int mca_pml_yalla_add_procs(struct ompi_proc_t **procs, size_t nprocs)
{
size_t i;
int ret;
void *address;
mxm_conn_h conn;
size_t addrlen;
mxm_error_t error;
if (OMPI_SUCCESS != (ret = mca_pml_base_pml_check_selected("yalla",
procs,
nprocs))) {
return ret;
}
for (i = 0; i < nprocs; ++i) {
ret = recv_ep_address(procs[i], &address, &addrlen);
if (ret < 0) {
return ret;
}
if (procs[i]->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_PML]) {
PML_YALLA_VERBOSE(3, "already connected to proc. %s",
OPAL_NAME_PRINT(procs[i]->super.proc_name));
continue;
}
PML_YALLA_VERBOSE(2, "connecting to proc. %s",
OPAL_NAME_PRINT(procs[i]->super.proc_name));
error = mxm_ep_connect(ompi_pml_yalla.mxm_ep, address, &conn);
free(address);
if (MXM_OK != error) {
PML_YALLA_ERROR("Failed to connect");
return OMPI_ERROR;
}
procs[i]->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_PML] = conn;
}
return OMPI_SUCCESS;
}
int mca_pml_ucx_add_procs(struct ompi_proc_t **procs, size_t nprocs)
{
ucp_address_t *address;
ucs_status_t status;
size_t addrlen;
ucp_ep_h ep;
size_t i;
int ret;
if (OMPI_SUCCESS != (ret = mca_pml_base_pml_check_selected("ucx",
procs,
nprocs))) {
return ret;
}
for (i = 0; i < nprocs; ++i) {
ret = mca_pml_ucx_recv_worker_address(procs[i], &address, &addrlen);
if (ret < 0) {
PML_UCX_ERROR("Failed to receive worker address from proc: %d", procs[i]->super.proc_name.vpid);
return ret;
}
if (procs[i]->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_PML]) {
PML_UCX_VERBOSE(3, "already connected to proc. %d", procs[i]->super.proc_name.vpid);
continue;
}
PML_UCX_VERBOSE(2, "connecting to proc. %d", procs[i]->super.proc_name.vpid);
status = ucp_ep_create(ompi_pml_ucx.ucp_worker, address, &ep);
free(address);
if (UCS_OK != status) {
PML_UCX_ERROR("Failed to connect to proc: %d, %s", procs[i]->super.proc_name.vpid,
ucs_status_string(status));
return OMPI_ERROR;
}
procs[i]->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_PML] = ep;
}
return OMPI_SUCCESS;
}
int
mca_pml_cm_add_procs(struct ompi_proc_t** procs, size_t nprocs)
{
int ret;
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
for (size_t i = 0 ; i < nprocs ; ++i) {
if (procs[i]->super.proc_arch != ompi_proc_local()->super.proc_arch) {
return OMPI_ERR_NOT_SUPPORTED;
}
}
#endif
/* make sure remote procs are using the same PML as us */
if (OMPI_SUCCESS != (ret = mca_pml_base_pml_check_selected("cm",
procs,
nprocs))) {
return ret;
}
ret = OMPI_MTL_CALL(add_procs(ompi_mtl, nprocs, procs));
return ret;
}
ucp_ep_h mca_pml_ucx_add_proc(ompi_communicator_t *comm, int dst)
{
ucp_address_t *address;
ucs_status_t status;
size_t addrlen;
ucp_ep_h ep;
int ret;
ompi_proc_t *proc0 = ompi_comm_peer_lookup(comm, 0);
ompi_proc_t *proc_peer = ompi_comm_peer_lookup(comm, dst);
/* Note, mca_pml_base_pml_check_selected, doesn't use 3rd argument */
if (OMPI_SUCCESS != (ret = mca_pml_base_pml_check_selected("ucx",
&proc0,
dst))) {
return NULL;
}
ret = mca_pml_ucx_recv_worker_address(proc_peer, &address, &addrlen);
if (ret < 0) {
PML_UCX_ERROR("Failed to receive worker address from proc: %d", proc_peer->super.proc_name.vpid);
return NULL;
}
PML_UCX_VERBOSE(2, "connecting to proc. %d", proc_peer->super.proc_name.vpid);
status = ucp_ep_create(ompi_pml_ucx.ucp_worker, address, &ep);
free(address);
if (UCS_OK != status) {
PML_UCX_ERROR("Failed to connect to proc: %d, %s", proc_peer->super.proc_name.vpid,
ucs_status_string(status));
return NULL;
}
proc_peer->proc_endpoints[OMPI_PROC_ENDPOINT_TAG_PML] = ep;
return ep;
}
int mca_pml_ob1_add_procs(ompi_proc_t** procs, size_t nprocs)
{
opal_bitmap_t reachable;
int rc;
opal_list_item_t *item;
if(nprocs == 0)
return OMPI_SUCCESS;
OBJ_CONSTRUCT(&reachable, opal_bitmap_t);
rc = opal_bitmap_init(&reachable, (int)nprocs);
if(OMPI_SUCCESS != rc)
return rc;
/*
* JJH: Disable this in FT enabled builds since
* we use a wrapper PML. It will cause this check to
* return failure as all processes will return the wrapper PML
* component in use instead of the wrapped PML component underneath.
