struct mca_sharedfp_base_module_1_0_0_t * mca_sharedfp_sm_component_file_query(mca_io_ompio_file_t *fh, int *priority)
{
int i;
ompi_proc_t *proc;
ompi_communicator_t * comm = fh->f_comm;
int size = ompi_comm_size(comm);
*priority = 0;
/* test, and update priority. All processes have to be
** on a single node.
** original test copied from mca/coll/sm/coll_sm_module.c:
*/
ompi_group_t *group = comm->c_local_group;
for (i = 0; i < size; ++i) {
proc = ompi_group_peer_lookup(group,i);
if (!OPAL_PROC_ON_LOCAL_NODE(proc->proc_flags)){
opal_output(1,"mca_sharedfp_sm_component_file_query: Disqualifying myself: (%d/%s) "
"not all processes are on the same node.",
comm->c_contextid, comm->c_name);
return NULL;
}
}
/* This module can run */
*priority = mca_sharedfp_sm_priority;
return &sm;
}
static void
mca_coll_hierarch_checkfor_sm ( struct ompi_communicator_t *comm, int *color, int *ncount )
{
int i, size;
int lncount=0;
struct ompi_proc_t** procs=NULL;
struct ompi_proc_t* my_proc=NULL;
*color = -1;
size = ompi_comm_size(comm);
my_proc = ompi_proc_local();
procs = comm->c_local_group->grp_proc_pointers;
for ( i = 0 ; i < size ; i++) {
if ( OMPI_CAST_RTE_NAME(&procs[i]->super.proc_name)->jobid == OMPI_CAST_RTE_NAME(&my_proc->super.proc_name)->jobid &&
( OPAL_PROC_ON_LOCAL_NODE(procs[i]->super.proc_flags)) ) {
lncount++;
if ( *color == -1){
*color = i;
}
}
}
/* we need to decrease ncount in order to make the other allreduce/allgather
operations work */
lncount--;
*ncount = lncount;
return;
}
static int vader_add_procs (struct mca_btl_base_module_t* btl,
size_t nprocs, struct ompi_proc_t **procs,
struct mca_btl_base_endpoint_t **peers,
opal_bitmap_t *reachability)
{
mca_btl_vader_component_t *component = &mca_btl_vader_component;
mca_btl_vader_t *vader_btl = (mca_btl_vader_t *) btl;
int32_t proc, local_rank;
ompi_proc_t *my_proc;
int rc;
/* initializion */
/* get pointer to my proc structure */
if (NULL == (my_proc = ompi_proc_local())) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* jump out if there's not someone we can talk to */
if (1 > MCA_BTL_VADER_NUM_LOCAL_PEERS) {
return OMPI_SUCCESS;
}
/* make sure that my local rank has been defined */
if (ORTE_LOCAL_RANK_INVALID == MCA_BTL_VADER_LOCAL_RANK) {
return OMPI_ERROR;
}
if (!vader_btl->btl_inited) {
rc = vader_btl_first_time_init (vader_btl, 1 + MCA_BTL_VADER_NUM_LOCAL_PEERS);
if (rc != OMPI_SUCCESS) {
return rc;
}
}
for (proc = 0, local_rank = 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 ||
!OPAL_PROC_ON_LOCAL_NODE(procs[proc]->proc_flags)) {
peers[proc] = NULL;
continue;
}
if (my_proc != procs[proc]) {
/* add this proc to shared memory accessibility list */
rc = opal_bitmap_set_bit (reachability, proc);
if(OMPI_SUCCESS != rc) {
return rc;
}
}
/* setup endpoint */
peers[proc] = component->endpoints + local_rank;
init_vader_endpoint (peers[proc], procs[proc], local_rank++);
}
return OMPI_SUCCESS;
}
static int _get_local_ranks(mca_scoll_fca_module_t *fca_module)
{
struct oshmem_group_t *comm = fca_module->comm;
oshmem_proc_t* proc;
int i, rank;
/* Count the local ranks */
fca_module->num_local_procs = 0;
for (rank = 0; rank < comm->proc_count; ++rank) {
proc = comm->proc_array[rank];
if (OPAL_PROC_ON_LOCAL_NODE(proc->super.proc_flags)) {
if (proc->super.proc_name.vpid == (uint32_t) fca_module->rank) {
fca_module->local_proc_idx = fca_module->num_local_procs;
}
++fca_module->num_local_procs;
}
}
/* Make a list of local ranks */
fca_module->local_ranks = calloc(fca_module->num_local_procs,
sizeof *fca_module->local_ranks);
if (!fca_module->local_ranks) {
FCA_ERROR("Failed to allocate memory for %d local ranks",
fca_module->num_local_procs);
return OSHMEM_ERROR;
}
i = 0;
for (rank = 0; rank < comm->proc_count; ++rank) {
proc = comm->proc_array[rank];
if (OPAL_PROC_ON_LOCAL_NODE(proc->super.proc_flags)) {
fca_module->local_ranks[i++] = rank;
}
}
FCA_MODULE_VERBOSE(fca_module,
3,
"i am %d/%d",
fca_module->local_proc_idx, fca_module->num_local_procs);
return OSHMEM_SUCCESS;
}
/**
* Fills local rank information in fca_module.
*/
static int __get_local_ranks(mca_coll_fca_module_t *fca_module)
{
ompi_communicator_t *comm = fca_module->comm;
ompi_proc_t* proc;
int i, rank;
/* Count the local ranks */
fca_module->num_local_procs = 0;
for (rank = 0; rank < ompi_comm_size(comm); ++rank) {
proc = __local_rank_lookup(comm, rank);
if (OPAL_PROC_ON_LOCAL_NODE(proc->super.proc_flags)) {
if (rank == fca_module->rank) {
fca_module->local_proc_idx = fca_module->num_local_procs;
}
++fca_module->num_local_procs;
}
}
/* Make a list of local ranks */
fca_module->local_ranks = calloc(fca_module->num_local_procs,
sizeof *fca_module->local_ranks);
if (!fca_module->local_ranks) {
FCA_ERROR("Failed to allocate memory for %d local ranks",
fca_module->num_local_procs);
return OMPI_ERROR;
}
i = 0;
for (rank = 0; rank < ompi_comm_size(comm); ++rank) {
proc = __local_rank_lookup(comm, rank);
if (OPAL_PROC_ON_LOCAL_NODE(proc->super.proc_flags)) {
fca_module->local_ranks[i++] = rank;
}
}
FCA_MODULE_VERBOSE(fca_module, 3, "i am %d/%d", fca_module->local_proc_idx,
fca_module->num_local_procs);
return OMPI_SUCCESS;
}
static bool have_local_peers(ompi_group_t *group, size_t size)
{
size_t i;
ompi_proc_t *proc;
for (i = 0; i < size; ++i) {
proc = ompi_group_peer_lookup(group,i);
if (!OPAL_PROC_ON_LOCAL_NODE(proc->proc_flags)) {
return false;
}
}
return true;
}
static void unpack_remote_mkeys(opal_buffer_t *msg, int remote_pe)
{
int32_t cnt;
int32_t n;
int32_t tr_id;
int i;
oshmem_proc_t *proc;
proc = oshmem_proc_group_find(oshmem_group_all, remote_pe);
cnt = 1;
opal_dss.unpack(msg, &n, &cnt, OPAL_UINT32);
for (i = 0; i < n; i++) {
cnt = 1;
opal_dss.unpack(msg, &tr_id, &cnt, OPAL_UINT32);
cnt = 1;
opal_dss.unpack(msg,
&memheap_oob.mkeys[tr_id].va_base,
&cnt,
OPAL_UINT64);
if (0 == memheap_oob.mkeys[tr_id].va_base) {
cnt = 1;
opal_dss.unpack(msg, &memheap_oob.mkeys[tr_id].u.key, &cnt, OPAL_UINT64);
if (OPAL_PROC_ON_LOCAL_NODE(proc->super.proc_flags)) {
memheap_attach_segment(&memheap_oob.mkeys[tr_id], tr_id);
}
} else {
cnt = 1;
opal_dss.unpack(msg, &memheap_oob.mkeys[tr_id].len, &cnt, OPAL_UINT16);
if (0 < memheap_oob.mkeys[tr_id].len) {
memheap_oob.mkeys[tr_id].u.data = malloc(memheap_oob.mkeys[tr_id].len);
if (NULL == memheap_oob.mkeys[tr_id].u.data) {
MEMHEAP_ERROR("Failed allocate %d bytes", memheap_oob.mkeys[tr_id].len);
oshmem_shmem_abort(-1);
}
cnt = memheap_oob.mkeys[tr_id].len;
opal_dss.unpack(msg, memheap_oob.mkeys[tr_id].u.data, &cnt, OPAL_BYTE);
MCA_SPML_CALL(rmkey_unpack(&memheap_oob.mkeys[tr_id], remote_pe));
} else {
memheap_oob.mkeys[tr_id].u.key = MAP_SEGMENT_SHM_INVALID;
}
}
MEMHEAP_VERBOSE(5,
"tr_id: %d %s",
tr_id, mca_spml_base_mkey2str(&memheap_oob.mkeys[tr_id]));
}
}
/*
* Initial query function that is invoked during MPI_INIT, allowing
* this component to disqualify itself if it doesn't support the
* required level of thread support. This function is invoked exactly
* once.
