static int orte_rmaps_rank_file_open(void)
{
/* ensure we flag mapping by user */
#if OPAL_HAVE_HWLOC
if (NULL != opal_hwloc_base_slot_list || NULL != orte_rankfile) {
#else
if (NULL != orte_rankfile) {
#endif
if (ORTE_MAPPING_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) {
/* if a non-default mapping is already specified, then we
* have an error
*/
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "mapping",
"RANK_FILE", orte_rmaps_base_print_mapping(orte_rmaps_base.mapping));
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_CONFLICTED);
return ORTE_ERR_SILENT;
}
ORTE_SET_MAPPING_POLICY(orte_rmaps_base.mapping, ORTE_MAPPING_BYUSER);
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_GIVEN);
/* we are going to bind to cpuset since the user is specifying the cpus */
OPAL_SET_BINDING_POLICY(opal_hwloc_binding_policy, OPAL_BIND_TO_CPUSET);
/* make us first */
my_priority = 10000;
}
return ORTE_SUCCESS;
}
static int orte_rmaps_rank_file_query(mca_base_module_t **module, int *priority)
{
*priority = my_priority;
*module = (mca_base_module_t *)&orte_rmaps_rank_file_module;
return ORTE_SUCCESS;
}
/*
* Sequentially map the ranks according to the placement in the
* specified hostfile
*/
static int orte_rmaps_seq_map(orte_job_t *jdata)
{
orte_job_map_t *map;
orte_app_context_t *app;
int i, n;
orte_std_cntr_t j;
opal_list_item_t *item;
orte_node_t *node, *nd, *save=NULL;
orte_vpid_t vpid;
orte_std_cntr_t num_nodes;
int rc;
opal_list_t *default_node_list=NULL;
opal_list_t *node_list=NULL;
orte_proc_t *proc;
mca_base_component_t *c = &mca_rmaps_seq_component.base_version;
OPAL_OUTPUT_VERBOSE((1, orte_rmaps_base.rmaps_output,
"%s rmaps:seq mapping job %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_JOBID_PRINT(jdata->jobid)));
/* this mapper can only handle initial launch
* when seq mapping is desired - allow
* restarting of failed apps
*/
if (ORTE_JOB_CONTROL_RESTART & jdata->controls) {
opal_output_verbose(5, orte_rmaps_base.rmaps_output,
"mca:rmaps:seq: job %s is being restarted - seq cannot map",
ORTE_JOBID_PRINT(jdata->jobid));
return ORTE_ERR_TAKE_NEXT_OPTION;
}
if (NULL != jdata->map->req_mapper &&
0 != strcasecmp(jdata->map->req_mapper, c->mca_component_name)) {
/* a mapper has been specified, and it isn't me */
opal_output_verbose(5, orte_rmaps_base.rmaps_output,
"mca:rmaps:seq: job %s not using sequential mapper",
ORTE_JOBID_PRINT(jdata->jobid));
return ORTE_ERR_TAKE_NEXT_OPTION;
}
if (ORTE_MAPPING_SEQ != ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
/* I don't know how to do these - defer */
opal_output_verbose(5, orte_rmaps_base.rmaps_output,
"mca:rmaps:seq: job %s not using seq mapper",
ORTE_JOBID_PRINT(jdata->jobid));
return ORTE_ERR_TAKE_NEXT_OPTION;
}
opal_output_verbose(5, orte_rmaps_base.rmaps_output,
"mca:rmaps:seq: mapping job %s",
ORTE_JOBID_PRINT(jdata->jobid));
/* flag that I did the mapping */
if (NULL != jdata->map->last_mapper) {
free(jdata->map->last_mapper);
}
jdata->map->last_mapper = strdup(c->mca_component_name);
/* conveniece def */
map = jdata->map;
/* if there is a default hostfile, go and get its ordered list of nodes */
if (NULL != orte_default_hostfile) {
default_node_list = OBJ_NEW(opal_list_t);
if (ORTE_SUCCESS != (rc = orte_util_get_ordered_host_list(default_node_list, orte_default_hostfile))) {
ORTE_ERROR_LOG(rc);
goto error;
}
}
/* start at the beginning... */
vpid = 0;
jdata->num_procs = 0;
if (NULL != default_node_list) {
save = (orte_node_t*)opal_list_get_first(default_node_list);
}
/* cycle through the app_contexts, mapping them sequentially */
for(i=0; i < jdata->apps->size; i++) {
if (NULL == (app = (orte_app_context_t*)opal_pointer_array_get_item(jdata->apps, i))) {
continue;
}
/* dash-host trumps hostfile */
if (NULL != app->dash_host) {
node_list = OBJ_NEW(opal_list_t);
if (ORTE_SUCCESS != (rc = orte_util_get_ordered_dash_host_list(node_list, app->dash_host))) {
ORTE_ERROR_LOG(rc);
goto error;
}
nd = (orte_node_t*)opal_list_get_first(node_list);
} else if (NULL != app->hostfile) {
node_list = OBJ_NEW(opal_list_t);
if (ORTE_SUCCESS != (rc = orte_util_get_ordered_host_list(node_list, app->hostfile))) {
ORTE_ERROR_LOG(rc);
goto error;
}
nd = (orte_node_t*)opal_list_get_first(node_list);
} else if (NULL != default_node_list) {
node_list = default_node_list;
nd = save;
} else {
/* can't do anything - no nodes available! */
orte_show_help("help-orte-rmaps-base.txt",
"orte-rmaps-base:no-available-resources",
true);
return ORTE_ERR_SILENT;
}
/* check for nolocal and remove the head node, if required */
if (map->mapping & ORTE_MAPPING_NO_USE_LOCAL) {
for (item = opal_list_get_first(node_list);
item != opal_list_get_end(node_list);
item = opal_list_get_next(item) ) {
node = (orte_node_t*)item;
/* need to check ifislocal because the name in the
* hostfile may not have been FQDN, while name returned
* by gethostname may have been (or vice versa)
*/
if (opal_ifislocal(node->name)) {
opal_list_remove_item(node_list, item);
OBJ_RELEASE(item); /* "un-retain" it */
}
}
}
if (NULL == node_list || 0 == (num_nodes = (orte_std_cntr_t)opal_list_get_size(node_list))) {
orte_show_help("help-orte-rmaps-base.txt",
"orte-rmaps-base:no-available-resources",
true);
return ORTE_ERR_SILENT;
}
/* if num_procs wasn't specified, set it now */
if (0 == app->num_procs) {
app->num_procs = num_nodes;
}
for (n=0; n < app->num_procs; n++) {
/* find this node on the global array - this is necessary so
* that our mapping gets saved on that array as the objects
* returned by the hostfile function are -not- on the array
*/
node = NULL;
for (j=0; j < orte_node_pool->size; j++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(orte_node_pool, j))) {
continue;
}
if (0 == strcmp(nd->name, node->name)) {
break;
}
}
if (NULL == node) {