*/
#if OPAL_ENABLE_FT_CR == 0
/* make sure remote procs are using the same PML as us */
if (OMPI_SUCCESS != (rc = mca_pml_base_pml_check_selected("ob1",
procs,
nprocs))) {
return rc;
}
#endif
rc = mca_bml.bml_add_procs( nprocs,
procs,
&reachable );
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
/* Check that values supplied by all initialized btls will work
for us. Note that this is the list of all initialized BTLs,
not the ones used for the just added procs. This is a little
overkill and inaccurate, as we may end up not using the BTL in
question and all add_procs calls after the first one are
duplicating an already completed check. But the final
initialization of the PML occurs before the final
initialization of the BTLs, and iterating through the in-use
BTLs requires iterating over the procs, as the BML does not
expose all currently in use btls. */
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_eager_limit < sizeof(mca_pml_ob1_hdr_t)) {
opal_show_help("help-mpi-pml-ob1.txt", "eager_limit_too_small",
true,
sm->btl_component->btl_version.mca_component_name,
ompi_process_info.nodename,
sm->btl_component->btl_version.mca_component_name,
sm->btl_module->btl_eager_limit,
sm->btl_component->btl_version.mca_component_name,
sizeof(mca_pml_ob1_hdr_t),
sm->btl_component->btl_version.mca_component_name);
rc = OMPI_ERR_BAD_PARAM;
goto cleanup_and_return;
}
#if OPAL_CUDA_SUPPORT_60
/* If size is SIZE_MAX, then we know we want to set this to the minimum possible
* value which is the size of the PML header. */
if (SIZE_MAX == sm->btl_module->btl_cuda_eager_limit) {
sm->btl_module->btl_cuda_eager_limit = sizeof(mca_pml_ob1_hdr_t);
}
if (0 != sm->btl_module->btl_cuda_eager_limit) {
if (sm->btl_module->btl_cuda_eager_limit < sizeof(mca_pml_ob1_hdr_t)) {
opal_show_help("help-mpi-pml-ob1.txt", "cuda_eager_limit_too_small",
true,
sm->btl_component->btl_version.mca_component_name,
ompi_process_info.nodename,
sm->btl_component->btl_version.mca_component_name,
sm->btl_module->btl_cuda_eager_limit,
sm->btl_component->btl_version.mca_component_name,
sizeof(mca_pml_ob1_hdr_t),
sm->btl_component->btl_version.mca_component_name);
rc = OMPI_ERR_BAD_PARAM;
goto cleanup_and_return;
}
}
if (0 == sm->btl_module->btl_cuda_rdma_limit) {
/* All is fine. 0 means to ignore value so set to SIZE_MAX */
sm->btl_module->btl_cuda_rdma_limit = SIZE_MAX;
} else {
if (sm->btl_module->btl_cuda_rdma_limit < sm->btl_module->btl_cuda_eager_limit) {
opal_show_help("help-mpi-pml-ob1.txt", "cuda_rdma_limit_too_small",
true,
sm->btl_component->btl_version.mca_component_name,
ompi_process_info.nodename,
sm->btl_component->btl_version.mca_component_name,
sm->btl_module->btl_cuda_rdma_limit,
sm->btl_component->btl_version.mca_component_name,
sm->btl_module->btl_cuda_eager_limit,
sm->btl_component->btl_version.mca_component_name);
rc = OMPI_ERR_BAD_PARAM;
goto cleanup_and_return;
}
}
#endif /* OPAL_CUDA_SUPPORT_60 */
}
/* TODO: Move these callback registration to another place */
rc = mca_bml.bml_register( MCA_PML_OB1_HDR_TYPE_MATCH,
mca_pml_ob1_recv_frag_callback_match,
NULL );
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
rc = mca_bml.bml_register( MCA_PML_OB1_HDR_TYPE_RNDV,
mca_pml_ob1_recv_frag_callback_rndv,
NULL );
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
rc = mca_bml.bml_register( MCA_PML_OB1_HDR_TYPE_RGET,
mca_pml_ob1_recv_frag_callback_rget,
NULL );
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
rc = mca_bml.bml_register( MCA_PML_OB1_HDR_TYPE_ACK,
mca_pml_ob1_recv_frag_callback_ack,
NULL );
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
rc = mca_bml.bml_register( MCA_PML_OB1_HDR_TYPE_FRAG,
mca_pml_ob1_recv_frag_callback_frag,
NULL );
