static int mca_bml_r2_add_procs( size_t nprocs,
struct ompi_proc_t** procs,
struct opal_bitmap_t* reachable )
{
size_t p, p_index, n_new_procs = 0;
struct mca_btl_base_endpoint_t ** btl_endpoints = NULL;
struct ompi_proc_t** new_procs = NULL;
struct ompi_proc_t *unreach_proc = 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 this on the same destination proc.
*/
for(p_index = 0; p_index < nprocs; p_index++) {
struct ompi_proc_t* proc = procs[p_index];
OBJ_RETAIN(proc);
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;
}
}
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(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, new_procs, btl_endpoints, reachable);
if(OMPI_SUCCESS != rc) {
/* This BTL has troubles adding the nodes. Let's continue maybe some other BTL
* can take care of this task.
*/
continue;
}
/* for each proc that is reachable */
for( p = 0; p < n_new_procs; p++ ) {
if(opal_bitmap_is_set_bit(reachable, p)) {
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;
size_t size;
if(NULL == bml_endpoint) {
/* allocate bml specific proc data */
bml_endpoint = OBJ_NEW(mca_bml_base_endpoint_t);
if (NULL == bml_endpoint) {
opal_output(0, "mca_bml_r2_add_procs: unable to allocate resources");
free(btl_endpoints);
free(new_procs);
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* 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;
proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML] = bml_endpoint;
bml_endpoint->btl_flags_or = 0;
}
/* 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) {
btl->btl_del_procs(btl, 1, &proc, &btl_endpoints[p]);
continue;
}
}
/* 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. Disard 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 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.
*/
bml_btl->btl_flags |= MCA_BTL_FLAGS_SEND;
}
/**
* calculate the bitwise OR of the btl flags
*/
bml_endpoint->btl_flags_or |= bml_btl->btl_flags;
/* 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;
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) {
mca_bml_r2.btl_progress[mca_bml_r2.num_btl_progress] =
btl->btl_component->btl_progress;
mca_bml_r2.num_btl_progress++;
opal_progress_register( btl->btl_component->btl_progress );
}
}
}
free(btl_endpoints);
/* iterate back through procs and compute metrics for registered r2s */
for(p=0; p<n_new_procs; p++) {
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];
double total_bandwidth = 0;
uint32_t latency = 0xffffffff;
size_t n_index;
size_t n_size;
/* skip over procs w/ no btl's registered */
if(NULL == bml_endpoint) {
continue;
}
/* (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_size = mca_bml_base_btl_array_get_size(&bml_endpoint->btl_send);
/* sort BTLs in descending order according to bandwidth value */
qsort(bml_endpoint->btl_send.bml_btls, n_size,
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;
}
}
/* (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++) {
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_size);
}
/* 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;
/* check flags - is rdma prefered */
if ((btl->btl_flags & (MCA_BTL_FLAGS_PUT|MCA_BTL_FLAGS_GET)) &&
!((proc->proc_arch != ompi_proc_local_proc->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;
*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;
}
}
}
}
/* see if we have a connection to everyone else */
for(p=0; p<n_new_procs; p++) {
ompi_proc_t *proc = new_procs[p];
if (NULL == proc->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_BML]) {
if (NULL == unreach_proc) {
unreach_proc = proc;
}
ret = OMPI_ERR_UNREACH;
}
}
if (mca_bml_r2.show_unreach_errors &&
OMPI_ERR_UNREACH == ret) {
opal_show_help("help-mca-bml-r2.txt",
"unreachable proc",
true,
OMPI_NAME_PRINT(&(ompi_proc_local_proc->proc_name)),
(NULL != ompi_proc_local_proc->proc_hostname ?
ompi_proc_local_proc->proc_hostname : "unknown!"),
OMPI_NAME_PRINT(&(unreach_proc->proc_name)),
(NULL != ompi_proc_local_proc->proc_hostname ?
ompi_proc_local_proc->proc_hostname : "unknown!"),
btl_names);
}
free(new_procs);
return ret;
}
int mca_bml_r2_add_procs(
size_t nprocs,
struct ompi_proc_t** procs,
struct mca_bml_base_endpoint_t** bml_endpoints,
struct ompi_bitmap_t* reachable
)
{
size_t p;
int rc;
size_t p_index;
struct mca_btl_base_endpoint_t ** btl_endpoints = NULL;
struct ompi_proc_t** new_procs = NULL;
size_t n_new_procs = 0;
int ret = OMPI_SUCCESS;
struct ompi_proc_t *unreach_proc = NULL;
if(0 == nprocs) {
return OMPI_SUCCESS;
}
if(OMPI_SUCCESS != (rc = mca_bml_r2_add_btls()) ) {
return rc;
}
new_procs = (struct ompi_proc_t **)
malloc(nprocs * sizeof(struct ompi_proc_t *));
if (NULL == new_procs ) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
memset(bml_endpoints, 0, nprocs * sizeof(struct mca_bml_base_endpoint_t*));
for(p_index = 0; p_index < nprocs; p_index++) {
struct ompi_proc_t* proc;
proc = procs[p_index];
OBJ_RETAIN(proc);
if(NULL != proc->proc_bml) {
bml_endpoints[p_index] =
(mca_bml_base_endpoint_t*) proc->proc_bml;
} else {
new_procs[n_new_procs++] = proc;
}
}
if ( 0 == n_new_procs ) {
return OMPI_SUCCESS;
}