/* wasn't found - that is an error */
orte_show_help("help-orte-rmaps-seq.txt",
"orte-rmaps-seq:resource-not-found",
true, nd->name);
rc = ORTE_ERR_SILENT;
goto error;
}
/* ensure the node is in the map */
if (!node->mapped) {
OBJ_RETAIN(node);
opal_pointer_array_add(map->nodes, node);
node->mapped = true;
}
proc = orte_rmaps_base_setup_proc(jdata, node, i);
if ((node->slots < (int)node->num_procs) ||
(0 < node->slots_max && node->slots_max < (int)node->num_procs)) {
if (ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, node->num_procs, app->app);
rc = ORTE_ERR_SILENT;
goto error;
}
/* flag the node as oversubscribed so that sched-yield gets
* properly set
*/
node->oversubscribed = true;
}
/* assign the vpid */
proc->name.vpid = vpid++;
#if OPAL_HAVE_HWLOC
/* assign the locale - okay for the topo to be null as
* it just means it wasn't returned
*/
if (NULL != node->topology) {
proc->locale = hwloc_get_root_obj(node->topology);
}
#endif
/* add to the jdata proc array */
if (ORTE_SUCCESS != (rc = opal_pointer_array_set_item(jdata->procs, proc->name.vpid, proc))) {
ORTE_ERROR_LOG(rc);
goto error;
}
/* move to next node */
nd = (orte_node_t*)opal_list_get_next((opal_list_item_t*)nd);
}
/** track the total number of processes we mapped */
jdata->num_procs += app->num_procs;
/* cleanup the node list if it came from this app_context */
if (node_list != default_node_list) {
while (NULL != (item = opal_list_remove_first(node_list))) {
OBJ_RELEASE(item);
}
OBJ_RELEASE(node_list);
} else {
save = nd;
}
}
return ORTE_SUCCESS;
error:
if (NULL != default_node_list) {
while (NULL != (item = opal_list_remove_first(default_node_list))) {
OBJ_RELEASE(item);
}
OBJ_RELEASE(default_node_list);
}
if (NULL != node_list) {
while (NULL != (item = opal_list_remove_first(node_list))) {
OBJ_RELEASE(item);
}
OBJ_RELEASE(node_list);
}
return rc;
}
/*
* Function for selecting one component from all those that are
* available.
*/
void orte_rmaps_base_map_job(int fd, short args, void *cbdata)
{
orte_job_t *jdata;
orte_job_map_t *map;
int rc;
bool did_map;
opal_list_item_t *item;
orte_rmaps_base_selected_module_t *mod;
orte_job_t *parent;
orte_state_caddy_t *caddy = (orte_state_caddy_t*)cbdata;
/* convenience */
jdata = caddy->jdata;
jdata->state = ORTE_JOB_STATE_MAP;
/* NOTE: NO PROXY COMPONENT REQUIRED - REMOTE PROCS ARE NOT
* ALLOWED TO CALL RMAPS INDEPENDENTLY. ONLY THE PLM CAN
* DO SO, AND ALL PLM COMMANDS ARE RELAYED TO HNP
*/
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: mapping job %s",
ORTE_JOBID_PRINT(jdata->jobid));
/* NOTE: CHECK FOR JDATA->MAP == NULL. IF IT IS, THEN USE
* THE VALUES THAT WERE READ BY THE LOCAL MCA PARAMS. THE
* PLM PROXY WILL SEND A JOB-OBJECT THAT WILL INCLUDE ANY
* MAPPING DIRECTIVES - OTHERWISE, THAT OBJECT WILL HAVE A
* NULL MAP FIELD
* LONE EXCEPTION - WE COPY DISPLAY MAP ACROSS IF THEY
* DIDN'T SET IT
*/
if (NULL == jdata->map) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: creating new map for job %s",
ORTE_JOBID_PRINT(jdata->jobid));
/* create a map object where we will store the results */
map = OBJ_NEW(orte_job_map_t);
if (NULL == map) {
ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
/* load it with the system defaults */
map->mapping = orte_rmaps_base.mapping;
map->ranking = orte_rmaps_base.ranking;
#if OPAL_HAVE_HWLOC
map->binding = opal_hwloc_binding_policy;
#endif
if (NULL != orte_rmaps_base.ppr) {
map->ppr = strdup(orte_rmaps_base.ppr);
}
map->cpus_per_rank = orte_rmaps_base.cpus_per_rank;
map->display_map = orte_rmaps_base.display_map;
/* assign the map object to this job */
jdata->map = map;
} else {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: setting mapping policies for job %s",
ORTE_JOBID_PRINT(jdata->jobid));
if (!jdata->map->display_map) {
jdata->map->display_map = orte_rmaps_base.display_map;
}
/* set the default mapping policy IFF it wasn't provided */
if (!ORTE_MAPPING_POLICY_IS_SET(jdata->map->mapping)) {
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, orte_rmaps_base.mapping);
}
if (!ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping)) {
ORTE_SET_MAPPING_DIRECTIVE(jdata->map->mapping, ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping));
}
/* ditto for rank and bind policies */
if (!ORTE_RANKING_POLICY_IS_SET(jdata->map->ranking)) {
ORTE_SET_RANKING_POLICY(jdata->map->ranking, orte_rmaps_base.ranking);
}
#if OPAL_HAVE_HWLOC
if (!OPAL_BINDING_POLICY_IS_SET(jdata->map->binding)) {
jdata->map->binding = opal_hwloc_binding_policy;
}
#endif
}
#if OPAL_HAVE_HWLOC
/* if we are not going to launch, then we need to set any
* undefined topologies to match our own so the mapper
* can operate
*/
if (orte_do_not_launch) {
orte_node_t *node;
hwloc_topology_t t0;
int i;
node = (orte_node_t*)opal_pointer_array_get_item(orte_node_pool, 0);
t0 = node->topology;
for (i=1; i < orte_node_pool->size; i++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(orte_node_pool, i))) {
continue;
}
if (NULL == node->topology) {
node->topology = t0;
}
}
}
#endif
/* cycle thru the available mappers until one agrees to map
* the job
*/
did_map = false;
if (1 == opal_list_get_size(&orte_rmaps_base.selected_modules)) {
/* forced selection */
mod = (orte_rmaps_base_selected_module_t*)opal_list_get_first(&orte_rmaps_base.selected_modules);
jdata->map->req_mapper = strdup(mod->component->mca_component_name);
}
for (item = opal_list_get_first(&orte_rmaps_base.selected_modules);
item != opal_list_get_end(&orte_rmaps_base.selected_modules);
item = opal_list_get_next(item)) {
mod = (orte_rmaps_base_selected_module_t*)item;
if (ORTE_SUCCESS == (rc = mod->module->map_job(jdata)) ||
ORTE_ERR_RESOURCE_BUSY == rc) {
did_map = true;
break;
}
/* mappers return "next option" if they didn't attempt to
* map the job. anything else is a true error.