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
rc = mca_bml.bml_register( MCA_PML_OB1_HDR_TYPE_PUT,
mca_pml_ob1_recv_frag_callback_put,
NULL );
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
rc = mca_bml.bml_register( MCA_PML_OB1_HDR_TYPE_FIN,
mca_pml_ob1_recv_frag_callback_fin,
NULL );
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
/* register error handlers */
rc = mca_bml.bml_register_error(mca_pml_ob1_error_handler);
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
cleanup_and_return:
OBJ_DESTRUCT(&reachable);
return rc;
}
int mca_pml_bfo_add_procs(ompi_proc_t** procs, size_t nprocs)
{
opal_bitmap_t reachable;
int rc;
size_t i;
opal_list_item_t *item;
if(nprocs == 0)
return OMPI_SUCCESS;
/* we don't have any endpoint data we need to cache on the
ompi_proc_t, so set proc_pml to NULL */
for (i = 0 ; i < nprocs ; ++i) {
procs[i]->proc_pml = NULL;
}
OBJ_CONSTRUCT(&reachable, opal_bitmap_t);
rc = opal_bitmap_init(&reachable, (int)nprocs);
if(OMPI_SUCCESS != rc)
return rc;
/*
* JJH: Disable this in FT enabled builds since
* we use a wrapper PML. It will cause this check to
* return failure as all processes will return the wrapper PML
* component in use instead of the wrapped PML component underneath.
*/
#if OPAL_ENABLE_FT_CR == 0
/* make sure remote procs are using the same PML as us */
if (OMPI_SUCCESS != (rc = mca_pml_base_pml_check_selected("bfo",
procs,
nprocs))) {
return rc;
}
#endif
rc = mca_bml.bml_add_procs( nprocs,
procs,
&reachable );
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
/* Check that values supplied by all initialized btls will work
for us. Note that this is the list of all initialized BTLs,
not the ones used for the just added procs. This is a little
overkill and inaccurate, as we may end up not using the BTL in
question and all add_procs calls after the first one are
duplicating an already completed check. But the final
initialization of the PML occurs before the final
initialization of the BTLs, and iterating through the in-use
BTLs requires iterating over the procs, as the BML does not
expose all currently in use btls. */
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_eager_limit < sizeof(mca_pml_bfo_hdr_t)) {
orte_show_help("help-mpi-pml-bfo.txt", "eager_limit_too_small",
true,
sm->btl_component->btl_version.mca_component_name,
orte_process_info.nodename,
sm->btl_component->btl_version.mca_component_name,
sm->btl_module->btl_eager_limit,
sm->btl_component->btl_version.mca_component_name,
sizeof(mca_pml_bfo_hdr_t),
sm->btl_component->btl_version.mca_component_name);
rc = OMPI_ERR_BAD_PARAM;
goto cleanup_and_return;
}
}
/* TODO: Move these callback registration to another place */
rc = mca_bml.bml_register( MCA_PML_BFO_HDR_TYPE_MATCH,
mca_pml_bfo_recv_frag_callback_match,
NULL );
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
rc = mca_bml.bml_register( MCA_PML_BFO_HDR_TYPE_RNDV,
mca_pml_bfo_recv_frag_callback_rndv,
NULL );
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
rc = mca_bml.bml_register( MCA_PML_BFO_HDR_TYPE_RGET,
mca_pml_bfo_recv_frag_callback_rget,
NULL );
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
rc = mca_bml.bml_register( MCA_PML_BFO_HDR_TYPE_ACK,
mca_pml_bfo_recv_frag_callback_ack,
NULL );
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
rc = mca_bml.bml_register( MCA_PML_BFO_HDR_TYPE_FRAG,
mca_pml_bfo_recv_frag_callback_frag,
NULL );
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
rc = mca_bml.bml_register( MCA_PML_BFO_HDR_TYPE_PUT,
mca_pml_bfo_recv_frag_callback_put,
NULL );
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
rc = mca_bml.bml_register( MCA_PML_BFO_HDR_TYPE_FIN,
mca_pml_bfo_recv_frag_callback_fin,
NULL );
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