*/
int mca_coll_sm_init_query(bool enable_progress_threads,
bool enable_mpi_threads)
{
ompi_proc_t *my_proc, **procs;
size_t i, size;
/* See if there are other procs in my job on this node. If not,
then don't bother going any further. */
if (NULL == (my_proc = ompi_proc_local()) ||
NULL == (procs = ompi_proc_all(&size))) {
opal_output_verbose(10, ompi_coll_base_framework.framework_output,
"coll:sm:init_query: weirdness on procs; disqualifying myself");
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* if no session directory was created, then we cannot be used */
if (NULL == ompi_process_info.job_session_dir) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
if (size <= 1) {
opal_output_verbose(10, ompi_coll_base_framework.framework_output,
"coll:sm:init_query: comm size too small; disqualifying myself");
free(procs);
return OMPI_ERR_NOT_AVAILABLE;
}
for (i = 0; i < size; ++i) {
if (procs[i] != my_proc &&
procs[i]->proc_name.jobid == my_proc->proc_name.jobid &&
OPAL_PROC_ON_LOCAL_NODE(procs[i]->proc_flags)) {
break;
}
}
free(procs);
if (i >= size) {
opal_output_verbose(10, ompi_coll_base_framework.framework_output,
"coll:sm:init_query: no other local procs; disqualifying myself");
return OMPI_ERR_NOT_AVAILABLE;
}
/* Don't do much here because we don't really want to allocate any
shared memory until this component is selected to be used. */
opal_output_verbose(10, ompi_coll_base_framework.framework_output,
"coll:sm:init_query: pick me! pick me!");
return OMPI_SUCCESS;
}
static int have_remote_peers(struct oshmem_group_t *group,
size_t size,
int *local_peers)
{
struct oshmem_proc_t *proc;
size_t i;
int ret;
*local_peers = 0;
ret = 0;
for (i = 0; i < size; ++i) {
proc = group->proc_array[i];
if (OPAL_PROC_ON_LOCAL_NODE(proc->super.proc_flags)) {
++*local_peers;
} else {
ret = 1;
}
}
return ret;
}
bool ompi_group_have_remote_peers (ompi_group_t *group)
{
for (int i = 0 ; i < group->grp_proc_count ; ++i) {
ompi_proc_t *proc = NULL;
#if OMPI_GROUP_SPARSE
proc = ompi_group_peer_lookup (group, i);
#else
proc = ompi_group_get_proc_ptr_raw (group, i);
if (ompi_proc_is_sentinel (proc)) {
/* the proc must be stored in the group or cached in the proc
* hash table if the process resides in the local node
* (see ompi_proc_complete_init) */
return true;
}
#endif
if (!OPAL_PROC_ON_LOCAL_NODE(proc->super.proc_flags)) {
return true;
}
}
return false;
}
char* opal_hwloc_base_print_locality(opal_hwloc_locality_t locality)
{
opal_hwloc_print_buffers_t *ptr;
int idx;
ptr = opal_hwloc_get_print_buffer();
if (NULL == ptr) {
return opal_hwloc_print_null;
}
/* cycle around the ring */
if (OPAL_HWLOC_PRINT_NUM_BUFS == ptr->cntr) {
ptr->cntr = 0;
}
idx = 0;
if (OPAL_PROC_ON_LOCAL_CLUSTER(locality)) {
ptr->buffers[ptr->cntr][idx++] = 'C';
ptr->buffers[ptr->cntr][idx++] = 'L';
ptr->buffers[ptr->cntr][idx++] = ':';
}
if (OPAL_PROC_ON_LOCAL_CU(locality)) {
ptr->buffers[ptr->cntr][idx++] = 'C';
ptr->buffers[ptr->cntr][idx++] = 'U';
ptr->buffers[ptr->cntr][idx++] = ':';
}
if (OPAL_PROC_ON_LOCAL_NODE(locality)) {
ptr->buffers[ptr->cntr][idx++] = 'N';
ptr->buffers[ptr->cntr][idx++] = ':';
}
if (OPAL_PROC_ON_LOCAL_BOARD(locality)) {
ptr->buffers[ptr->cntr][idx++] = 'B';
ptr->buffers[ptr->cntr][idx++] = ':';
}
if (OPAL_PROC_ON_LOCAL_NUMA(locality)) {
ptr->buffers[ptr->cntr][idx++] = 'N';
ptr->buffers[ptr->cntr][idx++] = 'u';
ptr->buffers[ptr->cntr][idx++] = ':';
}
if (OPAL_PROC_ON_LOCAL_SOCKET(locality)) {
ptr->buffers[ptr->cntr][idx++] = 'S';
ptr->buffers[ptr->cntr][idx++] = ':';
}
if (OPAL_PROC_ON_LOCAL_L3CACHE(locality)) {
ptr->buffers[ptr->cntr][idx++] = 'L';
ptr->buffers[ptr->cntr][idx++] = '3';
ptr->buffers[ptr->cntr][idx++] = ':';
}
if (OPAL_PROC_ON_LOCAL_L2CACHE(locality)) {
ptr->buffers[ptr->cntr][idx++] = 'L';
ptr->buffers[ptr->cntr][idx++] = '2';
ptr->buffers[ptr->cntr][idx++] = ':';
}
if (OPAL_PROC_ON_LOCAL_L1CACHE(locality)) {
ptr->buffers[ptr->cntr][idx++] = 'L';
ptr->buffers[ptr->cntr][idx++] = '1';
ptr->buffers[ptr->cntr][idx++] = ':';
}
if (OPAL_PROC_ON_LOCAL_CORE(locality)) {
ptr->buffers[ptr->cntr][idx++] = 'C';
ptr->buffers[ptr->cntr][idx++] = ':';
}
if (OPAL_PROC_ON_LOCAL_HWTHREAD(locality)) {
ptr->buffers[ptr->cntr][idx++] = 'H';
ptr->buffers[ptr->cntr][idx++] = 'w';
ptr->buffers[ptr->cntr][idx++] = 't';
ptr->buffers[ptr->cntr][idx++] = ':';
}
if (0 < idx) {
ptr->buffers[ptr->cntr][idx-1] = '\0';
} else if (OPAL_PROC_NON_LOCAL & locality) {
ptr->buffers[ptr->cntr][idx++] = 'N';
ptr->buffers[ptr->cntr][idx++] = 'O';
ptr->buffers[ptr->cntr][idx++] = 'N';
ptr->buffers[ptr->cntr][idx++] = '\0';
} else {
/* must be an unknown locality */
ptr->buffers[ptr->cntr][idx++] = 'U';
ptr->buffers[ptr->cntr][idx++] = 'N';
ptr->buffers[ptr->cntr][idx++] = 'K';
ptr->buffers[ptr->cntr][idx++] = '\0';
}
return ptr->buffers[ptr->cntr];