*/
if (ORTE_ERR_TAKE_NEXT_OPTION != rc) {
ORTE_ERROR_LOG(rc);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
}
if (did_map && ORTE_ERR_RESOURCE_BUSY == rc) {
/* the map was done but nothing could be mapped
* for launch as all the resources were busy
*/
OBJ_RELEASE(caddy);
return;
}
/* if we get here without doing the map, or with zero procs in
* the map, then that's an error
*/
if (!did_map || 0 == jdata->num_procs) {
orte_show_help("help-orte-rmaps-base.txt", "failed-map", true);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
/* compute and save local ranks */
if (ORTE_SUCCESS != (rc = orte_rmaps_base_compute_local_ranks(jdata))) {
ORTE_ERROR_LOG(rc);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
#if OPAL_HAVE_HWLOC
/* compute and save bindings */
if (ORTE_SUCCESS != (rc = orte_rmaps_base_compute_bindings(jdata))) {
ORTE_ERROR_LOG(rc);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
#endif
/* set the offset so shared memory components can potentially
* connect to any spawned jobs
*/
jdata->offset = orte_total_procs;
/* track the total number of procs launched by us */
orte_total_procs += jdata->num_procs;
/* if it is a dynamic spawn, save the bookmark on the parent's job too */
if (ORTE_JOBID_INVALID != jdata->originator.jobid) {
if (NULL != (parent = orte_get_job_data_object(jdata->originator.jobid))) {
parent->bookmark = jdata->bookmark;
}
}
/* if we wanted to display the map, now is the time to do it - ignore
* daemon job
*/
if (jdata->map->display_map) {
char *output=NULL;
int i, j;
orte_node_t *node;
orte_proc_t *proc;
if (orte_display_diffable_output) {
/* intended solely to test mapping methods, this output
* can become quite long when testing at scale. Rather
* than enduring all the malloc/free's required to
* create an arbitrary-length string, custom-generate
* the output a line at a time here
*/
/* display just the procs in a diffable format */
opal_output(orte_clean_output, "<map>\n\t<jobid=%s>\n\t<offset=%s>",
ORTE_JOBID_PRINT(jdata->jobid), ORTE_VPID_PRINT(jdata->offset));
fflush(stderr);
/* loop through nodes */
for (i=0; i < jdata->map->nodes->size; i++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(jdata->map->nodes, i))) {
continue;
}
opal_output(orte_clean_output, "\t<host name=%s>", (NULL == node->name) ? "UNKNOWN" : node->name);
fflush(stderr);
for (j=0; j < node->procs->size; j++) {
if (NULL == (proc = (orte_proc_t*)opal_pointer_array_get_item(node->procs, j))) {
continue;
}
#if OPAL_HAVE_HWLOC
{
char locale[64];
if (NULL != proc->locale) {
hwloc_bitmap_list_snprintf(locale, 64, proc->locale->cpuset);
}
opal_output(orte_clean_output, "\t\t<process rank=%s app_idx=%ld local_rank=%lu node_rank=%lu locale=%s binding=%s>",
ORTE_VPID_PRINT(proc->name.vpid), (long)proc->app_idx,
(unsigned long)proc->local_rank,
(unsigned long)proc->node_rank, locale,
(NULL == proc->cpu_bitmap) ? "NULL" : proc->cpu_bitmap);
}
#else
opal_output(orte_clean_output, "\t\t<process rank=%s app_idx=%ld local_rank=%lu node_rank=%lu>",
ORTE_VPID_PRINT(proc->name.vpid), (long)proc->app_idx,
(unsigned long)proc->local_rank,
(unsigned long)proc->node_rank);
#endif
fflush(stderr);
}
opal_output(orte_clean_output, "\t</host>");
fflush(stderr);
}
#if OPAL_HAVE_HWLOC
{
opal_hwloc_locality_t locality;
orte_proc_t *p0;
/* test locality - for the first node, print the locality of each proc relative to the first one */
node = (orte_node_t*)opal_pointer_array_get_item(jdata->map->nodes, 0);
p0 = (orte_proc_t*)opal_pointer_array_get_item(node->procs, 0);
opal_output(orte_clean_output, "\t<locality>");
for (j=1; j < node->procs->size; j++) {
if (NULL == (proc = (orte_proc_t*)opal_pointer_array_get_item(node->procs, j))) {
continue;
}
locality = opal_hwloc_base_get_relative_locality(node->topology,
p0->cpu_bitmap,
proc->cpu_bitmap);
opal_output(orte_clean_output, "\t\t<rank=%s rank=%s locality=%s>",
ORTE_VPID_PRINT(p0->name.vpid),
ORTE_VPID_PRINT(proc->name.vpid),
opal_hwloc_base_print_locality(locality));
}
opal_output(orte_clean_output, "\t</locality>\n</map>");
fflush(stderr);
}
#else
opal_output(orte_clean_output, "\n</map>");
fflush(stderr);
#endif
} else {
opal_output(orte_clean_output, " Data for JOB %s offset %s", ORTE_JOBID_PRINT(jdata->jobid), ORTE_VPID_PRINT(jdata->offset));
opal_dss.print(&output, NULL, jdata->map, ORTE_JOB_MAP);
if (orte_xml_output) {
fprintf(orte_xml_fp, "%s\n", output);
fflush(orte_xml_fp);
} else {
opal_output(orte_clean_output, "%s", output);
}
free(output);
}
}
/* set the job state to the next position */
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_MAP_COMPLETE);
/* cleanup */
OBJ_RELEASE(caddy);
}
int orte_rmaps_rr_byslot(orte_job_t *jdata,
