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

bml/r2: add all rdma btls even if another btl has higher exclusivity

Просмотр исходного кода 
Background: In order to support atomics each btl needs to provide support
for communicating with self unless the btl module can guarantee global
atomicity. Before this commit bml/r2 discarded any BTL with lower
exclusivity than an existing send btl. This would cause the BML to
discard any btl other than self.

The new behavior is as follows:

 - If an exisiting send btl has higher exclusivity then the btl will not be
   added to the send btl list for the endpoint.

 - If a btl provides RDMA support then it is always added to the rdma btl
   list.

 - bml_btl weight for send btls is now calculated across all send btls.

 - bml_btl weight for rdma btls is now calculated across all rdma btls.

With this change self should still win as the only send btl for loopback
without disqualifying other btls (ugni, openib) for atomic operations.

Signed-off-by: Nathan Hjelm <hjelmn@lanl.gov>
Этот коммит содержится в:
Nathan Hjelm 2014-11-14 13:16:45 -07:00 коммит произвёл Nathan Hjelm
родитель 3847025540
Коммит 4bf7a207e9
1 изменённых файлов: 105 добавлений и 63 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

168
ompi/mca/bml/r2/bml_r2.c
Просмотреть файл

@ -17,6 +17,8 @@
* 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.
* $COPYRIGHT$
*
* Additional copyrights may follow
@ -86,9 +88,7 @@ static int mca_bml_r2_add_btls( void )
return OMPI_ERR_OUT_OF_RESOURCE;
}
for(selected_btl = (mca_btl_base_selected_module_t*)opal_list_get_first(btls);
selected_btl != (mca_btl_base_selected_module_t*)opal_list_get_end(btls);
selected_btl = (mca_btl_base_selected_module_t*)opal_list_get_next(selected_btl)) {
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) {
@ -127,6 +127,23 @@ static int btl_bandwidth_compare(const void *v1, const void *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;
}
}
}
/*
* For each proc setup a datastructure that indicates the BTLs
* that can be used to reach the destination.
@ -189,6 +206,7 @@ static int mca_bml_r2_add_procs( size_t nprocs,
for(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;
int btl_flags;
/* if the r2 can reach the destination proc it sets the
* corresponding bit (proc index) in the reachable bitmap
@ -212,7 +230,7 @@ static int mca_bml_r2_add_procs( size_t nprocs,
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];
mca_bml_base_btl_t* bml_btl;
mca_bml_base_btl_t* bml_btl = NULL;
size_t size;
if(NULL == bml_endpoint) {
@ -236,12 +254,35 @@ static int mca_bml_r2_add_procs( size_t nprocs,
bml_endpoint->btl_flags_or = 0;
}
btl_flags = btl->btl_flags;
if( (btl_flags & MCA_BTL_FLAGS_PUT) && (NULL == btl->btl_put) ) {
opal_output(0, "mca_bml_r2_add_procs: The PUT flag is specified for"
" the %s BTL without any PUT function attached. Discard the flag !",
bml_btl->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, "mca_bml_r2_add_procs: The GET flag is specified for"
" the %s BTL without any GET function attached. Discard the flag !",
bml_btl->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.
*/
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);
if(size > 0) {
bml_btl = mca_bml_base_btl_array_get_index(&bml_endpoint->btl_send, size-1);
/* skip this btl if the exclusivity is less than the previous */
if(bml_btl->btl->btl_exclusivity > btl->btl_exclusivity) {
/* skip this btl if the exclusivity is less than the previous only if the btl does not provide full rdma (for one-sided) */
if(bml_btl->btl->btl_exclusivity > btl->btl_exclusivity && ((btl_flags & MCA_BTL_FLAGS_RDMA) != MCA_BTL_FLAGS_RDMA)) {
btl->btl_del_procs(btl, 1, (opal_proc_t**)&proc, &btl_endpoints[p]);
opal_output_verbose(20, opal_btl_base_framework.framework_output,
"mca: bml: Not using %s btl to %s on node %s "
@ -261,39 +302,44 @@ static int mca_bml_r2_add_procs( size_t nprocs,
proc->super.proc_hostname);
/* cache the endpoint on the proc */
bml_btl = mca_bml_base_btl_array_insert(&bml_endpoint->btl_send);
