ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
958e29f929
openmpi/ompi/mca/bml/r2/bml_r2.c
Ralph Castain 0e433eaa78 Silence warning
2016-07-11 19:43:02 -07:00

987 строки
35 KiB
C
Исходник Ответственный История

/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2004-2007 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2016 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) 2004-2006 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2007-2015 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2008-2016 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2013 Intel, Inc. All rights reserved
* Copyright (c) 2014 NVIDIA Corporation. All rights reserved.
* Copyright (c) 2014 Research Organization for Information Science
* and Technology (RIST). All rights reserved.
* Copyright (c) 2014-2015 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2016 Intel, Inc. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include <stdlib.h>
#include <string.h>
#include "opal/class/opal_bitmap.h"
#include "opal/util/argv.h"
#include "opal/util/show_help.h"
#include "opal/util/output.h"
#include "ompi/mca/bml/bml.h"
#include "ompi/mca/bml/base/base.h"
#include "opal/mca/btl/btl.h"
#include "opal/mca/btl/base/base.h"
#include "ompi/mca/bml/base/bml_base_btl.h"
#include "bml_r2.h"
#include "ompi/proc/proc.h"
extern mca_bml_base_component_t mca_bml_r2_component;
/* Names of all the BTL components that this BML is aware of */
static char *btl_names = NULL;
static int btl_exclusivity_compare(const void* arg1, const void* arg2)
{
mca_btl_base_module_t* btl1 = *(struct mca_btl_base_module_t**)arg1;
mca_btl_base_module_t* btl2 = *(struct mca_btl_base_module_t**)arg2;
if( btl1->btl_exclusivity > btl2->btl_exclusivity ) {
return -1;
} else if (btl1->btl_exclusivity == btl2->btl_exclusivity ) {
return 0;
} else {
return 1;
}
}
static int mca_bml_r2_add_btls( void )
{
int i;
opal_list_t *btls = NULL;
mca_btl_base_selected_module_t* selected_btl;
size_t num_btls = 0;
char **btl_names_argv = NULL;
if(true == mca_bml_r2.btls_added) {
return OMPI_SUCCESS;
}
/* build an array of r2s and r2 modules */
btls = &mca_btl_base_modules_initialized;
num_btls = opal_list_get_size(btls);
mca_bml_r2.num_btl_modules = 0;
mca_bml_r2.num_btl_progress = 0;
mca_bml_r2.btl_modules = (mca_btl_base_module_t **)malloc(sizeof(mca_btl_base_module_t*) * num_btls);
mca_bml_r2.btl_progress = (mca_btl_base_component_progress_fn_t*)malloc(sizeof(mca_btl_base_component_progress_fn_t) * num_btls);
if (NULL == mca_bml_r2.btl_modules ||
NULL == mca_bml_r2.btl_progress) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
OPAL_LIST_FOREACH(selected_btl, btls, mca_btl_base_selected_module_t) {
mca_btl_base_module_t *btl = selected_btl->btl_module;
mca_bml_r2.btl_modules[mca_bml_r2.num_btl_modules++] = btl;
for (i = 0; NULL != btl_names_argv && NULL != btl_names_argv[i]; ++i) {
if (0 ==
strcmp(btl_names_argv[i],
btl->btl_component->btl_version.mca_component_name)) {
break;
}
}
if (NULL == btl_names_argv || NULL == btl_names_argv[i]) {
opal_argv_append_nosize(&btl_names_argv,
btl->btl_component->btl_version.mca_component_name);
}
}
if (NULL != btl_names_argv) {
btl_names = opal_argv_join(btl_names_argv, ' ');
opal_argv_free(btl_names_argv);
} else {
btl_names = strdup("no devices available");
}
/* sort r2 list by exclusivity */
qsort(mca_bml_r2.btl_modules,
mca_bml_r2.num_btl_modules,
sizeof(struct mca_btl_base_module_t*),
btl_exclusivity_compare);
mca_bml_r2.btls_added = true;
return OMPI_SUCCESS;
}
static int btl_bandwidth_compare(const void *v1, const void *v2)
{
mca_bml_base_btl_t *b1 = (mca_bml_base_btl_t*)v1,
*b2 = (mca_bml_base_btl_t*)v2;