}
/* This routine is used to find the list of procs that run on the
** same host as the calling process.
*/
static mca_sbgp_base_module_t *mca_sbgp_basesmuma_select_procs(struct ompi_proc_t ** procs,
int n_procs_in,
struct ompi_communicator_t *comm,
char *key,
void *output_data
)
{
/* local variables */
int cnt,proc,local,last_local_proc;
mca_sbgp_basesmuma_module_t *module;
module=OBJ_NEW(mca_sbgp_basesmuma_module_t);
if (!module ) {
return NULL;
}
module->super.group_size=0;
module->super.group_comm = comm;
module->super.group_list = NULL;
module->super.group_net = OMPI_SBGP_MUMA;
for (proc = 0, cnt = 0, last_local_proc = 0 ; proc < n_procs_in ; ++proc) {
local = OPAL_PROC_ON_LOCAL_NODE(procs[proc]->super.proc_flags);
if (local) {
last_local_proc = proc;
cnt++;
}
}
/* if no other local procs found skip to end */
if( 2 > cnt ) {
/* There's always at least one - namely myself */
assert(1 == cnt);
module->super.group_size = 1;
module->super.group_list = (int *) malloc (sizeof (int));
module->super.group_list[0] = last_local_proc;
/* let ml handle this case */
goto OneLocalPeer;
}
/* generate list of local ranks */
module->super.group_size=cnt;
if( cnt > 0 ) {
module->super.group_list=(int *)malloc(sizeof(int)*cnt);
if(NULL == module->super.group_list){
goto Error;
}
}
for (proc = 0, cnt = 0 ; proc < n_procs_in ; ++proc) {
local = OPAL_PROC_ON_LOCAL_NODE(procs[proc]->super.proc_flags);
if( local ) {
module->super.group_list[cnt++] = proc;
}
}
OneLocalPeer:
/* successful completion */
return (mca_sbgp_base_module_t *)module;
/* return with error */
Error:
/* clean up */
if( NULL != module->super.group_list ) {
free(module->super.group_list);
module->super.group_list=NULL;
}
OBJ_RELEASE(module);
return NULL;
}
int ompi_mtl_mxm_module_init(void)
{
#if MXM_API < MXM_VERSION(2,0)
ompi_mtl_mxm_ep_conn_info_t ep_info;
#endif
void *ep_address;
size_t ep_address_len;
mxm_error_t err;
uint32_t jobid;
uint64_t mxlr;
ompi_proc_t **procs;
unsigned ptl_bitmap;
size_t totps, proc;
int lr, nlps;
int rc;
mxlr = 0;
lr = -1;
jobid = ompi_mtl_mxm_get_job_id();
if (0 == jobid) {
MXM_ERROR("Failed to generate jobid");
return OMPI_ERROR;
}
if (NULL == (procs = ompi_proc_world(&totps))) {
MXM_ERROR("Unable to obtain process list");
return OMPI_ERROR;
}
if (totps < (size_t)ompi_mtl_mxm.mxm_np) {
MXM_VERBOSE(1, "MXM support will be disabled because of total number "
"of processes (%lu) is less than the minimum set by the "
"mtl_mxm_np MCA parameter (%u)", totps, ompi_mtl_mxm.mxm_np);
return OMPI_ERR_NOT_SUPPORTED;
}
MXM_VERBOSE(1, "MXM support enabled");
if (ORTE_NODE_RANK_INVALID == (lr = ompi_process_info.my_node_rank)) {
MXM_ERROR("Unable to obtain local node rank");
return OMPI_ERROR;
}
nlps = ompi_process_info.num_local_peers + 1;
for (proc = 0; proc < totps; proc++) {
if (OPAL_PROC_ON_LOCAL_NODE(procs[proc]->proc_flags)) {
mxlr = max(mxlr, procs[proc]->proc_name.vpid);
}
}
/* Setup the endpoint options and local addresses to bind to. */
#if MXM_API < MXM_VERSION(2,0)
ptl_bitmap = ompi_mtl_mxm.mxm_ctx_opts->ptl_bitmap;
#else
ptl_bitmap = 0;
#endif
/* Open MXM endpoint */
err = ompi_mtl_mxm_create_ep(ompi_mtl_mxm.mxm_context, &ompi_mtl_mxm.ep,
ptl_bitmap, lr, jobid, mxlr, nlps);
if (MXM_OK != err) {
opal_show_help("help-mtl-mxm.txt", "unable to create endpoint", true,
mxm_error_string(err));
return OMPI_ERROR;
}
/*
* Get address for each PTL on this endpoint, and share it with other ranks.
*/
#if MXM_API < MXM_VERSION(2,0)
if ((ptl_bitmap & MXM_BIT(MXM_PTL_SELF)) &&
OMPI_SUCCESS != ompi_mtl_mxm_get_ep_address(&ep_info, MXM_PTL_SELF)) {
return OMPI_ERROR;
}
if ((ptl_bitmap & MXM_BIT(MXM_PTL_RDMA)) &&
OMPI_SUCCESS != ompi_mtl_mxm_get_ep_address(&ep_info, MXM_PTL_RDMA)) {
return OMPI_ERROR;
}
if ((ptl_bitmap & MXM_BIT(MXM_PTL_SHM)) &&
OMPI_SUCCESS != ompi_mtl_mxm_get_ep_address(&ep_info, MXM_PTL_SHM)) {
return OMPI_ERROR;
}
ep_address = &ep_info;
ep_address_len = sizeof(ep_info);
#else
rc = ompi_mtl_mxm_get_ep_address(&ep_address, &ep_address_len);
if (OMPI_SUCCESS != rc) {
return rc;
}
#endif
rc = ompi_mtl_mxm_send_ep_address(ep_address, ep_address_len);
if (OMPI_SUCCESS != rc) {
MXM_ERROR("Modex session failed.");
return rc;
}
#if MXM_API >= MXM_VERSION(2,0)
free(ep_address);
#endif
/* Register the MXM progress function */
opal_progress_register(ompi_mtl_mxm_progress);
#if MXM_API >= MXM_VERSION(2,0)
if (ompi_mtl_mxm.using_mem_hooks) {
opal_mem_hooks_register_release(ompi_mtl_mxm_mem_release_cb, NULL);
}
#endif
return OMPI_SUCCESS;
}
static mca_sbgp_base_module_t *mca_sbgp_basesmsocket_select_procs(struct ompi_proc_t ** procs,
int n_procs_in,
struct ompi_communicator_t *comm,
char *key,
void *output_data
)
{
/* local variables */
mca_sbgp_basesmsocket_module_t *module;
int ret;
int my_socket_index;
int proc, cnt, local, n_local_peers, my_rank;
ompi_proc_t* my_proc;
int *local_ranks_in_comm=NULL;
int *socket_info=NULL, my_socket_info;
int i_cnt, lp_cnt, my_local_index = -1, comm_size=ompi_comm_size(comm);
/* initialize data */
output_data=NULL;
my_rank=ompi_comm_rank(comm);
my_proc=ompi_comm_peer_lookup(comm,my_rank);
/*create a new module*/
module=OBJ_NEW(mca_sbgp_basesmsocket_module_t);
if (!module ) {
return NULL;
}
module->super.group_size=0;
module->super.group_comm = comm;
module->super.group_list = NULL;
module->super.group_net = OMPI_SBGP_SOCKET;
/* test to see if process is bound */
if( OPAL_BIND_TO_NONE == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy) ) {
/* pa affinity not set, so socket index will be set to -1 */
my_socket_index=-1;
/*debug print*/
/* */
BASESMSOCKET_VERBOSE(10, ("[%d] FAILED to set basesmsocket group, processes are not bound!!!\n",my_rank));
/*end debug*/
goto NoLocalPeers;
} else {
my_socket_index=-1;
/* this should find my logical socket id which is the socket id we want
* physical socket ids are not necessarily unique, logical ones, as defined
* by the hwloc API are unique.
*/
if( OMPI_SUCCESS != mca_sbgp_map_to_logical_socket_id(&my_socket_index)) {
BASESMSOCKET_VERBOSE(10, ("[%d] FAILED to set basesmsocket group !!!\n",my_rank));
goto NoLocalPeers;
}
}
/* Debug prints */
/*
{
fprintf(stderr,"Number of processors per node: %d\n",num_processors);
fprintf(stderr,"I am rank %d and my socket index is %d\n and my core index is %d\n",my_rank,my_socket_index,core_index);
fprintf(stderr,"n_proc_in = %d\n",n_procs_in);
fprintf(stderr,"\n");
fflush(stderr);
}
end debug prints */
/*get my socket index*/
cnt=0;
for( proc=0 ; proc < n_procs_in ; proc++) {
local=OPAL_PROC_ON_LOCAL_NODE(procs[proc]->proc_flags);
if( local ) {
cnt++;
}
}
/*debug print */
/*
fprintf(stderr,"Number of local processors %d\n",cnt);
end debug print*/
/* if no other local procs found skip to end */
if( 1 >= cnt ) {
goto NoLocalPeers;
}
/* allocate structure to hold the list of local ranks */
local_ranks_in_comm=(int *)malloc(sizeof(int)*cnt);
if(NULL == local_ranks_in_comm ) {
goto Error;
}
/* figure out which ranks from the input communicator - comm - will
* particiapte in the local socket determination.
*/
n_local_peers=0;
i_cnt=0;
for( proc = 0; proc < n_procs_in; proc++) {
local = OPAL_PROC_ON_LOCAL_NODE(procs[proc]->proc_flags);
if ( local ) {
/* set the rank within the on-host ranks - this will be used for tha
* allgather
*/
if( my_proc == procs[proc] ) {
my_local_index=n_local_peers;
}
/* find the rank of the current proc in comm. We take advantage
* of the fact that ranks in a group have the same relative
* ordering as they do within the communicator.