bml_btl->btl = btl;
bml_btl->btl_endpoint = btl_endpoints[p];
bml_btl->btl_weight = 0;
bml_btl->btl_flags = btl->btl_flags;
if( (bml_btl->btl_flags & MCA_BTL_FLAGS_PUT) && (NULL == btl->btl_put) ) {
opal_output(0, "mca_bml_r2_add_procs: The PUT flag is specified for"
" the %s BTL without any PUT function attached. Discard the flag !",
bml_btl->btl->btl_component->btl_version.mca_component_name);
bml_btl->btl_flags ^= MCA_BTL_FLAGS_PUT;
}
if( (bml_btl->btl_flags & MCA_BTL_FLAGS_GET) && (NULL == btl->btl_get) ) {
opal_output(0, "mca_bml_r2_add_procs: The GET flag is specified for"
" the %s BTL without any GET function attached. Discard the flag !",
bml_btl->btl->btl_component->btl_version.mca_component_name);
bml_btl->btl_flags ^= MCA_BTL_FLAGS_GET;
}
if( (bml_btl->btl_flags & (MCA_BTL_FLAGS_PUT | MCA_BTL_FLAGS_GET | MCA_BTL_FLAGS_SEND)) == 0 ) {
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_endpoints[p];
bml_btl->btl_weight = 0;
bml_btl->btl_flags = btl_flags;
/**
* 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.
* calculate the bitwise OR of the btl flags
*/
bml_btl->btl_flags |= MCA_BTL_FLAGS_SEND;
bml_endpoint->btl_flags_or |= bml_btl->btl_flags;
}
/**
* calculate the bitwise OR of the btl flags
*/
bml_endpoint->btl_flags_or |= bml_btl->btl_flags;
/* always add rdma endpoints */
if ((btl_flags & MCA_BTL_FLAGS_RDMA) &&
!((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_endpoints[p];
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;
}
}
/* This BTL is in use, allow the progress registration */
btl_inuse++;
}
}
if(btl_inuse > 0 && NULL != btl->btl_component->btl_progress) {
size_t p;
bool found = false;
@ -319,9 +365,8 @@ static int mca_bml_r2_add_procs( size_t nprocs,
mca_bml_base_endpoint_t* bml_endpoint =
(mca_bml_base_endpoint_t*) proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML];
double total_bandwidth = 0;
uint32_t latency = 0xffffffff;
size_t n_index;
size_t n_size;
uint32_t latency;
size_t n_send, n_rdma;
/* skip over procs w/ no btl's registered */
if(NULL == bml_endpoint) {
@ -335,28 +380,22 @@ static int mca_bml_r2_add_procs( size_t nprocs,
* weighting. Once the left over is smaller than this number we will
* start using the weight to compute the correct amount.
*/
n_size = mca_bml_base_btl_array_get_size(&bml_endpoint->btl_send);
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_size,
qsort(bml_endpoint->btl_send.bml_btls, n_send,
sizeof(mca_bml_base_btl_t), btl_bandwidth_compare);
bml_endpoint->btl_rdma_index = 0;
for(n_index = 0; n_index < n_size; 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;
total_bandwidth += bml_btl->btl->btl_bandwidth;
if(btl->btl_latency < latency) {
latency = btl->btl_latency;
}
}
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(n_index = 0; n_index < n_size; n_index++) {
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;
@ -365,7 +404,7 @@ static int mca_bml_r2_add_procs( size_t nprocs,
if(btl->btl_bandwidth > 0) {
bml_btl->btl_weight = (float)(btl->btl_bandwidth / total_bandwidth);
} else {
bml_btl->btl_weight = (float)(1.0 / n_size);
bml_btl->btl_weight = (float)(1.0 / n_send);
}
/* check to see if this r2 is already in the array of r2s
@ -380,21 +419,24 @@ static int mca_bml_r2_add_procs( size_t nprocs,
/* 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;
}
/* check flags - is rdma prefered */
if ((btl->btl_flags & (MCA_BTL_FLAGS_PUT|MCA_BTL_FLAGS_GET)) &&
!((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);
mca_btl_base_module_t* btl_rdma = bml_btl->btl;
/* 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);
*bml_btl_rdma = *bml_btl;
if(bml_endpoint->btl_pipeline_send_length < btl_rdma->btl_rdma_pipeline_send_length) {
bml_endpoint->btl_pipeline_send_length = btl_rdma->btl_rdma_pipeline_send_length;
}
if(bml_endpoint->btl_send_limit < btl_rdma->btl_min_rdma_pipeline_size) {
bml_endpoint->btl_send_limit = btl_rdma->btl_min_rdma_pipeline_size;
}
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);
}
}
}