return b2->btl->btl_bandwidth - b1->btl->btl_bandwidth;
}
static void mca_bml_r2_calculate_bandwidth_latency (mca_bml_base_btl_array_t *btl_array, double *total_bandwidth, uint32_t *latency)
{
const size_t array_length = mca_bml_base_btl_array_get_size (btl_array);
*latency = UINT_MAX;
*total_bandwidth = 0.;
for (size_t i = 0 ; i < array_length ; ++i) {
mca_bml_base_btl_t *bml_btl = mca_bml_base_btl_array_get_index (btl_array, i);
mca_btl_base_module_t *btl = bml_btl->btl;
*total_bandwidth += btl->btl_bandwidth;
if (btl->btl_latency < *latency) {
*latency = btl->btl_latency;
}
}
}
static mca_bml_base_endpoint_t *mca_bml_r2_allocate_endpoint (ompi_proc_t *proc) {
mca_bml_base_endpoint_t *bml_endpoint;
/* allocate bml specific proc data */
bml_endpoint = OBJ_NEW(mca_bml_base_endpoint_t);
if (NULL == bml_endpoint) {
opal_output(0, "%s: unable to allocate resources", __func__);
return NULL;
}
/* preallocate space in array for max number of r2s */
mca_bml_base_btl_array_reserve(&bml_endpoint->btl_eager, mca_bml_r2.num_btl_modules);
mca_bml_base_btl_array_reserve(&bml_endpoint->btl_send, mca_bml_r2.num_btl_modules);
mca_bml_base_btl_array_reserve(&bml_endpoint->btl_rdma, mca_bml_r2.num_btl_modules);
bml_endpoint->btl_max_send_size = -1;
bml_endpoint->btl_proc = proc;
bml_endpoint->btl_flags_or = 0;
return bml_endpoint;
}
static void mca_bml_r2_register_progress (mca_btl_base_module_t *btl, bool hp)
{
if (NULL != btl->btl_component->btl_progress) {
bool found = false;
size_t p;
for (p = 0 ; p < mca_bml_r2.num_btl_progress ; ++p) {
if(mca_bml_r2.btl_progress[p] == btl->btl_component->btl_progress) {
found = true;
break;
}
}
if (found == false || hp) {
if (found == false) {
mca_bml_r2.btl_progress[mca_bml_r2.num_btl_progress++] =
btl->btl_component->btl_progress;
}
if (hp) {
opal_progress_register (btl->btl_component->btl_progress);
} else {
opal_progress_register_lp (btl->btl_component->btl_progress);
}
}
}
}
static int mca_bml_r2_endpoint_add_btl (struct ompi_proc_t *proc, mca_bml_base_endpoint_t *bml_endpoint,
mca_btl_base_module_t *btl, struct mca_btl_base_endpoint_t *btl_endpoint)
{
mca_bml_base_btl_t* bml_btl = NULL;
int btl_flags = btl->btl_flags;
bool btl_in_use = false;
size_t size;
/* NTH: these flags should have been sanitized by the btl. Once that is verified these
* checks can be safely removed. */
if ((btl_flags & MCA_BTL_FLAGS_PUT) && (NULL == btl->btl_put)) {
opal_output(0, "%s: The PUT flag is specified for"
" the %s BTL without any PUT function attached. Discard the flag !",
__func__,
btl->btl_component->btl_version.mca_component_name);
btl_flags ^= MCA_BTL_FLAGS_PUT;
}
if ((btl_flags & MCA_BTL_FLAGS_GET) && (NULL == btl->btl_get)) {
opal_output(0, "%s: The GET flag is specified for"
" the %s BTL without any GET function attached. Discard the flag !",
__func__, btl->btl_component->btl_version.mca_component_name);
btl_flags ^= MCA_BTL_FLAGS_GET;
}
if ((btl_flags & (MCA_BTL_FLAGS_PUT | MCA_BTL_FLAGS_GET | MCA_BTL_FLAGS_SEND)) == 0) {
/* If no protocol specified, we have 2 choices: we ignore the BTL
* as we don't know which protocl to use, or we suppose that all
* BTLs support the send protocol. This is really a btl error as
* these flags should have been sanitized by the btl. */
btl_flags |= MCA_BTL_FLAGS_SEND;
}
if (btl_flags & MCA_BTL_FLAGS_SEND) {
/* dont allow an additional BTL with a lower exclusivity ranking */
size = mca_bml_base_btl_array_get_size (&bml_endpoint->btl_send);
bml_btl = mca_bml_base_btl_array_get_index (&bml_endpoint->btl_send, size - 1);
if (!bml_btl || bml_btl->btl->btl_exclusivity <= btl->btl_exclusivity) {
/* this btl has higher exclusivity than an existing btl or none exists */