orte_app_context_t *app,
opal_list_t *node_list,
orte_std_cntr_t num_slots,
orte_vpid_t num_procs)
{
int rc, i, nprocs_mapped;
orte_node_t *node;
orte_proc_t *proc;
int num_procs_to_assign, extra_procs_to_assign=0, nxtra_nodes=0;
#if OPAL_HAVE_HWLOC
hwloc_obj_t obj=NULL;
#endif
float balance;
bool add_one=false;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr: mapping by slot for job %s slots %d num_procs %lu",
ORTE_JOBID_PRINT(jdata->jobid), (int)num_slots, (unsigned long)num_procs);
/* check to see if we can map all the procs */
if (num_slots < ((int)app->num_procs * orte_rmaps_base.cpus_per_rank)) {
if (ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, app->num_procs, app->app);
return ORTE_ERR_SILENT;
}
#if OPAL_HAVE_HWLOC
/* if we will and are allowed to oversubscribe, and binding was given, then
* we really should warn the user that we cannot bind
*/
if (OPAL_BINDING_POLICY_IS_SET(jdata->map->binding)) {
if ((OPAL_BIND_TO_CORE == OPAL_GET_BINDING_POLICY(jdata->map->binding) ||
OPAL_BIND_TO_HWTHREAD == OPAL_GET_BINDING_POLICY(jdata->map->binding)) &&
!OPAL_BIND_OVERLOAD_ALLOWED(jdata->map->binding)) {
/* RHC: don't emit this warning at this time while we try to
* determine the best path forward. See
* https://svn.open-mpi.org/trac/ompi/ticket/4345
* for an explanation
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:oversubscribed",
true, num_slots, app->num_procs * orte_rmaps_base.cpus_per_rank);
OPAL_SET_BINDING_POLICY(jdata->map->binding, OPAL_BIND_TO_NONE);
*/
}
} else {
/* don't default to bound */
OPAL_SET_BINDING_POLICY(jdata->map->binding, OPAL_BIND_TO_NONE);
}
#endif
}
/* first pass: map the number of procs to each node until we
* map all specified procs or use all allocated slots
*/
nprocs_mapped = 0;
OPAL_LIST_FOREACH(node, node_list, orte_node_t) {
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot working node %s",
node->name);
#if OPAL_HAVE_HWLOC
/* get the root object as we are not assigning
* locale here except at the node level
*/
if (NULL != node->topology) {
obj = hwloc_get_root_obj(node->topology);
}
#endif
if (node->slots <= node->slots_inuse) {
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot node %s is full - skipping",
node->name);
continue;
}
/* assign a number of procs equal to the number of available
* slots divided by the number of cpus/rank the user
* requested
*/
num_procs_to_assign = (node->slots - node->slots_inuse) / orte_rmaps_base.cpus_per_rank;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot assigning %d procs to node %s",
(int)num_procs_to_assign, node->name);
for (i=0; i < num_procs_to_assign && nprocs_mapped < app->num_procs; i++) {
/* add this node to the map - do it only once */
if (!node->mapped) {
if (ORTE_SUCCESS > (rc = opal_pointer_array_add(jdata->map->nodes, (void*)node))) {
ORTE_ERROR_LOG(rc);
return rc;
}
node->mapped = true;
OBJ_RETAIN(node); /* maintain accounting on object */
++(jdata->map->num_nodes);
}
if (NULL == (proc = orte_rmaps_base_setup_proc(jdata, node, app->idx))) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
nprocs_mapped++;
#if OPAL_HAVE_HWLOC
proc->locale = obj;
#endif
}
}
/*
* Query the registry for all nodes allocated to a specified app_context
*/
int orte_rmaps_base_get_target_nodes(opal_list_t *allocated_nodes, orte_std_cntr_t *total_num_slots,
orte_app_context_t *app, orte_mapping_policy_t policy,
bool initial_map, bool silent)
{
opal_list_item_t *item, *next;
orte_node_t *node, *nd, *nptr;
orte_std_cntr_t num_slots;
orte_std_cntr_t i;
int rc;
orte_job_t *daemons;
bool novm;
opal_list_t nodes;
char *hosts;
/** set default answer */
*total_num_slots = 0;
/* get the daemon job object */
daemons = orte_get_job_data_object(ORTE_PROC_MY_NAME->jobid);
/* see if we have a vm or not */
novm = orte_get_attribute(&daemons->attributes, ORTE_JOB_NO_VM, NULL, OPAL_BOOL);
/* if this is NOT a managed allocation, then we use the nodes
* that were specified for this app - there is no need to collect
* all available nodes and "filter" them
*/
if (!orte_managed_allocation) {
OBJ_CONSTRUCT(&nodes, opal_list_t);
/* if the app provided a dash-host, and we are not treating
* them as requested or "soft" locations, then use those nodes
*/
hosts = NULL;
if (!orte_soft_locations &&
orte_get_attribute(&app->attributes, ORTE_APP_DASH_HOST, (void**)&hosts, OPAL_STRING)) {
OPAL_OUTPUT_VERBOSE((5, orte_rmaps_base_framework.framework_output,
"%s using dash_host %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), hosts));
if (ORTE_SUCCESS != (rc = orte_util_add_dash_host_nodes(&nodes, hosts, false))) {
ORTE_ERROR_LOG(rc);
free(hosts);
return rc;
}
free(hosts);