*/
for( lp_cnt=proc; lp_cnt < comm_size ; lp_cnt++ ) {
if(procs[proc] == ompi_comm_peer_lookup(comm,lp_cnt) ) {
local_ranks_in_comm[i_cnt]=lp_cnt;
/* lp_cnt has alrady been checked */
i_cnt++;
/* found the corresponding rank in comm, so don't need
* to search any more */
break;
}
/*i_cnt++;*/
/*fprintf(stderr,"QQQ i_cnt %d \n",i_cnt);*/
}
n_local_peers++;
}
}
/*fprintf(stderr,"YYY n_local_peers %d\n",n_local_peers);*/
socket_info=(int *)malloc(sizeof(int)*n_local_peers);
/*fprintf(stderr,"XXX got socket info\n");*/
if(NULL == socket_info ) {
goto Error;
}
my_socket_info=my_socket_index;
/* Allgather data over the communicator */
ret=comm_allgather_pml(&my_socket_info, socket_info, 1,
MPI_INT, my_local_index, n_local_peers, local_ranks_in_comm,comm);
if (OMPI_SUCCESS != ret ) {
BASESMSOCKET_VERBOSE(10, ("comm_allgather_pml returned error %d\n",ret));
return NULL;
}
/*allocate memory to the group_list probably an overestimation
of the necessary resources */
module->super.group_list=(int *)malloc(sizeof(int)*cnt);
if(NULL == module->super.group_list) {
goto Error;
}
/* figure out who is sharing the same socket */
cnt=0;
for (proc = 0; proc < n_local_peers; proc++) {
int rem_rank=local_ranks_in_comm[proc];
int rem_socket_index=socket_info[proc];
/*Populate the list*/
if (rem_socket_index == my_socket_index) {
module->super.group_list[cnt]=rem_rank;
cnt++;
}
}
module->super.group_size=cnt;
#if 0
/*debug print*/
{
int ii;
fprintf(stderr,"Ranks per socket: %d\n",cnt);
fprintf(stderr,"Socket %d owns ranks: ", my_socket_index);
for (ii=0; ii < cnt; ii++)
fprintf(stderr,"%d ",module->super.group_list[ii]);
fprintf(stderr,"\n");
fflush(stderr);
}
#endif
/* end debug*/
/*Free resources*/
free(local_ranks_in_comm);
free(socket_info);
/*Return the module*/
return (mca_sbgp_base_module_t *) module;
NoLocalPeers:
/* nothing to store, so just free the module and return */
/*fprintf(stderr,"No local socket peers\n");*/
/*free(module);*/
if(socket_info) {
free(socket_info);
socket_info=NULL;
}
if(local_ranks_in_comm) {
free(local_ranks_in_comm);
}
OBJ_RELEASE(module);
return NULL;
Error:
/*clean up*/
if( NULL != module->super.group_list) {
free(module->super.group_list);
module->super.group_list=NULL;
}
if(socket_info) {
free(socket_info);
socket_info=NULL;
}
if(local_ranks_in_comm) {
free(local_ranks_in_comm);
}
OBJ_RELEASE(module);
return NULL;
}
static mca_mtl_base_module_t*
ompi_mtl_psm_component_init(bool enable_progress_threads,
bool enable_mpi_threads)
{
psm_error_t err;
int rc;
int verno_major = PSM_VERNO_MAJOR;
int verno_minor = PSM_VERNO_MINOR;
ompi_proc_t *my_proc, **procs;
size_t num_total_procs, proc;
int local_rank = -1, num_local_procs = 0;
/* Compute the total number of processes on this host and our local rank
* on that node. We need to provide PSM with these values so it can
* allocate hardware contexts appropriately across processes.
*/
if ((rc = ompi_proc_refresh()) != OMPI_SUCCESS) {
return NULL;
}
my_proc = ompi_proc_local();
if (NULL == (procs = ompi_proc_world(&num_total_procs))) {
return NULL;
}
for (proc = 0; proc < num_total_procs; proc++) {
if (my_proc == procs[proc]) {
local_rank = num_local_procs++;
continue;
}
if (OPAL_PROC_ON_LOCAL_NODE(procs[proc]->proc_flags)) {
num_local_procs++;
}
}
assert(local_rank >= 0 && num_local_procs > 0);
free(procs);
err = psm_error_register_handler(NULL /* no ep */,
PSM_ERRHANDLER_NOP);
if (err) {
opal_output(0, "Error in psm_error_register_handler (error %s)\n",
psm_error_get_string(err));
return NULL;
}
#if PSM_VERNO >= 0x010c
/* Set infinipath debug level */
err = psm_setopt(PSM_COMPONENT_CORE, 0, PSM_CORE_OPT_DEBUG,
(const void*) &ompi_mtl_psm.debug_level,
sizeof(unsigned));
if (err) {
/* Non fatal error. Can continue */
orte_show_help("help-mtl-psm.txt",
"psm init", false,
psm_error_get_string(err));
}
#endif
/* Only allow for shm and ipath devices in 2.0 and earlier releases
* (unless the user overrides the setting).
*/
if (PSM_VERNO >= 0x0104) {
setenv("PSM_DEVICES", "self,shm,ipath", 0);
}
else {
setenv("PSM_DEVICES", "shm,ipath", 0);
}
err = psm_init(&verno_major, &verno_minor);
if (err) {
orte_show_help("help-mtl-psm.txt",
"psm init", true,
psm_error_get_string(err));
return NULL;
}
/* Complete PSM initialization */
ompi_mtl_psm_module_init(local_rank, num_local_procs);
ompi_mtl_psm.super.mtl_request_size =
sizeof(mca_mtl_psm_request_t) -
sizeof(struct mca_mtl_request_t);
return &ompi_mtl_psm.super;
}
static mca_sbgp_base_module_t *mca_sbgp_basesmsocket_select_procs(struct ompi_proc_t ** procs,
int n_procs_in,
struct ompi_communicator_t *comm,
char *key,
void *output_data
)
{
/* local variables */
mca_sbgp_basesmsocket_module_t *module;
/*
opal_buffer_t* sbuffer = OBJ_NEW(opal_buffer_t);
opal_buffer_t* rbuffer = OBJ_NEW(opal_buffer_t);
*/
opal_paffinity_base_cpu_set_t my_cpu_set;
bool bound;
int ret;
int num_processors;
int socket_tmp;
int my_socket_index;
int core_index=-1;
int proc, cnt, local, n_local_peers, my_index, my_rank;
ompi_proc_t* my_proc;
int *local_ranks_in_comm=NULL;
int *socket_info=NULL, my_socket_info;
int i_cnt, lp_cnt, my_local_index, comm_size=ompi_comm_size(comm);
/* initialize data */
output_data=NULL;
my_rank=ompi_comm_rank(comm);
my_proc=ompi_comm_peer_lookup(comm,my_rank);
for( proc=0 ; proc < n_procs_in ; proc++) {
if( procs[proc]==my_proc) {
my_index=proc;
}
}
/*create a new module*/
module=OBJ_NEW(mca_sbgp_basesmsocket_module_t);
if (!module ) {
return NULL;
}
module->super.group_size=0;
module->super.group_comm = comm;
module->super.group_list = NULL;
module->super.group_net = OMPI_SBGP_SOCKET;
/*
** get my process affinity information
** */
/* get the number of processors on this node */
ret=opal_paffinity_base_get_processor_info(&num_processors);
/* get process affinity mask */
OPAL_PAFFINITY_CPU_ZERO(my_cpu_set);
ret=opal_paffinity_base_get(&my_cpu_set);
OPAL_PAFFINITY_PROCESS_IS_BOUND(my_cpu_set,&bound);
/*debug process affinity*/
/*
{
ret=opal_paffinity_base_get_socket_info(&num_socket);
fprintf(stderr,"Number of sockets %d\n",num_socket);
fprintf(stderr,"Test if rank %d is bound %d\n", my_rank, bound);
fprintf(stderr,"return from opal_paffinity_base_get: %d\n\n",ret);
fprintf(stderr,"bitmask elements: ");
unsigned int long jj;
for(jj=0; jj < OPAL_PAFFINITY_BITMASK_NUM_ELEMENTS; jj++)
fprintf(stderr," %d ",my_cpu_set.bitmask[jj]);
fprintf(stderr,"\n");
fflush(stderr);
}
end debug process affinity*/
if( !bound ) {
/* pa affinity not set, so socket index will be set to -1 */
my_socket_index=-1;
/*debug print*/
/* */
fprintf(stderr,"[%d]FAILED to set basesmsocket group !!!\n",my_rank);
fflush(stderr);
/*end debug*/
goto NoLocalPeers;
} else {
my_socket_index=-1;
/* loop over number of processors */
for ( proc=0 ; proc < num_processors ; proc++ ) {
if (OPAL_PAFFINITY_CPU_ISSET(proc,my_cpu_set)) {
ret=opal_paffinity_base_get_map_to_socket_core(proc,&socket_tmp,&core_index);
if( my_socket_index != socket_tmp ) {
my_socket_index=socket_tmp;
break;
}
}
} /* end of proc loop */
}
/* Debug prints */
/*
{
fprintf(stderr,"Number of processors per node: %d\n",num_processors);
fprintf(stderr,"I am rank %d and my socket index is %d\n and my core index is %d\n",my_rank,my_socket_index,core_index);
fprintf(stderr,"n_proc_in = %d\n",n_procs_in);
fprintf(stderr,"\n");
fflush(stderr);
}
end debug prints */
/*get my socket index*/
cnt=0;
for( proc=0 ; proc < n_procs_in ; proc++) {
local=OPAL_PROC_ON_LOCAL_NODE(procs[proc]->proc_flags);
if( local ) {
cnt++;
}
}
/*debug print */
/*
fprintf(stderr,"Number of local processors %d\n",cnt);
end debug print*/
/* if no other local procs found skip to end */
if( 1 >= cnt ) {
goto NoLocalPeers;
}
#if 0
int *local_ranks_in_comm;
int32_t *socket_info, *my_socket_info;
int my_local_index;
#endif
/* allocate structure to hold the list of local ranks */
local_ranks_in_comm=(int *)malloc(sizeof(int)*cnt);
if(NULL == local_ranks_in_comm ) {
goto Error;
}
/* figure out which ranks from the input communicator - comm - will
* particiapte in the local socket determination.