#if PML_BFO
rc = mca_pml_bfo_register_callbacks();
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
#endif /* PML_BFO */
/* register error handlers */
rc = mca_bml.bml_register_error(mca_pml_bfo_error_handler);
if(OMPI_SUCCESS != rc)
goto cleanup_and_return;
cleanup_and_return:
OBJ_DESTRUCT(&reachable);
return rc;
}
int mca_pml_dr_add_procs(ompi_proc_t** procs, size_t nprocs)
{
opal_bitmap_t reachable;
int rc;
size_t i;
opal_list_item_t *item;
if(nprocs == 0)
return OMPI_SUCCESS;
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
for (i = 0 ; i < nprocs ; ++i) {
if (procs[i]->proc_arch != ompi_proc_local()->proc_arch) {
return OMPI_ERR_NOT_SUPPORTED;
}
}
#endif
/* make sure remote procs are using the same PML as us */
if (OMPI_SUCCESS != (rc = mca_pml_base_pml_check_selected("dr",
procs,
nprocs))) {
return rc;
}
OBJ_CONSTRUCT(&reachable, opal_bitmap_t);
rc = opal_bitmap_init(&reachable, (int)nprocs);
if(OMPI_SUCCESS != rc)
return rc;
/* initialize bml endpoint data */
rc = mca_bml.bml_add_procs(
nprocs,
procs,
&reachable
);
if(OMPI_SUCCESS != rc)
return rc;
/* Check that values supplied by all initialized btls will work
for us. Note that this is the list of all initialized BTLs,
not the ones used for the just added procs. This is a little
overkill and inaccurate, as we may end up not using the BTL in
question and all add_procs calls after the first one are
duplicating an already completed check. But the final
initialization of the PML occurs before the final
initialization of the BTLs, and iterating through the in-use
BTLs requires iterating over the procs, as the BML does not
expose all currently in use btls. */
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_eager_limit < sizeof(mca_pml_dr_hdr_t)) {
orte_show_help("help-mpi-pml-dr.txt", "eager_limit_too_small",
true,
sm->btl_component->btl_version.mca_component_name,
orte_process_info.nodename,
sm->btl_component->btl_version.mca_component_name,
sm->btl_module->btl_eager_limit,
sm->btl_component->btl_version.mca_component_name,
sizeof(mca_pml_dr_hdr_t),
sm->btl_component->btl_version.mca_component_name);
rc = OMPI_ERR_BAD_PARAM;
return rc;
}
}
/* register recv handler */
rc = mca_bml.bml_register(
MCA_BTL_TAG_PML,
mca_pml_dr_recv_frag_callback,
NULL);
if(OMPI_SUCCESS != rc)
return rc;
/* register error handlers */
rc = mca_bml.bml_register_error(mca_pml_dr_error_handler);
if(OMPI_SUCCESS != rc)
return rc;
ompi_free_list_init_new(
&mca_pml_dr.buffers,
sizeof(mca_pml_dr_buffer_t) + mca_pml_dr.eager_limit,
opal_cache_line_size,
OBJ_CLASS(mca_pml_dr_buffer_t),
0,opal_cache_line_size,
0,
mca_pml_dr.free_list_max,
mca_pml_dr.free_list_inc,
NULL);
/* initialize pml endpoint data */
for (i = 0 ; i < nprocs ; ++i) {
int idx;
mca_pml_dr_endpoint_t *endpoint;
endpoint = OBJ_NEW(mca_pml_dr_endpoint_t);
endpoint->proc_ompi = procs[i];
procs[i]->proc_pml = (struct mca_pml_endpoint_t*) endpoint;
MCA_PML_DR_DEBUG(10, (0, "%s:%d: adding endpoint %p to proc_pml %p\n",
__FILE__, __LINE__, (void*)endpoint, (void*)procs[i]));
/* this won't work for comm spawn and other dynamic
processes, but will work for initial job start */
idx = opal_pointer_array_add(&mca_pml_dr.endpoints, (void*) endpoint);
if(OPAL_EQUAL == orte_util_compare_name_fields(ORTE_NS_CMP_ALL,
ORTE_PROC_MY_NAME,
&(endpoint->proc_ompi->proc_name))) {
mca_pml_dr.my_rank = idx;
}
endpoint->local = endpoint->dst = idx;
MCA_PML_DR_DEBUG(10, (0, "%s:%d: setting endpoint->dst to %d\n",
__FILE__, __LINE__, idx));
endpoint->bml_endpoint = procs[i]->proc_bml;
}
for(i = 0; i < nprocs; i++) {
mca_pml_dr_endpoint_t* ep = (mca_pml_dr_endpoint_t*)
opal_pointer_array_get_item(&mca_pml_dr.endpoints, i);
ep->src = mca_pml_dr.my_rank;
}
return rc;
}