} else if (orte_get_attribute(&app->attributes, ORTE_APP_HOSTFILE, (void**)&hosts, OPAL_STRING)) {
/* otherwise, if the app provided a hostfile, then use that */
OPAL_OUTPUT_VERBOSE((5, orte_rmaps_base_framework.framework_output,
"%s using hostfile %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), hosts));
if (ORTE_SUCCESS != (rc = orte_util_add_hostfile_nodes(&nodes, hosts))) {
free(hosts);
ORTE_ERROR_LOG(rc);
return rc;
}
free(hosts);
} else if (NULL != orte_rankfile) {
/* use the rankfile, if provided */
OPAL_OUTPUT_VERBOSE((5, orte_rmaps_base_framework.framework_output,
"%s using rankfile %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
orte_rankfile));
if (ORTE_SUCCESS != (rc = orte_util_add_hostfile_nodes(&nodes,
orte_rankfile))) {
ORTE_ERROR_LOG(rc);
return rc;
}
if (0 == opal_list_get_size(&nodes)) {
OPAL_OUTPUT_VERBOSE((5, orte_rmaps_base_framework.framework_output,
"%s nothing found in given rankfile",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
OBJ_DESTRUCT(&nodes);
return ORTE_ERR_BAD_PARAM;
}
} else if (NULL != orte_default_hostfile) {
/* fall back to the default hostfile, if provided */
OPAL_OUTPUT_VERBOSE((5, orte_rmaps_base_framework.framework_output,
"%s using default hostfile %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
orte_default_hostfile));
if (ORTE_SUCCESS != (rc = orte_util_add_hostfile_nodes(&nodes,
orte_default_hostfile))) {
ORTE_ERROR_LOG(rc);
return rc;
}
/* this is a special case - we always install a default
* hostfile, but it is empty. If the user didn't remove it
* or put something into it, then we will have pursued that
* option and found nothing. This isn't an error, we just need
* to add all the known nodes
*/
if (0 == opal_list_get_size(&nodes)) {
OPAL_OUTPUT_VERBOSE((5, orte_rmaps_base_framework.framework_output,
"%s nothing in default hostfile - using known nodes",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
goto addknown;
}
} else {
/* if nothing else was available, then use all known nodes, which
* will include ourselves
*/
OPAL_OUTPUT_VERBOSE((5, orte_rmaps_base_framework.framework_output,
"%s using known nodes",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
goto addknown;
}
/** if we still don't have anything */
if (0 == opal_list_get_size(&nodes)) {
if (!silent) {
orte_show_help("help-orte-rmaps-base.txt",
"orte-rmaps-base:no-available-resources",
true);
}
OBJ_DESTRUCT(&nodes);
return ORTE_ERR_SILENT;
}
/* find the nodes in our node array and assemble them
* in daemon order if the vm was launched
*/
while (NULL != (item = opal_list_remove_first(&nodes))) {
nptr = (orte_node_t*)item;
nd = NULL;
for (i=0; i < orte_node_pool->size; i++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(orte_node_pool, i))) {
continue;
}
if (0 != strcmp(node->name, nptr->name)) {
OPAL_OUTPUT_VERBOSE((10, orte_rmaps_base_framework.framework_output,
"NODE %s DOESNT MATCH NODE %s",
node->name, nptr->name));
continue;
}
/* ignore nodes that are marked as do-not-use for this mapping */
if (ORTE_NODE_STATE_DO_NOT_USE == node->state) {
OPAL_OUTPUT_VERBOSE((10, orte_rmaps_base_framework.framework_output,
"NODE %s IS MARKED NO_USE", node->name));
/* reset the state so it can be used another time */
node->state = ORTE_NODE_STATE_UP;
continue;
}
if (ORTE_NODE_STATE_DOWN == node->state) {
OPAL_OUTPUT_VERBOSE((10, orte_rmaps_base_framework.framework_output,
"NODE %s IS DOWN", node->name));
continue;
}
if (ORTE_NODE_STATE_NOT_INCLUDED == node->state) {
OPAL_OUTPUT_VERBOSE((10, orte_rmaps_base_framework.framework_output,
"NODE %s IS MARKED NO_INCLUDE", node->name));
/* not to be used */
continue;
}
/* if this node wasn't included in the vm (e.g., by -host), ignore it,
* unless we are mapping prior to launching the vm
*/
if (NULL == node->daemon && !novm) {
OPAL_OUTPUT_VERBOSE((10, orte_rmaps_base_framework.framework_output,
"NODE %s HAS NO DAEMON", node->name));
continue;
}
/* retain a copy for our use in case the item gets
* destructed along the way
*/
OBJ_RETAIN(node);
if (initial_map) {
/* if this is the first app_context we
* are getting for an initial map of a job,
* then mark all nodes as unmapped
*/
ORTE_FLAG_UNSET(node, ORTE_NODE_FLAG_MAPPED);
}
if (NULL == nd || NULL == nd->daemon ||
NULL == node->daemon ||
nd->daemon->name.vpid < node->daemon->name.vpid) {
/* just append to end */
opal_list_append(allocated_nodes, &node->super);
nd = node;
} else {
/* starting from end, put this node in daemon-vpid order */
while (node->daemon->name.vpid < nd->daemon->name.vpid) {
if (opal_list_get_begin(allocated_nodes) == opal_list_get_prev(&nd->super)) {
/* insert at beginning */
opal_list_prepend(allocated_nodes, &node->super);
goto moveon1;
}
nd = (orte_node_t*)opal_list_get_prev(&nd->super);