*/
n_local_peers=0;
i_cnt=0;
for( proc = 0; proc < n_procs_in; proc++) {
local = OPAL_PROC_ON_LOCAL_NODE(procs[proc]->proc_flags);
if ( local ) {
/* set the rank within the on-host ranks - this will be used for tha
* allgather
*/
if( my_proc == procs[proc] ) {
my_local_index=n_local_peers;
}
/* find the rank of the current proc in comm. We take advantage
* of the fact that ranks in a group have the same relative
* ordering as they do within the communicator.
*/
#if 1
/*for( lp_cnt=i_cnt; lp_cnt < comm_size ; lp_cnt++ ) {*/
for( lp_cnt=proc; lp_cnt < comm_size ; lp_cnt++ ) {
if(procs[proc] == ompi_comm_peer_lookup(comm,lp_cnt) ) {
local_ranks_in_comm[i_cnt]=lp_cnt;
/* lp_cnt has alrady been checked */
i_cnt++;
/* found the corresponding rank in comm, so don't need
* to search any more */
break;
}
/*i_cnt++;*/
/*fprintf(stderr,"QQQ i_cnt %d \n",i_cnt);*/
}
#endif
n_local_peers++;
}
}
/*fprintf(stderr,"YYY n_local_peers %d\n",n_local_peers);*/
socket_info=(int *)malloc(sizeof(int)*n_local_peers);
/*fprintf(stderr,"XXX got socket info\n");*/
if(NULL == socket_info ) {
goto Error;
}
my_socket_info=my_socket_index;
/* Allgather data over the communicator */
ret=comm_allgather_pml(&my_socket_info, socket_info, 1,
MPI_INT, my_local_index, n_local_peers, local_ranks_in_comm,comm);
if (OMPI_SUCCESS != ret ) {
fprintf(stderr," comm_allgather_pml returned error %d \n", ret);
fflush(stderr);
return NULL;
}
/*allocate memory to the group_list probably an overestimation
of the necessary resources */
module->super.group_list=(int *)malloc(sizeof(int)*cnt);
if(NULL == module->super.group_list) {
goto Error;
}
/* figure out who is sharing the same socket */
cnt=0;
for (proc = 0; proc < n_local_peers; proc++) {
int rem_rank=local_ranks_in_comm[proc];
int rem_socket_index=socket_info[proc];
/*Populate the list*/
if (rem_socket_index == my_socket_index) {
module->super.group_list[cnt]=rem_rank;
cnt++;
}
}
module->super.group_size=cnt;
/*debug print*/
/*
{
int ii;
fprintf(stderr,"Ranks per socket: %d\n",cnt);
fprintf(stderr,"Socket %d owns ranks: ", my_socket_index);
for (ii=0; ii < cnt; ii++)
fprintf(stderr,"%d ",module->super.group_list[ii]);
fprintf(stderr,"\n");
fflush(stderr);
}
{
cpu_set_t set;
unsigned int len = sizeof(set);
int i;
unsigned long mask = 0;
CPU_ZERO(&set);
if (sched_getaffinity(0, len, &set) < 0) {
perror("sched_getaffinity");
return -1;
}
for (i = 0; i < CPU_SETSIZE; i++) {
int cpu = CPU_ISSET(i, &set);
if (cpu) {
mask |= 1<< i;
}
}
opal_output(0,"%d: my affinity mask is: %08lx\n", my_local_index,mask);
}
end debug*/
/*Free resources*/
free(local_ranks_in_comm);
free(socket_info);
/*Return the module*/
return (mca_sbgp_base_module_t *) module;
NoLocalPeers:
/* nothing to store, so just free the module and return */
/*fprintf(stderr,"No local socket peers\n");*/
/*free(module);*/
if(socket_info) {
free(socket_info);
socket_info=NULL;
}
if(local_ranks_in_comm) {
free(local_ranks_in_comm);
}
OBJ_RELEASE(module);
return NULL;
Error:
/*clean up*/
if( NULL != module->super.group_list) {
free(module->super.group_list);
module->super.group_list=NULL;
}
if(socket_info) {
free(socket_info);
socket_info=NULL;
}
if(local_ranks_in_comm) {
free(local_ranks_in_comm);
}
OBJ_RELEASE(module);
return NULL;
}
static int hier_allgather(opal_buffer_t *sbuf, opal_buffer_t *rbuf)
{
int rc=ORTE_SUCCESS;
opal_list_item_t *item;
orte_namelist_t *nm;
opal_buffer_t tmp_buf;
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base.output,
"%s grpcomm:hier entering allgather",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* have I initialized my local info? */
if (!coll_initialized) {
orte_process_name_t proc;
orte_vpid_t v;
/* get my local rank so I can locally cache it */
my_local_rank = orte_ess.get_local_rank(ORTE_PROC_MY_NAME);
/* if I am local_rank=0 for this node and job, then setup
* my array of local_rank=0 peers
*/
if (0 == my_local_rank) {
/* we need one entry/node in this job */
my_coll_peers = (orte_vpid_t*)malloc(orte_process_info.num_nodes * sizeof(orte_vpid_t));
cpeers = 0;
}
/* cycle through the procs to create a list of those that are local to me */
proc.jobid = ORTE_PROC_MY_NAME->jobid;
for (v=0; v < orte_process_info.num_procs; v++) {
proc.vpid = v;
ORTE_EPOCH_SET(proc.epoch,orte_util_lookup_epoch(&proc));
/* is this proc local_rank=0 on its node? */
if (0 == my_local_rank && 0 == orte_ess.get_local_rank(&proc)) {
my_coll_peers[cpeers++] = v;
}
/* if this is me, or this proc isn't on our node, ignore it */
if (v == ORTE_PROC_MY_NAME->vpid ||
!OPAL_PROC_ON_LOCAL_NODE(orte_ess.proc_get_locality(&proc))) {
continue;
}
/* add this proc to our list of local peers */
nm = OBJ_NEW(orte_namelist_t);
nm->name.jobid = proc.jobid;
nm->name.vpid = proc.vpid;
ORTE_EPOCH_SET(nm->name.epoch,proc.epoch);
opal_list_append(&my_local_peers, &nm->item);
/* if I am not local_rank=0, is this one? */
if (0 != my_local_rank &&
0 == orte_ess.get_local_rank(&proc)) {
my_local_rank_zero_proc.jobid = proc.jobid;
my_local_rank_zero_proc.vpid = proc.vpid;
ORTE_EPOCH_SET(my_local_rank_zero_proc.epoch,proc.epoch);
}
}
/* compute the number of local peers - note that this number
* does not include me!!