opal_output_verbose(1, opal_btl_base_framework.framework_output,
"mca: bml: Using %s btl for send to %s on node %s",
btl->btl_component->btl_version.mca_component_name,
OMPI_NAME_PRINT(&proc->super.proc_name),
proc->super.proc_hostname);
/* cache the endpoint on the proc */
if (NULL == bml_btl || (bml_btl->btl->btl_exclusivity <= btl->btl_exclusivity)) {
bml_btl = mca_bml_base_btl_array_insert (&bml_endpoint->btl_send);
bml_btl->btl = btl;
bml_btl->btl_endpoint = btl_endpoint;
bml_btl->btl_weight = 0;
bml_btl->btl_flags = btl_flags;
/**
* calculate the bitwise OR of the btl flags
*/
bml_endpoint->btl_flags_or |= bml_btl->btl_flags;
} else {
opal_output_verbose(20, opal_btl_base_framework.framework_output,
"mca: bml: Not using %s btl for send to %s on node %s "
"because %s btl has higher exclusivity (%d > %d)",
btl->btl_component->btl_version.mca_component_name,
OMPI_NAME_PRINT(&proc->super.proc_name), proc->super.proc_hostname,
bml_btl->btl->btl_component->btl_version.mca_component_name,
bml_btl->btl->btl_exclusivity,
btl->btl_exclusivity);
}
btl_in_use = true;
}
}
/* always add rdma endpoints if they support full rdma */
if (((btl_in_use && (btl_flags & MCA_BTL_FLAGS_RDMA)) ||
(btl_flags & (MCA_BTL_FLAGS_RDMA | MCA_BTL_FLAGS_ATOMIC_FOPS)) == (MCA_BTL_FLAGS_RDMA | MCA_BTL_FLAGS_ATOMIC_FOPS)) &&
!((proc->super.proc_arch != ompi_proc_local_proc->super.proc_arch) &&
(0 == (btl->btl_flags & MCA_BTL_FLAGS_HETEROGENEOUS_RDMA)))) {
mca_bml_base_btl_t *bml_btl_rdma = mca_bml_base_btl_array_insert(&bml_endpoint->btl_rdma);
bml_btl_rdma->btl = btl;
bml_btl_rdma->btl_endpoint = btl_endpoint;
bml_btl_rdma->btl_weight = 0;
bml_btl_rdma->btl_flags = btl_flags;
if (bml_endpoint->btl_pipeline_send_length < btl->btl_rdma_pipeline_send_length) {
bml_endpoint->btl_pipeline_send_length = btl->btl_rdma_pipeline_send_length;
}
if (bml_endpoint->btl_send_limit < btl->btl_min_rdma_pipeline_size) {
bml_endpoint->btl_send_limit = btl->btl_min_rdma_pipeline_size;
}
btl_in_use = true;
}
return btl_in_use ? OMPI_SUCCESS : OMPI_ERR_NOT_AVAILABLE;
}
static void mca_bml_r2_compute_endpoint_metrics (mca_bml_base_endpoint_t *bml_endpoint)
{
double total_bandwidth = 0;
uint32_t latency;
size_t n_send, n_rdma;
/* (1) determine the total bandwidth available across all btls
* note that we need to do this here, as we may already have btls configured
* (2) determine the highest priority ranking for latency
* (3) compute the maximum amount of bytes that can be send without any
* weighting. Once the left over is smaller than this number we will
* start using the weight to compute the correct amount.
*/
n_send = mca_bml_base_btl_array_get_size (&bml_endpoint->btl_send);
n_rdma = mca_bml_base_btl_array_get_size (&bml_endpoint->btl_rdma);
/* sort BTLs in descending order according to bandwidth value */
qsort (bml_endpoint->btl_send.bml_btls, n_send,
sizeof(mca_bml_base_btl_t), btl_bandwidth_compare);
bml_endpoint->btl_rdma_index = 0;
mca_bml_r2_calculate_bandwidth_latency (&bml_endpoint->btl_send, &total_bandwidth, &latency);
/* (1) set the weight of each btl as a percentage of overall bandwidth
* (2) copy all btl instances at the highest priority ranking into the
* list of btls used for first fragments
*/
for (size_t n_index = 0 ; n_index < n_send ; ++n_index) {
mca_bml_base_btl_t *bml_btl =
mca_bml_base_btl_array_get_index(&bml_endpoint->btl_send, n_index);
mca_btl_base_module_t *btl = bml_btl->btl;
/* compute weighting factor for this r2 */
if(btl->btl_bandwidth > 0) {
bml_btl->btl_weight = (float)(btl->btl_bandwidth / total_bandwidth);
} else {
bml_btl->btl_weight = (float)(1.0 / n_send);
}
/* check to see if this r2 is already in the array of r2s
* used for first fragments - if not add it.
