_sm_component.verbose,
"mca_mpool_sm_init: unable to create shared memory mapping (%s)", file_name);
opal_output(mca_mpool_sm_component.verbose, "mca_mpool_sm_init: "
"unable to create shared memory mapping (%s)", file_name);
free(file_name);
free(mpool_module);
free(procs);
@ -231,7 +246,7 @@ static mca_mpool_base_module_t* mca_mpool_sm_init(
allocator_component->allocator_init(true,
mca_common_sm_seg_alloc,
NULL, &(mpool_module->super));
if(NULL == mpool_module->sm_allocator) {
if (NULL == mpool_module->sm_allocator) {
opal_output(0, "mca_mpool_sm_init: unable to initialize allocator");
free(mpool_module);
return NULL;