}
item = opal_list_get_next(&nd->super);
if (item == opal_list_get_end(allocated_nodes)) {
/* we are at the end - just append */
opal_list_append(allocated_nodes, &node->super);
} else {
nd = (orte_node_t*)item;
opal_list_insert_pos(allocated_nodes, item, &node->super);
}
moveon1:
/* reset us back to the end for the next node */
nd = (orte_node_t*)opal_list_get_last(allocated_nodes);
}
}
OBJ_RELEASE(nptr);
}
OBJ_DESTRUCT(&nodes);
/* now prune for usage and compute total slots */
goto complete;
}
addknown:
/* if the hnp was allocated, include it unless flagged not to */
if (orte_hnp_is_allocated && !(ORTE_GET_MAPPING_DIRECTIVE(policy) & ORTE_MAPPING_NO_USE_LOCAL)) {
if (NULL != (node = (orte_node_t*)opal_pointer_array_get_item(orte_node_pool, 0))) {
if (ORTE_NODE_STATE_DO_NOT_USE == node->state) {
OPAL_OUTPUT_VERBOSE((10, orte_rmaps_base_framework.framework_output,
"HNP IS MARKED NO_USE"));
/* clear this for future use, but don't include it */
node->state = ORTE_NODE_STATE_UP;
} else if (ORTE_NODE_STATE_NOT_INCLUDED != node->state) {
OBJ_RETAIN(node);
if (initial_map) {
/* if this is the first app_context we
* are getting for an initial map of a job,
* then mark all nodes as unmapped
*/
ORTE_FLAG_UNSET(node, ORTE_NODE_FLAG_MAPPED);
}
opal_list_append(allocated_nodes, &node->super);
}
}
}
/* add everything in the node pool that can be used - add them
* in daemon order, which may be different than the order in the
* node pool. Since an empty list is passed into us, the list at
* this point either has the HNP node or nothing, and the HNP
* node obviously has a daemon on it (us!)
*/
if (0 == opal_list_get_size(allocated_nodes)) {
/* the list is empty */
nd = NULL;
} else {
nd = (orte_node_t*)opal_list_get_last(allocated_nodes);
}
for (i=1; i < orte_node_pool->size; i++) {
if (NULL != (node = (orte_node_t*)opal_pointer_array_get_item(orte_node_pool, i))) {
/* ignore nodes that are marked as do-not-use for this mapping */
if (ORTE_NODE_STATE_DO_NOT_USE == node->state) {
OPAL_OUTPUT_VERBOSE((10, orte_rmaps_base_framework.framework_output,
"NODE %s IS MARKED NO_USE", node->name));
/* reset the state so it can be used another time */
node->state = ORTE_NODE_STATE_UP;
continue;
}
if (ORTE_NODE_STATE_DOWN == node->state) {
OPAL_OUTPUT_VERBOSE((10, orte_rmaps_base_framework.framework_output,
"NODE %s IS MARKED DOWN", node->name));
continue;
}
if (ORTE_NODE_STATE_NOT_INCLUDED == node->state) {
OPAL_OUTPUT_VERBOSE((10, orte_rmaps_base_framework.framework_output,
"NODE %s IS MARKED NO_INCLUDE", node->name));
/* not to be used */
continue;
}
/* if this node wasn't included in the vm (e.g., by -host), ignore it,
* unless we are mapping prior to launching the vm
*/
if (NULL == node->daemon && !novm) {
OPAL_OUTPUT_VERBOSE((10, orte_rmaps_base_framework.framework_output,
"NODE %s HAS NO DAEMON", node->name));
continue;
}
/* retain a copy for our use in case the item gets
* destructed along the way
*/
OBJ_RETAIN(node);
if (initial_map) {
/* if this is the first app_context we
* are getting for an initial map of a job,
* then mark all nodes as unmapped
*/
ORTE_FLAG_UNSET(node, ORTE_NODE_FLAG_MAPPED);
}
if (NULL == nd || NULL == nd->daemon ||
NULL == node->daemon ||
nd->daemon->name.vpid < node->daemon->name.vpid) {
/* just append to end */
opal_list_append(allocated_nodes, &node->super);
nd = node;
} else {
/* starting from end, put this node in daemon-vpid order */
while (node->daemon->name.vpid < nd->daemon->name.vpid) {
if (opal_list_get_begin(allocated_nodes) == opal_list_get_prev(&nd->super)) {
/* insert at beginning */
opal_list_prepend(allocated_nodes, &node->super);
goto moveon;
}
nd = (orte_node_t*)opal_list_get_prev(&nd->super);
}
item = opal_list_get_next(&nd->super);
if (item == opal_list_get_end(allocated_nodes)) {
/* we are at the end - just append */
opal_list_append(allocated_nodes, &node->super);
} else {
nd = (orte_node_t*)item;
opal_list_insert_pos(allocated_nodes, item, &node->super);
}
moveon:
/* reset us back to the end for the next node */
nd = (orte_node_t*)opal_list_get_last(allocated_nodes);
}
}
}
OPAL_OUTPUT_VERBOSE((5, orte_rmaps_base_framework.framework_output,
"%s Starting with %d nodes in list",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(int)opal_list_get_size(allocated_nodes)));
/** check that anything is here */
if (0 == opal_list_get_size(allocated_nodes)) {
if (!silent) {
orte_show_help("help-orte-rmaps-base.txt",
"orte-rmaps-base:no-available-resources",
true);
}
return ORTE_ERR_SILENT;
}
/* filter the nodes thru any hostfile and dash-host options */
OPAL_OUTPUT_VERBOSE((5, orte_rmaps_base_framework.framework_output,