*/
if(btl->btl_latency == latency) {
mca_bml_base_btl_t* bml_btl_new =
mca_bml_base_btl_array_insert(&bml_endpoint->btl_eager);
*bml_btl_new = *bml_btl;
}
/* set endpoint max send size as min of available btls */
if (bml_endpoint->btl_max_send_size > btl->btl_max_send_size)
bml_endpoint->btl_max_send_size = btl->btl_max_send_size;
}
/* sort BTLs in descending order according to bandwidth value */
qsort(bml_endpoint->btl_rdma.bml_btls, n_rdma,
sizeof(mca_bml_base_btl_t), btl_bandwidth_compare);
mca_bml_r2_calculate_bandwidth_latency (&bml_endpoint->btl_rdma, &total_bandwidth, &latency);
/* set rdma btl weights */
for (size_t n_index = 0 ; n_index < n_rdma ; ++n_index) {
mca_bml_base_btl_t *bml_btl =
mca_bml_base_btl_array_get_index(&bml_endpoint->btl_rdma, n_index);
/* compute weighting factor for this r2 */
if (bml_btl->btl->btl_bandwidth > 0.0) {
bml_btl->btl_weight = (float)(bml_btl->btl->btl_bandwidth / total_bandwidth);
} else {
bml_btl->btl_weight = (float)(1.0 / n_rdma);
}
}
}
static int mca_bml_r2_add_proc (struct ompi_proc_t *proc)
{
mca_bml_base_endpoint_t *bml_endpoint;
/* at least one btl is in use */
bool btl_in_use = false;
int rc;
if (OPAL_UNLIKELY(NULL == proc)) {
return OMPI_ERR_BAD_PARAM;
}
/* check if this endpoint is already set up */
if (NULL != proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML]) {
OBJ_RETAIN(proc);
return OMPI_SUCCESS;
}
/* add btls if not already done */
if (OMPI_SUCCESS != (rc = mca_bml_r2_add_btls())) {
return rc;
}
bml_endpoint = mca_bml_r2_allocate_endpoint (proc);
if (OPAL_UNLIKELY(NULL == bml_endpoint)) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
for (size_t p_index = 0 ; p_index < mca_bml_r2.num_btl_modules ; ++p_index) {
mca_btl_base_module_t *btl = mca_bml_r2.btl_modules[p_index];
struct mca_btl_base_endpoint_t *btl_endpoint = NULL;
/* if the r2 can reach the destination proc it sets the
* corresponding bit (proc index) in the reachable bitmap
* and can return addressing information for each proc
* that is passed back to the r2 on data transfer calls
*/
rc = btl->btl_add_procs (btl, 1, (opal_proc_t **) &proc, &btl_endpoint, NULL);
if (OMPI_SUCCESS != rc || NULL == btl_endpoint) {
/* This BTL has troubles adding the nodes. Let's continue maybe some other BTL
* can take care of this task. */
continue;
}
rc = mca_bml_r2_endpoint_add_btl (proc, bml_endpoint, btl, btl_endpoint);
if (OMPI_SUCCESS != rc) {
btl->btl_del_procs (btl, 1, (opal_proc_t **) &proc, &btl_endpoint);
} else {
mca_bml_r2_register_progress (btl, true);
btl_in_use = true;
}
}
if (!btl_in_use) {
proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML] = NULL;
OBJ_RELEASE(bml_endpoint);
/* no btl is available for this proc */
if (mca_bml_r2.show_unreach_errors) {
opal_show_help ("help-mca-bml-r2.txt", "unreachable proc", true,
OMPI_NAME_PRINT(&(ompi_proc_local_proc->super.proc_name)),
(NULL != ompi_proc_local_proc->super.proc_hostname ?
ompi_proc_local_proc->super.proc_hostname : "unknown!"),
OMPI_NAME_PRINT(&(proc->super.proc_name)),
(NULL != proc->super.proc_hostname ?
proc->super.proc_hostname : "unknown!"),
btl_names);
}
return OMPI_ERR_UNREACH;
}
/* compute metrics for registered btls */
mca_bml_r2_compute_endpoint_metrics (bml_endpoint);
/* do it last, for the lazy initialization check in bml_base_get* */
opal_atomic_wmb();
proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML] = bml_endpoint;
return OMPI_SUCCESS;
}
/*
* For each proc setup a datastructure that indicates the BTLs
* that can be used to reach the destination.
*
*/
static int mca_bml_r2_add_procs( size_t nprocs,
struct ompi_proc_t** procs,
struct opal_bitmap_t* reachable )
{
size_t n_new_procs = 0;
struct mca_btl_base_endpoint_t ** btl_endpoints = NULL;
struct ompi_proc_t** new_procs = NULL;
int rc, ret = OMPI_SUCCESS;
if(0 == nprocs) {
return OMPI_SUCCESS;
}
if(OMPI_SUCCESS != (rc = mca_bml_r2_add_btls()) ) {
return rc;
}
/* Select only the procs that don't yet have the BML proc struct. This prevent
* us from calling btl->add_procs several times on the same destination proc.