*/
num_local_peers = opal_list_get_size(&my_local_peers);
/* flag that I have initialized things */
coll_initialized = true;
}
/* if I am not local rank = 0 */
if (0 != my_local_rank) {
if (ORTE_VPID_INVALID == my_local_rank_zero_proc.vpid) {
/* something is broken */
ORTE_ERROR_LOG(ORTE_ERR_FATAL);
return ORTE_ERR_FATAL;
}
/* setup the collective */
OPAL_THREAD_LOCK(&allgather.lock);
allgather.recvd = 0;
/* reset the collector */
OBJ_DESTRUCT(&allgather.results);
OBJ_CONSTRUCT(&allgather.results, opal_buffer_t);
OPAL_THREAD_UNLOCK(&allgather.lock);
/* send our data to the local_rank=0 proc on this node */
if (0 > (rc = orte_rml.send_buffer(&my_local_rank_zero_proc, sbuf, ORTE_RML_TAG_ALLGATHER, 0))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* now receive the final result. Be sure to do this in
* a manner that allows us to return without being in a recv!
*/
rc = orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_ALLGATHER,
ORTE_RML_NON_PERSISTENT, allgather_recv, &allgather);
if (rc != ORTE_SUCCESS) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* wait to complete - we will receive a single message
* sent from our local_rank=0 peer
*/
OPAL_THREAD_LOCK(&allgather.lock);
while (allgather.recvd < 1) {
opal_condition_wait(&allgather.cond, &allgather.lock);
}
/* copy payload to the caller's buffer */
if (ORTE_SUCCESS != (rc = opal_dss.copy_payload(rbuf, &allgather.results))) {
ORTE_ERROR_LOG(rc);
}
OPAL_THREAD_UNLOCK(&allgather.lock);
} else {
/* I am local_rank = 0 on this node! */
/* setup the collective */
OPAL_THREAD_LOCK(&allgather.lock);
allgather.recvd = 0;
/* reset the collector */
OBJ_DESTRUCT(&allgather.results);
OBJ_CONSTRUCT(&allgather.results, opal_buffer_t);
/* seed with my data */
opal_dss.copy_payload(&allgather.results, sbuf);
OPAL_THREAD_UNLOCK(&allgather.lock);
/* wait to receive their data. Be sure to do this in
* a manner that allows us to return without being in a recv!
*/
rc = orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_ALLGATHER,
ORTE_RML_PERSISTENT, allgather_recv, &allgather);
if (rc != ORTE_SUCCESS) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* wait to complete - we need to receive input from every
* local peer (excluding myself)
*/
OPAL_THREAD_LOCK(&allgather.lock);
while (allgather.recvd < num_local_peers) {
opal_condition_wait(&allgather.cond, &allgather.lock);
}
/* xfer to the tmp buf in case another allgather comes along */
OBJ_CONSTRUCT(&tmp_buf, opal_buffer_t);
opal_dss.copy_payload(&tmp_buf, &allgather.results);
OPAL_THREAD_UNLOCK(&allgather.lock);
/* cancel the lingering recv */
orte_rml.recv_cancel(ORTE_NAME_WILDCARD, ORTE_RML_TAG_ALLGATHER);
/* take the recv'd data and use one of the base collectives
* to exchange it with all other local_rank=0 procs in a scalable
* manner - the exact collective will depend upon the number of
* nodes in the job
*/
if (ORTE_SUCCESS != (rc = orte_grpcomm_base_allgather(&tmp_buf, rbuf, num_local_peers + 1,
ORTE_PROC_MY_NAME->jobid,
cpeers, my_coll_peers))) {
ORTE_ERROR_LOG(rc);
OBJ_DESTRUCT(&tmp_buf);
return rc;
}
OBJ_DESTRUCT(&tmp_buf); /* done with this */
/* distribute the results to our local peers */
for (item = opal_list_get_first(&my_local_peers);
item != opal_list_get_end(&my_local_peers);
item = opal_list_get_next(item)) {
nm = (orte_namelist_t*)item;
if (0 > (rc = orte_rml.send_buffer(&nm->name, rbuf, ORTE_RML_TAG_ALLGATHER, 0))) {
ORTE_ERROR_LOG(rc);
return rc;
}
}
}
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base.output,
"%s grpcomm:hier allgather completed",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
return ORTE_SUCCESS;
}
int mca_spml_ikrit_add_procs(ompi_proc_t** procs, size_t nprocs)
{
spml_ikrit_mxm_ep_conn_info_t *ep_info = NULL;
spml_ikrit_mxm_ep_conn_info_t *ep_hw_rdma_info = NULL;
spml_ikrit_mxm_ep_conn_info_t my_ep_info;
size_t mxm_addr_len = MXM_MAX_ADDR_LEN;
mxm_error_t err;
size_t i, n;
int rc = OSHMEM_ERROR;
ompi_proc_t *proc_self;
int my_rank = oshmem_my_proc_id();
OBJ_CONSTRUCT(&mca_spml_ikrit.active_peers, opal_list_t);
/* Allocate connection requests */
ep_info = calloc(sizeof(spml_ikrit_mxm_ep_conn_info_t), nprocs);
if (NULL == ep_info) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
if (mca_spml_ikrit.hw_rdma_channel) {
ep_hw_rdma_info = calloc(sizeof(spml_ikrit_mxm_ep_conn_info_t), nprocs);
if (NULL == ep_hw_rdma_info) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
}
mca_spml_ikrit.mxm_peers = (mxm_peer_t *) calloc(nprocs , sizeof(mxm_peer_t));
if (NULL == mca_spml_ikrit.mxm_peers) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
memset(&my_ep_info, 0, sizeof(my_ep_info));
if (mca_spml_ikrit.hw_rdma_channel) {
err = mxm_ep_get_address(mca_spml_ikrit.mxm_hw_rdma_ep, &my_ep_info.addr.ep_addr, &mxm_addr_len);
if (MXM_OK != err) {
orte_show_help("help-oshmem-spml-ikrit.txt", "unable to get endpoint address", true,
mxm_error_string(err));
rc = OSHMEM_ERROR;
goto bail;
}
oshmem_shmem_allgather(&my_ep_info, ep_hw_rdma_info,
sizeof(spml_ikrit_mxm_ep_conn_info_t));
}
err = mxm_ep_get_address(mca_spml_ikrit.mxm_ep, &my_ep_info.addr.ep_addr, &mxm_addr_len);
if (MXM_OK != err) {
orte_show_help("help-oshmem-spml-ikrit.txt", "unable to get endpoint address", true,
mxm_error_string(err));
rc = OSHMEM_ERROR;
goto bail;
}
oshmem_shmem_allgather(&my_ep_info, ep_info,
sizeof(spml_ikrit_mxm_ep_conn_info_t));
opal_progress_register(spml_ikrit_progress);
/* Get the EP connection requests for all the processes from modex */
for (n = 0; n < nprocs; ++n) {
/* mxm 2.0 keeps its connections on a list. Make sure
* that list have different order on every rank */
i = (my_rank + n) % nprocs;
mxm_peer_construct(&mca_spml_ikrit.mxm_peers[i]);
err = mxm_ep_connect(mca_spml_ikrit.mxm_ep, ep_info[i].addr.ep_addr, &mca_spml_ikrit.mxm_peers[i].mxm_conn);
if (MXM_OK != err) {
SPML_ERROR("MXM returned connect error: %s\n", mxm_error_string(err));
goto bail;
}
mxm_conn_ctx_set(mca_spml_ikrit.mxm_peers[i].mxm_conn, &mca_spml_ikrit.mxm_peers[i]);
if (mca_spml_ikrit.hw_rdma_channel) {
err = mxm_ep_connect(mca_spml_ikrit.mxm_hw_rdma_ep, ep_hw_rdma_info[i].addr.ep_addr, &mca_spml_ikrit.mxm_peers[i].mxm_hw_rdma_conn);
if (MXM_OK != err) {
SPML_ERROR("MXM returned connect error: %s\n", mxm_error_string(err));
goto bail;
}
} else {
mca_spml_ikrit.mxm_peers[i].mxm_hw_rdma_conn = mca_spml_ikrit.mxm_peers[i].mxm_conn;
}
}
if (ep_info)
free(ep_info);
if (ep_hw_rdma_info)
free(ep_hw_rdma_info);
if (mca_spml_ikrit.bulk_connect) {
/* Need a barrier to ensure remote peers already created connection */
oshmem_shmem_barrier();
mxm_ep_wireup(mca_spml_ikrit.mxm_ep);
}
proc_self = oshmem_proc_group_find(oshmem_group_all, my_rank);
