void chpl_comm_ofi_oob_allgather(const void* mine, void* all, size_t size) {
DBG_PRINTF(DBG_OOB, "OOB allGather: %zd", size);
//
// PMI doesn't provide an ordered allGather, so we build one here
// by concatenating the node index and the payload and using that
// to scatter the unordered PMI_Allgather() results.
//
typedef struct {
int nodeID;
uint64_t info[];
} gather_t;
const size_t g_size = offsetof(gather_t, info) + size;
gather_t* g_mine;
CHK_SYS_CALLOC_SZ(g_mine, 1, g_size);
g_mine->nodeID = chpl_nodeID;
memcpy(&g_mine->info, mine, size);
gather_t* g_all;
CHK_SYS_CALLOC_SZ(g_all, chpl_numNodes, g_size);
PMI_CHK(PMI_Allgather(g_mine, g_all, g_size));
for (int g_i = 0; g_i < chpl_numNodes; g_i++) {
char* g_pa = (char*) g_all + g_i * g_size;
int i;
memcpy(&i, g_pa + offsetof(gather_t, nodeID), sizeof(i));
char* p_a = (char*) all + i * size;
memcpy(p_a, g_pa + offsetof(gather_t, info), size);
}
sys_free(g_all);
sys_free(g_mine);
}
static int cray_fence(opal_list_t *procs, int collect_data)
{
int rc, cnt;
int32_t i;
int *all_lens = NULL;
opal_value_t *kp, kvn;
opal_buffer_t *send_buffer = NULL;
opal_buffer_t *buf = NULL;
void *sbuf_ptr;
char *cptr, *rcv_buff = NULL;
opal_process_name_t id;
typedef struct {
uint32_t pmix_rank;
opal_process_name_t name;
int32_t nbytes;
} bytes_and_rank_t;
int32_t rcv_nbytes_tot;
bytes_and_rank_t s_bytes_and_rank;
bytes_and_rank_t *r_bytes_and_ranks = NULL;
opal_hwloc_locality_t locality;
opal_list_t vals;
char *cpuset = NULL;
opal_process_name_t pname;
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:cray executing fence cache_global %p cache_local %p",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
(void *)mca_pmix_cray_component.cache_global,
(void *)mca_pmix_cray_component.cache_local);
/* get the modex data from each local process and set the
* localities to avoid having the MPI layer fetch data
* for every process in the job */
pname.jobid = OPAL_PROC_MY_NAME.jobid;
/*
* "unload" the cache_local/cache_global buffers, first copy
* it so we can continue to use the local buffers if further
* calls to put can be made
*/
send_buffer = OBJ_NEW(opal_buffer_t);
if (NULL == send_buffer) {
return OPAL_ERR_OUT_OF_RESOURCE;
}
opal_dss.copy_payload(send_buffer, mca_pmix_cray_component.cache_global);
opal_dss.unload(send_buffer, &sbuf_ptr, &s_bytes_and_rank.nbytes);
s_bytes_and_rank.pmix_rank = pmix_rank;
s_bytes_and_rank.name = OPAL_PROC_MY_NAME;
r_bytes_and_ranks = (bytes_and_rank_t *)malloc(pmix_size * sizeof(bytes_and_rank_t));
if (NULL == r_bytes_and_ranks) {
rc = OPAL_ERR_OUT_OF_RESOURCE;
goto fn_exit;
}
/*
* gather up all the buffer sizes and rank order.
* doing this step below since the cray pmi PMI_Allgather doesn't deliver
* the gathered data necessarily in PMI rank order, although the order stays
* the same for the duration of a job - assuming no node failures.
*/
if (PMI_SUCCESS != (rc = PMI_Allgather(&s_bytes_and_rank,r_bytes_and_ranks,sizeof(bytes_and_rank_t)))) {
OPAL_PMI_ERROR(rc,"PMI_Allgather");
rc = OPAL_ERR_COMM_FAILURE;
goto fn_exit;
}
for (rcv_nbytes_tot=0,i=0; i < pmix_size; i++) {
rcv_nbytes_tot += r_bytes_and_ranks[i].nbytes;
}
opal_output_verbose(20, opal_pmix_base_framework.framework_output,
"%s pmix:cray total number of bytes to receive %d",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME), rcv_nbytes_tot);
rcv_buff = (char *) malloc(rcv_nbytes_tot * sizeof(char));
if (NULL == rcv_buff) {
rc = OPAL_ERR_OUT_OF_RESOURCE;
goto fn_exit;
}
all_lens = (int *)malloc(sizeof(int) * pmix_size);
if (NULL == all_lens) {
rc = OPAL_ERR_OUT_OF_RESOURCE;
goto fn_exit;
}
for (i=0; i< pmix_size; i++) {
all_lens[r_bytes_and_ranks[i].pmix_rank] = r_bytes_and_ranks[i].nbytes;
}
if (PMI_SUCCESS != (rc = PMI_Allgatherv(sbuf_ptr,s_bytes_and_rank.nbytes,rcv_buff,all_lens))) {
OPAL_PMI_ERROR(rc,"PMI_Allgatherv");
rc = OPAL_ERR_COMM_FAILURE;
goto fn_exit;
}
OBJ_RELEASE(send_buffer);
send_buffer = NULL;
buf = OBJ_NEW(opal_buffer_t);
if (buf == NULL) {
rc = OPAL_ERR_OUT_OF_RESOURCE;
goto fn_exit;
}
for (cptr = rcv_buff, i=0; i < pmix_size; i++) {
id = r_bytes_and_ranks[i].name;
buf->base_ptr = NULL; /* TODO: ugh */