*/
for (size_t p_index = 0 ; p_index < nprocs ; ++p_index) {
struct ompi_proc_t* proc = procs[p_index];
if(NULL != proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML]) {
continue; /* go to the next proc */
}
/* Allocate the new_procs on demand */
if( NULL == new_procs ) {
new_procs = (struct ompi_proc_t **)malloc(nprocs * sizeof(struct ompi_proc_t *));
if( NULL == new_procs ) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
}
OBJ_RETAIN(proc);
new_procs[n_new_procs++] = proc;
}
if ( 0 == n_new_procs ) {
return OMPI_SUCCESS;
}
/* Starting from here we only work on the unregistered procs */
procs = new_procs;
nprocs = n_new_procs;
/* attempt to add all procs to each r2 */
btl_endpoints = (struct mca_btl_base_endpoint_t **)
malloc(nprocs * sizeof(struct mca_btl_base_endpoint_t*));
if (NULL == btl_endpoints) {
free(new_procs);
return OMPI_ERR_OUT_OF_RESOURCE;
}
for (size_t p_index = 0 ; p_index < mca_bml_r2.num_btl_modules ; ++p_index) {
mca_btl_base_module_t *btl = mca_bml_r2.btl_modules[p_index];
int btl_inuse = 0;
/* if the r2 can reach the destination proc it sets the
* corresponding bit (proc index) in the reachable bitmap
* and can return addressing information for each proc
* that is passed back to the r2 on data transfer calls
*/
opal_bitmap_clear_all_bits(reachable);
memset(btl_endpoints, 0, nprocs *sizeof(struct mca_btl_base_endpoint_t*));
rc = btl->btl_add_procs(btl, n_new_procs, (opal_proc_t**)new_procs, btl_endpoints, reachable);
if (OMPI_SUCCESS != rc) {
/* This BTL encountered an error while adding procs. Continue in case some other
* BTL(s) can be used. */
continue;
}
/* for each proc that is reachable */
for (size_t p = 0 ; p < n_new_procs ; ++p) {
if (!opal_bitmap_is_set_bit(reachable, p)) {
continue;
}
ompi_proc_t *proc = new_procs[p];
mca_bml_base_endpoint_t *bml_endpoint =
(mca_bml_base_endpoint_t *) proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML];
if (NULL == bml_endpoint) {
bml_endpoint = mca_bml_r2_allocate_endpoint (proc);
proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML] = bml_endpoint;
if (NULL == bml_endpoint) {
free(btl_endpoints);
free(new_procs);
return OPAL_ERR_OUT_OF_RESOURCE;
}
}
rc = mca_bml_r2_endpoint_add_btl (proc, bml_endpoint, btl, btl_endpoints[p]);
if (OMPI_SUCCESS != rc) {
btl->btl_del_procs(btl, 1, (opal_proc_t**)&proc, &btl_endpoints[p]);
continue;
}
/* This BTL is in use, allow the progress registration */
btl_inuse++;
}
mca_bml_r2_register_progress (btl, !!(btl_inuse));
}
free(btl_endpoints);
/* iterate back through procs and compute metrics for registered r2s */
for (size_t p = 0; p < n_new_procs ; ++p) {
mca_bml_base_endpoint_t *bml_endpoint =
(mca_bml_base_endpoint_t *) new_procs[p]->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML];
/* skip over procs w/ no btl's registered */
if (NULL != bml_endpoint) {
mca_bml_r2_compute_endpoint_metrics (bml_endpoint);
}
}
/* see if we have a connection to everyone else */
for(size_t p = 0; p < n_new_procs ; ++p) {
ompi_proc_t *proc = new_procs[p];
if (NULL == proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML]) {
ret = OMPI_ERR_UNREACH;
if (mca_bml_r2.show_unreach_errors) {
opal_show_help("help-mca-bml-r2.txt", "unreachable proc", true,
OMPI_NAME_PRINT(&(ompi_proc_local_proc->super.proc_name)),
(NULL != ompi_proc_local_proc->super.proc_hostname ?
ompi_proc_local_proc->super.proc_hostname : "unknown!"),
OMPI_NAME_PRINT(&(proc->super.proc_name)),
(NULL != proc->super.proc_hostname ?