/* identify local processes and change transport to SHM */
for (i = 0; i < nprocs; i++) {
if (procs[i]->super.proc_name.jobid != proc_self->super.proc_name.jobid ||
!OPAL_PROC_ON_LOCAL_NODE(procs[i]->super.proc_flags)) {
continue;
}
if (procs[i] == proc_self)
continue;
/* use zcopy for put/get via sysv shared memory with fallback to RDMA */
mca_spml_ikrit.mxm_peers[i].ptl_id = MXM_PTL_SHM;
}
SPML_VERBOSE(50, "*** ADDED PROCS ***");
return OSHMEM_SUCCESS;
bail:
if (ep_info)
free(ep_info);
if (ep_hw_rdma_info)
free(ep_hw_rdma_info);
SPML_ERROR("add procs FAILED rc=%d", rc);
return rc;
}
static mca_mpool_base_module_t *
mca_mpool_sm_init(struct mca_mpool_base_resources_t *resources)
{
mca_mpool_sm_module_t *mpool_module;
mca_allocator_base_component_t* allocator_component;
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 */
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++;
}
}
/* Make a new mpool module */
mpool_module =
(mca_mpool_sm_module_t *)malloc(sizeof(mca_mpool_sm_module_t));
mca_mpool_sm_module_init(mpool_module);
/* set sm_size */
mpool_module->sm_size = resources->size;
/* clip at the min size */
if (mpool_module->sm_size < (long) ompi_mpool_sm_min_size) {
mpool_module->sm_size = (long) ompi_mpool_sm_min_size;
}
allocator_component = mca_allocator_component_lookup(
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(&ompi_allocator_base_framework.framework_components) == 0) {
mca_base_component_list_item_t *item =
(mca_base_component_list_item_t *)
opal_list_get_first(&ompi_allocator_base_framework.framework_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);
} else {
opal_output(0, "mca_mpool_sm_init: "
"unable to locate allocator: %s\n",
mca_mpool_sm_component.sm_allocator_name);
free(procs);
return NULL;
}
}
mpool_module->mem_node = resources->mem_node;
opal_output(mca_mpool_sm_component.verbose,
"mca_mpool_sm_init: shared memory size used: (%ld)",
mpool_module->sm_size);
if (NULL == (mpool_module->sm_common_module =
mca_common_sm_module_attach(&resources->bs_meta_buf,
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)",
resources->bs_meta_buf.seg_name);
free(mpool_module);
free(procs);
return NULL;
}
free(procs);
/* setup allocator */
mpool_module->sm_allocator =
allocator_component->allocator_init(true,
mca_common_sm_seg_alloc,
NULL, &(mpool_module->super));
if (NULL == mpool_module->sm_allocator) {
opal_output(0, "mca_mpool_sm_init: unable to initialize allocator");
free(mpool_module);
return NULL;
}
return &mpool_module->super;
}
static int ec_on_local_node (rte_ec_handle_t ec, rte_grp_handle_t group) {
ompi_proc_t *proc = (ompi_proc_t *)ec.handle;
return OPAL_PROC_ON_LOCAL_NODE(proc->super.proc_flags);
}
int mca_btl_sm_add_procs(
struct mca_btl_base_module_t* btl,
size_t nprocs,
struct ompi_proc_t **procs,
struct mca_btl_base_endpoint_t **peers,
opal_bitmap_t* reachability)
{
int return_code = OMPI_SUCCESS;
int32_t n_local_procs = 0, proc, j, my_smp_rank = -1;
ompi_proc_t* my_proc; /* pointer to caller's proc structure */
mca_btl_sm_t *sm_btl;
bool have_connected_peer = false;
char **bases;
/* for easy access to the mpool_sm_module */
mca_mpool_sm_module_t *sm_mpool_modp = NULL;
/* initializion */
sm_btl = (mca_btl_sm_t *)btl;
/* get pointer to my proc structure */
if(NULL == (my_proc = ompi_proc_local()))
return OMPI_ERR_OUT_OF_RESOURCE;
/* Get unique host identifier for each process in the list,
* 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++) {
/* 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 ||
!OPAL_PROC_ON_LOCAL_NODE(procs[proc]->proc_flags)) {
peers[proc] = NULL;
continue;
}
/* check to see if this is me */
if(my_proc == procs[proc]) {
my_smp_rank = mca_btl_sm_component.my_smp_rank = n_local_procs++;
continue;
}
/* sm doesn't support heterogeneous yet... */
if (procs[proc]->proc_arch != my_proc->proc_arch) {
continue;
}
/* we have someone to talk to */
have_connected_peer = true;
if(!(peers[proc] = create_sm_endpoint(n_local_procs, procs[proc]))) {
return_code = OMPI_ERROR;
goto CLEANUP;
}
n_local_procs++;
/* add this proc to shared memory accessibility list */
return_code = opal_bitmap_set_bit(reachability, proc);
if(OMPI_SUCCESS != return_code)
goto CLEANUP;
}
/* jump out if there's not someone we can talk to */
if (!have_connected_peer)
goto CLEANUP;
/* make sure that my_smp_rank has been defined */
if (-1 == my_smp_rank) {
return_code = OMPI_ERROR;
goto CLEANUP;
}
if (!sm_btl->btl_inited) {
return_code =
sm_btl_first_time_init(sm_btl, my_smp_rank,
mca_btl_sm_component.sm_max_procs);
if (return_code != OMPI_SUCCESS) {
goto CLEANUP;
}
}
/* set local proc's smp rank in the peers structure for
* rapid access and calculate reachability */
for(proc = 0; proc < (int32_t)nprocs; proc++) {
if(NULL == peers[proc])
continue;
mca_btl_sm_component.sm_peers[peers[proc]->peer_smp_rank] = peers[proc];
peers[proc]->my_smp_rank = my_smp_rank;
}
bases = mca_btl_sm_component.shm_bases;
sm_mpool_modp = (mca_mpool_sm_module_t *)mca_btl_sm_component.sm_mpool;
/* initialize own FIFOs */
/*
* The receiver initializes all its FIFOs. All components will
* be allocated near the receiver. Nothing will be local to
* "the sender" since there will be many senders.
*/
for(j = mca_btl_sm_component.num_smp_procs;
j < mca_btl_sm_component.num_smp_procs + FIFO_MAP_NUM(n_local_procs); j++) {
return_code = sm_fifo_init( mca_btl_sm_component.fifo_size,
mca_btl_sm_component.sm_mpool,
&mca_btl_sm_component.fifo[my_smp_rank][j],
mca_btl_sm_component.fifo_lazy_free);
if(return_code != OMPI_SUCCESS)
goto CLEANUP;
}
opal_atomic_wmb();
/* 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);
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");
}
/* SKG - another abstraction violation here, but I don't want to add
* extra code in the sm mpool for further synchronization. */
/* at this point, all processes have attached to the mpool segment. so
* it is safe to unlink it here. */
if (OMPI_SUCCESS !=
mca_common_sm_module_unlink(sm_mpool_modp->sm_common_module)) {
opal_output(0, "WARNING: common_sm_module_unlink failed.\n");
}
if (-1 == unlink(mca_btl_sm_component.sm_mpool_rndv_file_name)) {
opal_output(0, "WARNING: %s unlink failed.\n",
mca_btl_sm_component.sm_mpool_rndv_file_name);
}
if (-1 == unlink(mca_btl_sm_component.sm_rndv_file_name)) {
opal_output(0, "WARNING: %s unlink failed.\n",
mca_btl_sm_component.sm_rndv_file_name);
}
}
/* free up some space used by the name buffers */
free(mca_btl_sm_component.sm_mpool_ctl_file_name);
free(mca_btl_sm_component.sm_mpool_rndv_file_name);
free(mca_btl_sm_component.sm_ctl_file_name);
free(mca_btl_sm_component.sm_rndv_file_name);
/* 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++) {
ptrdiff_t diff;