"%s Filtering thru apps",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
if (ORTE_SUCCESS != (rc = orte_rmaps_base_filter_nodes(app, allocated_nodes, true))
&& ORTE_ERR_TAKE_NEXT_OPTION != rc) {
ORTE_ERROR_LOG(rc);
return rc;
}
OPAL_OUTPUT_VERBOSE((5, orte_rmaps_base_framework.framework_output,
"%s Retained %d nodes in list",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(int)opal_list_get_size(allocated_nodes)));
complete:
/* remove all nodes that are already at max usage, and
* compute the total number of allocated slots while
* we do so */
num_slots = 0;
item = opal_list_get_first(allocated_nodes);
while (item != opal_list_get_end(allocated_nodes)) {
/** save the next pointer in case we remove this node */
next = opal_list_get_next(item);
/** check to see if this node is fully used - remove if so */
node = (orte_node_t*)item;
if (0 != node->slots_max && node->slots_inuse > node->slots_max) {
OPAL_OUTPUT_VERBOSE((5, orte_rmaps_base_framework.framework_output,
"%s Removing node %s: max %d inuse %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
node->name, node->slots_max, node->slots_inuse));
opal_list_remove_item(allocated_nodes, item);
OBJ_RELEASE(item); /* "un-retain" it */
} else if (node->slots <= node->slots_inuse &&
(ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(policy))) {
/* remove the node as fully used */
OPAL_OUTPUT_VERBOSE((5, orte_rmaps_base_framework.framework_output,
"%s Removing node %s slots %d inuse %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
node->name, node->slots, node->slots_inuse));
opal_list_remove_item(allocated_nodes, item);
OBJ_RELEASE(item); /* "un-retain" it */
} else if (node->slots > node->slots_inuse) {
/* add the available slots */
OPAL_OUTPUT_VERBOSE((5, orte_rmaps_base_framework.framework_output,
"%s node %s has %d slots available",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
node->name, node->slots - node->slots_inuse));
num_slots += node->slots - node->slots_inuse;
} else if (!(ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(policy))) {
/* nothing needed to do here - we don't add slots to the
* count as we don't have any available. Just let the mapper
* do what it needs to do to meet the request
*/
OPAL_OUTPUT_VERBOSE((5, orte_rmaps_base_framework.framework_output,
"%s node %s is fully used, but available for oversubscrition",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
node->name));
} else {
/* if we cannot use it, remove it from list */
opal_list_remove_item(allocated_nodes, item);
OBJ_RELEASE(item); /* "un-retain" it */
}
/** go on to next item */
item = next;
}
/* Sanity check to make sure we have resources available */
if (0 == opal_list_get_size(allocated_nodes)) {
if (silent) {
/* let the caller know that the resources exist,
* but are currently busy
*/
return ORTE_ERR_RESOURCE_BUSY;
} else {
orte_show_help("help-orte-rmaps-base.txt",
"orte-rmaps-base:all-available-resources-used", true);
return ORTE_ERR_SILENT;
}
}
/* pass back the total number of available slots */
*total_num_slots = num_slots;
if (4 < opal_output_get_verbosity(orte_rmaps_base_framework.framework_output)) {
opal_output(0, "AVAILABLE NODES FOR MAPPING:");
for (item = opal_list_get_first(allocated_nodes);
item != opal_list_get_end(allocated_nodes);
item = opal_list_get_next(item)) {
node = (orte_node_t*)item;
opal_output(0, " node: %s daemon: %s", node->name,
(NULL == node->daemon) ? "NULL" : ORTE_VPID_PRINT(node->daemon->name.vpid));
}
}
return ORTE_SUCCESS;
}
int orte_rmaps_rr_bynode(orte_job_t *jdata,
orte_app_context_t *app,
opal_list_t *node_list,
orte_std_cntr_t num_slots,
orte_vpid_t num_procs)
{
int j, nprocs_mapped, lag, delta;
orte_node_t *node;
orte_proc_t *proc;
opal_list_item_t *item;
int num_procs_to_assign, navg, idx;
int extra_procs_to_assign=0, nxtra_nodes=0;
#if OPAL_HAVE_HWLOC
hwloc_obj_t obj=NULL;
#endif
float balance;
bool add_one=false;
bool oversubscribed=false;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr: mapping by node for job %s app %d slots %d num_procs %lu",
ORTE_JOBID_PRINT(jdata->jobid), (int)app->idx,
(int)num_slots, (unsigned long)num_procs);
/* quick check to see if we can map all the procs */
if (num_slots < ((int)app->num_procs * orte_rmaps_base.cpus_per_rank)) {
if (ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, app->num_procs, app->app);
return ORTE_ERR_SILENT;
}
oversubscribed = true;
}
/* divide the procs evenly across all nodes - this is the
* average we have to maintain as we go, but we adjust
* the number on each node to reflect its available slots.
* Obviously, if all nodes have the same number of slots,
* then the avg is what we get on each node - this is
* the most common situation.