proc->super.proc_hostname : "unknown!"),
btl_names);
}
break;
}
}
free(new_procs);
return ret;
}
/*
* iterate through each proc and notify any BTLs associated
* with the proc that it is/has gone away
*/
static int mca_bml_r2_del_procs(size_t nprocs,
struct ompi_proc_t** procs)
{
for (size_t p = 0 ; p < nprocs ; ++p) {
ompi_proc_t *proc = procs[p];
mca_bml_base_endpoint_t *bml_endpoint =
(mca_bml_base_endpoint_t*) proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML];
if (!bml_endpoint) {
/* NTH: I would think this is a developer bug and should not be ignored. */
continue;
}
/* notify each btl that the proc is going away */
size_t f_size = mca_bml_base_btl_array_get_size (&bml_endpoint->btl_send);
for (size_t f_index = 0 ; f_index < f_size ; ++f_index) {
mca_bml_base_btl_t* bml_btl = mca_bml_base_btl_array_get_index(&bml_endpoint->btl_send, f_index);
mca_btl_base_module_t *btl = bml_btl->btl;
int rc = btl->btl_del_procs (btl, 1, (opal_proc_t **) &proc, &bml_btl->btl_endpoint);
if (OPAL_SUCCESS != rc) {
return rc;
}
/* The reference stored in btl_eager and btl_rdma will automatically
* dissapear once the btl_array destructor is called. Thus, there is
* no need for extra cleaning here.
*/
}
/* some btl endpoints may only be in the btl_rdma array. call del_procs on those as well */
size_t r_size = mca_bml_base_btl_array_get_size (&bml_endpoint->btl_rdma);
for (size_t r_index = 0 ; r_index < r_size ; ++r_index) {
mca_bml_base_btl_t *rdma_btl = mca_bml_base_btl_array_get_index (&bml_endpoint->btl_rdma, r_index);
mca_btl_base_module_t *btl = rdma_btl->btl;
bool needs_del = true;
for (size_t f_index = 0 ; f_index < f_size ; ++f_index) {
mca_bml_base_btl_t *bml_btl = mca_bml_base_btl_array_get_index (&bml_endpoint->btl_send, f_index);
if (bml_btl->btl_endpoint == rdma_btl->btl_endpoint) {
needs_del = false;
break;
}
}
if (needs_del) {
int rc = btl->btl_del_procs (btl, 1, (opal_proc_t **) &proc, &rdma_btl->btl_endpoint);
if (OPAL_SUCCESS != rc) {
return rc;
}
}
}
proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML] = NULL;
/* release the bml endpoint's reference to the proc */
OBJ_RELEASE(proc);
/* do any required cleanup */
OBJ_RELEASE(bml_endpoint);
}
return OMPI_SUCCESS;
}
static inline int bml_r2_remove_btl_progress(mca_btl_base_module_t* btl)
{
unsigned int p;
if(NULL == btl->btl_component->btl_progress) {
return OMPI_SUCCESS;
}
for(p = 0; p < mca_bml_r2.num_btl_progress; p++) {
if(btl->btl_component->btl_progress != mca_bml_r2.btl_progress[p])
continue;
opal_progress_unregister( btl->btl_component->btl_progress );
if( p < (mca_bml_r2.num_btl_progress-1) ) {
mca_bml_r2.btl_progress[p] = mca_bml_r2.btl_progress[mca_bml_r2.num_btl_progress-1];
}
mca_bml_r2.num_btl_progress--;
return OMPI_SUCCESS;
}
return OMPI_ERR_NOT_FOUND;
}
static int mca_bml_r2_del_proc_btl(ompi_proc_t* proc, mca_btl_base_module_t* btl)
{
mca_bml_base_endpoint_t* ep = (mca_bml_base_endpoint_t*)proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML];
mca_bml_base_btl_t* bml_btl;
mca_btl_base_module_t* ep_btl;
double total_bandwidth = 0;
size_t b;
if(NULL == ep)
return OMPI_SUCCESS;
/* remove btl from eager list */
mca_bml_base_btl_array_remove(&ep->btl_eager, btl);
/* remove btl from send list */
if(mca_bml_base_btl_array_remove(&ep->btl_send, btl)) {
/* compute total_bandwidth and
reset max_send_size to the min of all btl's */
total_bandwidth = 0;
ep->btl_max_send_size = -1;
for(b=0; b< mca_bml_base_btl_array_get_size(&ep->btl_send); b++) {