/* spin until this element is allocated */
/* doesn't really wait for that process... FIFO might be allocated, but not initialized */
opal_atomic_rmb();
while(NULL == mca_btl_sm_component.shm_fifo[j]) {
opal_progress();
opal_atomic_rmb();
}
/* Calculate the difference as (my_base - their_base) */
diff = ADDR2OFFSET(bases[my_smp_rank], bases[j]);
/* store local address of remote fifos */
mca_btl_sm_component.fifo[j] =
(sm_fifo_t*)OFFSET2ADDR(diff, mca_btl_sm_component.shm_fifo[j]);
/* cache local copy of peer memory node number */
mca_btl_sm_component.mem_nodes[j] = mca_btl_sm_component.shm_mem_nodes[j];
}
/* update the local smp process count */
mca_btl_sm_component.num_smp_procs += n_local_procs;
/* make sure we have enough eager fragmnents for each process */
return_code = ompi_free_list_resize_mt(&mca_btl_sm_component.sm_frags_eager,
mca_btl_sm_component.num_smp_procs * 2);
if (OMPI_SUCCESS != return_code)
goto CLEANUP;
CLEANUP:
return return_code;
}
int mca_spml_ikrit_add_procs(ompi_proc_t** procs, size_t nprocs)
{
spml_ikrit_mxm_ep_conn_info_t *ep_info = NULL;
spml_ikrit_mxm_ep_conn_info_t *ep_hw_rdma_info = NULL;
spml_ikrit_mxm_ep_conn_info_t my_ep_info = {{0}};
#if MXM_API < MXM_VERSION(2,0)
mxm_conn_req_t *conn_reqs;
int timeout;
#else
size_t mxm_addr_len = MXM_MAX_ADDR_LEN;
#endif
mxm_error_t err;
size_t i, n;
int rc = OSHMEM_ERROR;
ompi_proc_t *proc_self;
int my_rank = oshmem_my_proc_id();
OBJ_CONSTRUCT(&mca_spml_ikrit.active_peers, opal_list_t);
/* Allocate connection requests */
#if MXM_API < MXM_VERSION(2,0)
conn_reqs = malloc(nprocs * sizeof(mxm_conn_req_t));
if (NULL == conn_reqs) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
memset(conn_reqs, 0x0, sizeof(mxm_conn_req_t));
#endif
ep_info = calloc(sizeof(spml_ikrit_mxm_ep_conn_info_t), nprocs);
if (NULL == ep_info) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
if (mca_spml_ikrit.hw_rdma_channel) {
ep_hw_rdma_info = calloc(sizeof(spml_ikrit_mxm_ep_conn_info_t), nprocs);
if (NULL == ep_hw_rdma_info) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
}
mca_spml_ikrit.mxm_peers = (mxm_peer_t **) malloc(nprocs
* sizeof(*(mca_spml_ikrit.mxm_peers)));
if (NULL == mca_spml_ikrit.mxm_peers) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
#if MXM_API < MXM_VERSION(2,0)
if (OSHMEM_SUCCESS
!= spml_ikrit_get_ep_address(&my_ep_info, MXM_PTL_SELF)) {
rc = OSHMEM_ERROR;
goto bail;
}
if (OSHMEM_SUCCESS
!= spml_ikrit_get_ep_address(&my_ep_info, MXM_PTL_RDMA)) {
rc = OSHMEM_ERROR;
goto bail;
}
#else
if (mca_spml_ikrit.hw_rdma_channel) {
err = mxm_ep_get_address(mca_spml_ikrit.mxm_hw_rdma_ep, &my_ep_info.addr.ep_addr, &mxm_addr_len);
if (MXM_OK != err) {
orte_show_help("help-oshmem-spml-ikrit.txt", "unable to get endpoint address", true,
mxm_error_string(err));
rc = OSHMEM_ERROR;
goto bail;
}
oshmem_shmem_allgather(&my_ep_info, ep_hw_rdma_info,
sizeof(spml_ikrit_mxm_ep_conn_info_t));
}
err = mxm_ep_get_address(mca_spml_ikrit.mxm_ep, &my_ep_info.addr.ep_addr, &mxm_addr_len);
if (MXM_OK != err) {
orte_show_help("help-oshmem-spml-ikrit.txt", "unable to get endpoint address", true,
mxm_error_string(err));
rc = OSHMEM_ERROR;
goto bail;
}
#endif
oshmem_shmem_allgather(&my_ep_info, ep_info,
sizeof(spml_ikrit_mxm_ep_conn_info_t));
opal_progress_register(spml_ikrit_progress);
/* Get the EP connection requests for all the processes from modex */
for (n = 0; n < nprocs; ++n) {
/* mxm 2.0 keeps its connections on a list. Make sure
* that list have different order on every rank */
i = (my_rank + n) % nprocs;
mca_spml_ikrit.mxm_peers[i] = OBJ_NEW(mxm_peer_t);
if (NULL == mca_spml_ikrit.mxm_peers[i]) {
rc = OSHMEM_ERR_OUT_OF_RESOURCE;
goto bail;
}
mca_spml_ikrit.mxm_peers[i]->pe = i;
#if MXM_API < MXM_VERSION(2,0)
conn_reqs[i].ptl_addr[MXM_PTL_SELF] =
(struct sockaddr *) &ep_info[i].addr.ptl_addr[MXM_PTL_SELF];
conn_reqs[i].ptl_addr[MXM_PTL_SHM] = NULL;
conn_reqs[i].ptl_addr[MXM_PTL_RDMA] =
(struct sockaddr *) &ep_info[i].addr.ptl_addr[MXM_PTL_RDMA];
#else
err = mxm_ep_connect(mca_spml_ikrit.mxm_ep, ep_info[i].addr.ep_addr, &mca_spml_ikrit.mxm_peers[i]->mxm_conn);
if (MXM_OK != err) {
SPML_ERROR("MXM returned connect error: %s\n", mxm_error_string(err));
goto bail;
}
if (OSHMEM_SUCCESS != create_ptl_idx(i))
goto bail;
mxm_conn_ctx_set(mca_spml_ikrit.mxm_peers[i]->mxm_conn, mca_spml_ikrit.mxm_peers[i]);
if (mca_spml_ikrit.hw_rdma_channel) {
err = mxm_ep_connect(mca_spml_ikrit.mxm_hw_rdma_ep, ep_hw_rdma_info[i].addr.ep_addr, &mca_spml_ikrit.mxm_peers[i]->mxm_hw_rdma_conn);
if (MXM_OK != err) {
SPML_ERROR("MXM returned connect error: %s\n", mxm_error_string(err));
goto bail;
}
} else {
mca_spml_ikrit.mxm_peers[i]->mxm_hw_rdma_conn = mca_spml_ikrit.mxm_peers[i]->mxm_conn;
}
#endif
}
#if MXM_API < MXM_VERSION(2,0)
/* Connect to remote peers */
if (mxm_get_version() < MXM_VERSION(1,5)) {
timeout = 1000;
} else {
timeout = -1;
}
err = mxm_ep_connect(mca_spml_ikrit.mxm_ep, conn_reqs, nprocs, timeout);
if (MXM_OK != err) {
SPML_ERROR("MXM returned connect error: %s\n", mxm_error_string(err));
for (i = 0; i < nprocs; ++i) {
if (MXM_OK != conn_reqs[i].error) {
SPML_ERROR("MXM EP connect to %s error: %s\n",
procs[i]->proc_hostname, mxm_error_string(conn_reqs[i].error));
}
}
rc = OSHMEM_ERR_CONNECTION_FAILED;
goto bail;
}
/* Save returned connections */
for (i = 0; i < nprocs; ++i) {
mca_spml_ikrit.mxm_peers[i]->mxm_conn = conn_reqs[i].conn;
if (OSHMEM_SUCCESS != create_ptl_idx(i)) {
rc = OSHMEM_ERR_CONNECTION_FAILED;
goto bail;
}
mxm_conn_ctx_set(conn_reqs[i].conn, mca_spml_ikrit.mxm_peers[i]);
}
if (conn_reqs)
free(conn_reqs);
#endif
if (ep_info)
free(ep_info);
if (ep_hw_rdma_info)
free(ep_hw_rdma_info);
#if MXM_API >= MXM_VERSION(2,0)
if (mca_spml_ikrit.bulk_connect) {
/* Need a barrier to ensure remote peers already created connection */
oshmem_shmem_barrier();
mxm_ep_wireup(mca_spml_ikrit.mxm_ep);
}
#endif
proc_self = oshmem_proc_group_find(oshmem_group_all, my_rank);
/* identify local processes and change transport to SHM */
for (i = 0; i < nprocs; i++) {
if (procs[i]->super.proc_name.jobid != proc_self->super.proc_name.jobid ||
!OPAL_PROC_ON_LOCAL_NODE(procs[i]->super.proc_flags)) {
continue;
}
if (procs[i] == proc_self)
continue;
/* use zcopy for put/get via sysv shared memory */
OSHMEM_PROC_DATA(procs[i])->transport_ids[0] = MXM_PTL_SHM;
OSHMEM_PROC_DATA(procs[i])->transport_ids[1] = MXM_PTL_RDMA;
OSHMEM_PROC_DATA(procs[i])->num_transports = 2;
}
SPML_VERBOSE(50, "*** ADDED PROCS ***");
return OSHMEM_SUCCESS;
bail:
#if MXM_API < MXM_VERSION(2,0)
if (conn_reqs)
free(conn_reqs);
#endif
if (ep_info)
free(ep_info);
if (ep_hw_rdma_info)
free(ep_hw_rdma_info);
SPML_ERROR("add procs FAILED rc=%d", rc);
return rc;
}