if (OPAL_SUCCESS != (rc = opal_dss.load(buf, (void *)cptr, r_bytes_and_ranks[i].nbytes))) {
OPAL_PMI_ERROR(rc,"pmix:cray opal_dss.load failed");
goto fn_exit;
}
/* unpack and stuff in to the dstore */
cnt = 1;
while (OPAL_SUCCESS == (rc = opal_dss.unpack(buf, &kp, &cnt, OPAL_VALUE))) {
opal_output_verbose(20, opal_pmix_base_framework.framework_output,
"%s pmix:cray unpacked kp with key %s type(%d) for id %s",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME), kp->key, kp->type, OPAL_NAME_PRINT(id));
if (OPAL_SUCCESS != (rc = opal_pmix_base_store(&id, kp))) {
OPAL_ERROR_LOG(rc);
goto fn_exit;
}
OBJ_RELEASE(kp);
cnt = 1;
}
cptr += r_bytes_and_ranks[i].nbytes;
}
buf->base_ptr = NULL; /* TODO: ugh */
OBJ_RELEASE(buf);
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:cray kvs_fence complete",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
#if OPAL_HAVE_HWLOC
/* fetch my cpuset */
OBJ_CONSTRUCT(&vals, opal_list_t);
if (OPAL_SUCCESS == (rc = opal_pmix_base_fetch(&pmix_pname,
OPAL_PMIX_CPUSET, &vals))) {
kp = (opal_value_t*)opal_list_get_first(&vals);
cpuset = strdup(kp->data.string);
} else {
cpuset = NULL;
}
OPAL_LIST_DESTRUCT(&vals);
#endif
/* we only need to set locality for each local rank as "not found"
* equates to "non-local" */
for (i=0; i < pmix_nlranks; i++) {
id.vpid = pmix_lranks[i];
id.jobid = pmix_jobid;
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s checking out if %s is local to me",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
OPAL_NAME_PRINT(id));
/* fetch cpuset for this vpid */
#if OPAL_HAVE_HWLOC
OBJ_CONSTRUCT(&vals, opal_list_t);
if (OPAL_SUCCESS != (rc = opal_pmix_base_fetch(&id,
OPAL_PMIX_CPUSET, &vals))) {
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s cpuset for local proc %s not found",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
OPAL_NAME_PRINT(id));
OPAL_LIST_DESTRUCT(&vals);
/* even though the cpuset wasn't found, we at least know it is
* on the same node with us */
locality = OPAL_PROC_ON_CLUSTER | OPAL_PROC_ON_CU | OPAL_PROC_ON_NODE;
} else {
kp = (opal_value_t*)opal_list_get_first(&vals);
if (NULL == kp->data.string) {
/* if we share a node, but we don't know anything more, then
* mark us as on the node as this is all we know
*/
locality = OPAL_PROC_ON_CLUSTER | OPAL_PROC_ON_CU | OPAL_PROC_ON_NODE;
} else {
/* determine relative location on our node */
locality = opal_hwloc_base_get_relative_locality(opal_hwloc_topology,
cpuset,
kp->data.string);
}
OPAL_LIST_DESTRUCT(&vals);
}
#else
/* all we know is we share a node */
locality = OPAL_PROC_ON_CLUSTER | OPAL_PROC_ON_CU | OPAL_PROC_ON_NODE;
#endif
OPAL_OUTPUT_VERBOSE((1, opal_pmix_base_framework.framework_output,
"%s pmix:cray proc %s locality %s",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
OPAL_NAME_PRINT(id),
opal_hwloc_base_print_locality(locality)));
OBJ_CONSTRUCT(&kvn, opal_value_t);
kvn.key = strdup(OPAL_PMIX_LOCALITY);
kvn.type = OPAL_UINT16;
kvn.data.uint16 = locality;
opal_pmix_base_store(&pname, &kvn);
OBJ_DESTRUCT(&kvn);
}
fn_exit:
#if OPAL_HAVE_HWLOC
if (NULL != cpuset) {
free(cpuset);
}
#endif
if (all_lens != NULL) {
free(all_lens);
}
if (rcv_buff != NULL) {
free(rcv_buff);
}
if (r_bytes_and_ranks != NULL) {
free(r_bytes_and_ranks);
}
return rc;
}
counter++;
#endif
#else
PMI_Barrier();
#endif
}
#if HAVE_PMI_ALLGATHER
static gasnet_node_t *gasnetc_pmi_allgather_order = NULL;
GASNETI_INLINE(gasnetc_pmi_allgather_init)
void gasnetc_pmi_allgather_init(void) {
/* perform (just once) an Allgather of node number to establish the order */
if_pf (!gasnetc_pmi_allgather_order) {
int rc;
gasnetc_pmi_allgather_order = gasneti_malloc(gasneti_nodes * sizeof(gasnet_node_t));
rc = PMI_Allgather(&gasneti_mynode, gasnetc_pmi_allgather_order, sizeof(gasnet_node_t));
gasneti_assert(PMI_SUCCESS == rc);
}
}
#endif
/* gasneti_bootstrapExchange
*/
void gasneti_bootstrapExchange_pmi(void *src, size_t len, void *dest) {
#if HAVE_PMI_ALLGATHER
uint8_t *unsorted = gasneti_malloc(len * gasneti_nodes); /* TODO: use alloca()? */
gasnet_node_t i;
int rc;
/* Allgather the callers data to a temporary array */
gasnetc_pmi_allgather_init();