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) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
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;
/* 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
*/
ompi_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, new_procs, btl_endpoints, reachable);
if(OMPI_SUCCESS != rc) {
free(btl_endpoints);
return rc;
}
/* for each proc that is reachable - add the endpoint to the bml_endpoints array(s) */
for(p=0; p<n_new_procs; p++) {
if(ompi_bitmap_is_set_bit(reachable, p)) {
ompi_proc_t *proc = new_procs[p];
mca_bml_base_endpoint_t * bml_endpoint = (mca_bml_base_endpoint_t*) proc->proc_bml;
mca_bml_base_btl_t* bml_btl;
size_t size;
btl_inuse++;
if(NULL == bml_endpoint) {
/* allocate bml specific proc data */
bml_endpoint = OBJ_NEW(mca_bml_base_endpoint_t);
if (NULL == bml_endpoint) {
opal_output(0, "mca_bml_r2_add_procs: unable to allocate resources");
free(btl_endpoints);
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* 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;
proc->proc_bml = bml_endpoint;
bml_endpoint->btl_flags_and = 0;
bml_endpoint->btl_flags_or = 0;
}
bml_endpoints[p] =(mca_bml_base_endpoint_t*) proc->proc_bml;
/* 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) {
if(btl_endpoints[p] != NULL) {
btl->btl_del_procs(btl, 1, &proc, &btl_endpoints[p]);
}
btl_inuse--;
continue;
}
}
/* 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_eager_limit = btl->btl_eager_limit;
bml_btl->btl_min_send_size = btl->btl_min_send_size;
bml_btl->btl_max_send_size = btl->btl_max_send_size;
bml_btl->btl_min_rdma_size = btl->btl_min_rdma_size;
bml_btl->btl_max_rdma_size = btl->btl_max_rdma_size;
bml_btl->btl_cache = NULL;
bml_btl->btl_endpoint = btl_endpoints[p];
bml_btl->btl_weight = 0;
bml_btl->btl_alloc = btl->btl_alloc;
bml_btl->btl_free = btl->btl_free;
bml_btl->btl_prepare_src = btl->btl_prepare_src;
bml_btl->btl_prepare_dst = btl->btl_prepare_dst;
bml_btl->btl_send = btl->btl_send;
bml_btl->btl_flags = btl->btl_flags;
bml_btl->btl_put = btl->btl_put;
if( (bml_btl->btl_flags & MCA_BTL_FLAGS_PUT) && (NULL == bml_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. Disard the flag !",
bml_btl->btl->btl_component->btl_version.mca_component_name);
bml_btl->btl_flags ^= MCA_BTL_FLAGS_PUT;
}
bml_btl->btl_get = btl->btl_get;
if( (bml_btl->btl_flags & MCA_BTL_FLAGS_GET) && (NULL == bml_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. Disard the flag !",
bml_btl->btl->btl_component->btl_version.mca_component_name);
bml_btl->btl_flags ^= MCA_BTL_FLAGS_GET;
}
bml_btl->btl_mpool = btl->btl_mpool;
if( (bml_btl->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.
*/
bml_btl->btl_flags |= MCA_BTL_FLAGS_SEND;
}
/**
* calculate the bitwise OR and AND of the btl flags
*/
bml_endpoint->btl_flags_or |= bml_btl->btl_flags;
bml_endpoint->btl_flags_and &= bml_btl->btl_flags;
}
}
if(btl_inuse > 0 && NULL != btl->btl_component->btl_progress) {
size_t p;
bool found = false;
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) {
mca_bml_r2.btl_progress[mca_bml_r2.num_btl_progress] =
btl->btl_component->btl_progress;
mca_bml_r2.num_btl_progress++;
}
}
}
free(btl_endpoints);
/* iterate back through procs and compute metrics for registered r2s */
for(p=0; p<n_new_procs; p++) {
ompi_proc_t *proc = new_procs[p];
mca_bml_base_endpoint_t* bml_endpoint = (mca_bml_base_endpoint_t*) proc->proc_bml;
double total_bandwidth = 0;
uint32_t latency = 0xffffffff;
size_t n_index;
size_t n_size;
/* skip over procs w/ no btl's registered */
if(NULL == bml_endpoint) {
continue;
}
/* (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_size = mca_bml_base_btl_array_get_size(&bml_endpoint->btl_send);
bml_endpoint->bml_max_send_length = 0;
bml_endpoint->bml_max_rdma_length = 0;
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;
}
bml_endpoint->bml_max_send_length += bml_btl->btl->btl_bandwidth;
}
/* (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++) {
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 = btl->btl_bandwidth / total_bandwidth;
} else {
bml_btl->btl_weight = 1.0 / n_size;
}
/* 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;
/* check flags - is rdma prefered */
if(btl->btl_flags & (MCA_BTL_FLAGS_PUT|MCA_BTL_FLAGS_GET) &&
proc->proc_arch == ompi_proc_local_proc->proc_arch) {
mca_bml_base_btl_t* bml_btl_rdma = mca_bml_base_btl_array_insert(&bml_endpoint->btl_rdma);
*bml_btl_rdma = *bml_btl;
if(bml_endpoint->btl_rdma_offset < bml_btl_rdma->btl_min_rdma_size) {
bml_endpoint->btl_rdma_offset = bml_btl_rdma->btl_min_rdma_size;
}
}
}
}
/* see if we have a connection to everyone else */
for(p=0; p<n_new_procs; p++) {
ompi_proc_t *proc = new_procs[p];
if (NULL == proc->proc_bml) {
if (NULL == unreach_proc) {
unreach_proc = proc;
}
ret = OMPI_ERR_UNREACH;
}
}
if (mca_bml_r2.show_unreach_errors &&
OMPI_ERR_UNREACH == ret) {
char *local, *remote;
orte_ns.get_proc_name_string(&local,
&(ompi_proc_local_proc->proc_name));
orte_ns.get_proc_name_string(&remote,
&(unreach_proc->proc_name));
opal_show_help("help-mca-bml-r2",
"unreachable proc",
true, local, remote, NULL);
free(local);
free(remote);
}
free(new_procs);
return ret;
}