bml_btl = mca_bml_base_btl_array_get_index(&ep->btl_send, b);
ep_btl = bml_btl->btl;
total_bandwidth += ep_btl->btl_bandwidth;
if (ep->btl_max_send_size > ep_btl->btl_max_send_size) {
ep->btl_max_send_size = ep_btl->btl_max_send_size;
}
}
/* compute weighting factor for this btl */
for(b=0; b< mca_bml_base_btl_array_get_size(&ep->btl_send); b++) {
bml_btl = mca_bml_base_btl_array_get_index(&ep->btl_send, b);
ep_btl = bml_btl->btl;
if(ep_btl->btl_bandwidth > 0) {
bml_btl->btl_weight = (float)(ep_btl->btl_bandwidth / total_bandwidth);
} else {
bml_btl->btl_weight = (float)(1.0 / mca_bml_base_btl_array_get_size(&ep->btl_send));
}
}
}
/* remove btl from RDMA list */
if(mca_bml_base_btl_array_remove(&ep->btl_rdma, btl)) {
/* compute total bandwidth */
total_bandwidth = 0;
ep->btl_pipeline_send_length = 0;
ep->btl_send_limit = 0;
for(b=0; b< mca_bml_base_btl_array_get_size(&ep->btl_rdma); b++) {
bml_btl = mca_bml_base_btl_array_get_index(&ep->btl_rdma, b);
ep_btl = bml_btl->btl;
/* update aggregate endpoint info */
total_bandwidth += ep_btl->btl_bandwidth;
if (ep->btl_pipeline_send_length < ep_btl->btl_rdma_pipeline_send_length) {
ep->btl_pipeline_send_length = ep_btl->btl_rdma_pipeline_send_length;
}
if (ep->btl_send_limit < ep_btl->btl_min_rdma_pipeline_size) {
ep->btl_send_limit = ep_btl->btl_min_rdma_pipeline_size;
}
}
/* compute weighting factor for this btl */
for(b=0; b< mca_bml_base_btl_array_get_size(&ep->btl_rdma); b++) {
bml_btl = mca_bml_base_btl_array_get_index(&ep->btl_rdma, b);
ep_btl = bml_btl->btl;
if(ep_btl->btl_bandwidth > 0) {
bml_btl->btl_weight = (float)(ep_btl->btl_bandwidth / total_bandwidth);
} else {
bml_btl->btl_weight = (float)(1.0 / mca_bml_base_btl_array_get_size(&ep->btl_rdma));
}
}
}
return OMPI_SUCCESS;
}
int mca_bml_r2_finalize( void )
{
ompi_proc_t** procs;
size_t p, num_procs;
opal_list_item_t* w_item;
if (NULL != btl_names) {
free(btl_names);
btl_names = NULL;
}
/* Similar to mca_bml_r2_del_btl ... */
procs = ompi_proc_all(&num_procs);
if(NULL == procs)
goto CLEANUP;
for (w_item = opal_list_get_first(&mca_btl_base_modules_initialized);
w_item != opal_list_get_end(&mca_btl_base_modules_initialized);
w_item = opal_list_get_next(w_item)) {
mca_btl_base_selected_module_t *sm = (mca_btl_base_selected_module_t *) w_item;
mca_btl_base_module_t* btl = sm->btl_module;
/* unregister the BTL progress function if any */
bml_r2_remove_btl_progress(btl);
/* dont use this btl for any peers */
for( p = 0; p < num_procs; p++ ) {
ompi_proc_t* proc = procs[p];
mca_bml_r2_del_proc_btl(proc, sm->btl_module);
}
}
/* Release the procs as the ompi_proc_all increase their ref_count */
for( p = 0; p < num_procs; p++ ) {
OBJ_RELEASE(procs[p]);
}
free(procs);
CLEANUP:
mca_bml_r2.num_btl_modules = 0;
mca_bml_r2.num_btl_progress = 0;
if( NULL != mca_bml_r2.btl_modules) {
free(mca_bml_r2.btl_modules);
mca_bml_r2.btl_modules = NULL;
}
if( NULL != mca_bml_r2.btl_progress ) {
free(mca_bml_r2.btl_progress);
mca_bml_r2.btl_progress = NULL;
}
/* Do not close the BTL base here; the BML upper layer will take
care of that. */
return OMPI_SUCCESS;
}
/*
* (1) Remove btl from each bml endpoint
* (2) Remove btl from the global list
*/
static int mca_bml_r2_del_btl(mca_btl_base_module_t* btl)
{
ompi_proc_t** procs;
size_t i, m, p, num_procs;
opal_list_item_t* item;
mca_btl_base_module_t** modules;
bool found = false;
if(opal_list_get_size(&mca_btl_base_modules_initialized) == 2) {
opal_output(0, "only one BTL left, can't failover");
return OMPI_SUCCESS;
}
procs = ompi_proc_all(&num_procs);
if(NULL == procs)
return OMPI_SUCCESS;
/* Get rid of the associated progress function */
bml_r2_remove_btl_progress(btl);
/* dont use this btl for any peers */
for( p = 0; p < num_procs; p++ ) {
ompi_proc_t* proc = procs[p];
mca_bml_r2_del_proc_btl(proc, btl);
}
/* remove from the btl list */
for (item = opal_list_get_first(&mca_btl_base_modules_initialized);
item != opal_list_get_end(&mca_btl_base_modules_initialized);
item = opal_list_get_next(item)) {
mca_btl_base_selected_module_t *sm = (mca_btl_base_selected_module_t *) item;
if(sm->btl_module == btl) {
opal_list_remove_item(&mca_btl_base_modules_initialized, item);
free(sm);
found = true;
break;
}
}
if(!found) {
/* doesn't even exist */
goto CLEANUP;
}
/* remove from bml list */
modules = (mca_btl_base_module_t**)malloc(sizeof(mca_btl_base_module_t*) * (mca_bml_r2.num_btl_modules-1));
for(i=0,m=0; i<mca_bml_r2.num_btl_modules; i++) {
if(mca_bml_r2.btl_modules[i] != btl) {
modules[m++] = mca_bml_r2.btl_modules[i];
}
}
free(mca_bml_r2.btl_modules);
mca_bml_r2.btl_modules = modules;
mca_bml_r2.num_btl_modules = m;
/* cleanup */
btl->btl_finalize(btl);
CLEANUP:
/* Decrease the ref_count increased by the call to ompi_proc_all */
for( p = 0; p < num_procs; p++ ) {
OBJ_RELEASE(procs[p]);
}
free(procs);
return OMPI_SUCCESS;
}
static int mca_bml_r2_add_btl(mca_btl_base_module_t* btl)
{
return OMPI_ERR_NOT_IMPLEMENTED;
}
/*
* Register callback w/ all active btls
*/
static int mca_bml_r2_register( mca_btl_base_tag_t tag,
mca_btl_base_module_recv_cb_fn_t cbfunc,
void* data )
{
mca_btl_base_active_message_trigger[tag].cbfunc = cbfunc;
mca_btl_base_active_message_trigger[tag].cbdata = data;
/* Give an oportunity to the BTLs to do something special
* for each registration.
*/
{
int i, rc;
mca_btl_base_module_t *btl;
for(i = 0; i < (int)mca_bml_r2.num_btl_modules; i++) {
btl = mca_bml_r2.btl_modules[i];
if( NULL == btl->btl_register )
continue;
rc = btl->btl_register(btl, tag, cbfunc, data);
if(OMPI_SUCCESS != rc) {
return rc;
}
}
}
return OMPI_SUCCESS;
}
/*
* Register an error handler with/ all active btls
* if they support error handlers..
*/
static int mca_bml_r2_register_error( mca_btl_base_module_error_cb_fn_t cbfunc)
{
uint32_t i;
int rc;
mca_btl_base_module_t *btl;
uint32_t ver;
for(i = 0; i < mca_bml_r2.num_btl_modules; i++) {
btl = mca_bml_r2.btl_modules[i];
/* this wont work for version numbers greater than 256... seems
reasonable.. */
ver = btl->btl_component->btl_version.mca_type_major_version << 16 |
btl->btl_component->btl_version.mca_type_minor_version << 8 |
btl->btl_component->btl_version.mca_type_release_version;
/* is version number greater than or equal to 1.0.1? */
if(ver >= ((1 << 16) | (0 << 8) | 1) &&
NULL != btl->btl_register_error) {
rc = btl->btl_register_error(btl, cbfunc);
if(OMPI_SUCCESS != rc) {
return rc;
}
}
}
return OMPI_SUCCESS;
}
int mca_bml_r2_component_fini(void)
{
return OMPI_SUCCESS;
}
mca_bml_r2_module_t mca_bml_r2 = {
.super = {
.bml_component = &mca_bml_r2_component,
.bml_add_proc = mca_bml_r2_add_proc,
.bml_add_procs = mca_bml_r2_add_procs,
.bml_del_procs = mca_bml_r2_del_procs,
.bml_add_btl = mca_bml_r2_add_btl,
.bml_del_btl = mca_bml_r2_del_btl,
.bml_del_proc_btl = mca_bml_r2_del_proc_btl,
.bml_register = mca_bml_r2_register,
.bml_register_error = mca_bml_r2_register_error,
.bml_finalize = mca_bml_r2_finalize,
.bml_ft_event = mca_bml_r2_ft_event,
},
};
Показать git blame Копировать постоянную ссылку