*/
navg = app->num_procs / opal_list_get_size(node_list);
if (0 == navg) {
/* if there are less procs than nodes, we have to
* place at least one/node
*/
navg = 1;
}
/* compute how many extra procs to put on each node */
balance = (float)(app->num_procs - (navg * (float)opal_list_get_size(node_list))) / (float)opal_list_get_size(node_list);
extra_procs_to_assign = (int)balance;
if (0 < (balance - (float)extra_procs_to_assign)) {
/* compute how many nodes need an extra proc */
nxtra_nodes = app->num_procs - ((navg + extra_procs_to_assign) * opal_list_get_size(node_list));
/* add one so that we add an extra proc to the first nodes
* until all procs are mapped
*/
extra_procs_to_assign++;
/* flag that we added one */
add_one = true;
}
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr: mapping by node navg %d extra_procs %d extra_nodes %d",
navg, extra_procs_to_assign, nxtra_nodes);
nprocs_mapped = 0;
lag = 0;
for (item = opal_list_get_first(node_list);
item != opal_list_get_end(node_list);
item = opal_list_get_next(item)) {
node = (orte_node_t*)item;
#if OPAL_HAVE_HWLOC
/* get the root object as we are not assigning
* locale except at the node level
*/
if (NULL != node->topology) {
obj = hwloc_get_root_obj(node->topology);
}
#endif
/* add this node to the map, but only do so once */
if (!node->mapped) {
if (ORTE_SUCCESS > (idx = opal_pointer_array_add(jdata->map->nodes, (void*)node))) {
ORTE_ERROR_LOG(idx);
return idx;
}
node->mapped = true;
OBJ_RETAIN(node); /* maintain accounting on object */
++(jdata->map->num_nodes);
}
/* compute the number of procs to go on this node */
if (add_one) {
if (0 == nxtra_nodes) {
--extra_procs_to_assign;
add_one = false;
} else {
--nxtra_nodes;
}
}
if (oversubscribed) {
/* everybody just takes their share */
num_procs_to_assign = navg + extra_procs_to_assign;
} else {
/* if we are not oversubscribed, then there are enough
* slots to handle all the procs. However, not every
* node will have the same number of slots, so we
* have to track how many procs to "shift" elsewhere
* to make up the difference
*/
if (node->slots <= node->slots_inuse) {
/* if there are no extras to take, then we can
* ignore this node
*/
if (0 == extra_procs_to_assign) {
/* update how many we are lagging behind */
lag += navg;
continue;
}
/* everybody has to take at least the extras */
num_procs_to_assign = extra_procs_to_assign;
/* update how many we are lagging behind */
lag += navg;
} else {
/* if slots < avg (adjusted for cpus/proc), then take all */
if ((node->slots - node->slots_inuse) < (navg * orte_rmaps_base.cpus_per_rank)) {
num_procs_to_assign = (node->slots - node->slots_inuse)/orte_rmaps_base.cpus_per_rank;
/* update how many we are lagging behind */
lag += navg - num_procs_to_assign;
} else {
/* take the avg plus as much of the "lag" as we can */
delta = 0;
if (0 < lag) {
delta = ((node->slots - node->slots_inuse)/orte_rmaps_base.cpus_per_rank) - navg;
if (lag < delta) {
delta = lag;
}
lag -= delta;
}
num_procs_to_assign = navg;
}
/* add in the extras */
num_procs_to_assign += extra_procs_to_assign;
}
}
for (j=0; j < num_procs_to_assign && nprocs_mapped < app->num_procs; j++) {
if (NULL == (proc = orte_rmaps_base_setup_proc(jdata, node, app->idx))) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
nprocs_mapped++;
#if OPAL_HAVE_HWLOC
proc->locale = obj;
#endif
}
/* not all nodes are equal, so only set oversubscribed for
* this node if it is in that state
*/
if (node->slots < ((int)node->num_procs * orte_rmaps_base.cpus_per_rank)) {
/* flag the node as oversubscribed so that sched-yield gets
* properly set
*/
node->oversubscribed = true;
}
if (nprocs_mapped == app->num_procs) {
/* we are done */
break;
}
}
/* if we have some remaining lag, then put one/node until
* all are assigned
*/
for (item = opal_list_get_first(node_list);
0 < lag && item != opal_list_get_end(node_list);
item = opal_list_get_next(item)) {
node = (orte_node_t*)item;
if (NULL == (proc = orte_rmaps_base_setup_proc(jdata, node, app->idx))) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
nprocs_mapped++;
lag--;
#if OPAL_HAVE_HWLOC
proc->locale = obj;
#endif
/* not all nodes are equal, so only set oversubscribed for
* this node if it is in that state
*/
if (node->slots < ((int)node->num_procs * orte_rmaps_base.cpus_per_rank)) {
/* flag the node as oversubscribed so that sched-yield gets
* properly set
*/
node->oversubscribed = true;
}
if (nprocs_mapped == app->num_procs) {
/* we are done */
break;
}
}
return ORTE_SUCCESS;
}
int orte_rmaps_base_set_mapping_policy(orte_mapping_policy_t *policy,
char **device, char *spec)
{
char **ck, **ck2;
orte_mapping_policy_t tmp;
int i;
size_t len;
/* set defaults */
tmp = 0;
*device = NULL;
if (NULL == spec) {
ORTE_SET_MAPPING_POLICY(tmp, ORTE_MAPPING_BYSOCKET);
ORTE_SET_MAPPING_DIRECTIVE(tmp, ORTE_MAPPING_SPAN);
ORTE_UNSET_MAPPING_DIRECTIVE(tmp, ORTE_MAPPING_GIVEN);
} else {
ck = opal_argv_split(spec, ':');
if (2 < opal_argv_count(ck)) {
/* incorrect format */
orte_show_help("help-orte-rmaps-base.txt", "unrecognized-policy", true, "mapping", rmaps_base_mapping_policy);
opal_argv_free(ck);
return ORTE_ERR_SILENT;
}
if (2 == opal_argv_count(ck)) {
/* if the policy is "dist", then we set the policy to that value
* and save the second argument as the device
*/
#if OPAL_HAVE_HWLOC
if (0 == strncasecmp(ck[0], "dist", strlen(ck[0]))) {
ORTE_SET_MAPPING_POLICY(tmp, ORTE_MAPPING_BYDIST);
ck2 = opal_argv_split(ck[1], ',');
if (ck2[0] != NULL) {
*device = strdup(ck2[0]);
for (i=1; NULL != ck2[i]; i++) {
if (0 == strncasecmp(ck2[i], "span", strlen(ck2[i]))) {
ORTE_SET_MAPPING_DIRECTIVE(tmp, ORTE_MAPPING_SPAN);
}
}
}
opal_argv_free(ck2);
goto setpolicy;
}
#endif
ck2 = opal_argv_split(ck[1], ',');
for (i=0; NULL != ck2[i]; i++) {
if (0 == strncasecmp(ck2[i], "span", strlen(ck2[i]))) {
ORTE_SET_MAPPING_DIRECTIVE(tmp, ORTE_MAPPING_SPAN);
} else if (0 == strncasecmp(ck2[i], "oversubscribe", strlen(ck2[i]))) {
if (ORTE_MAPPING_SUBSCRIBE_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(tmp)) {
ORTE_UNSET_MAPPING_DIRECTIVE(tmp, ORTE_MAPPING_NO_OVERSUBSCRIBE);
ORTE_SET_MAPPING_DIRECTIVE(tmp, ORTE_MAPPING_SUBSCRIBE_GIVEN);
} else if (0 == strncasecmp(ck2[i], "nooversubscribe", strlen(ck2[i]))) {
ORTE_SET_MAPPING_DIRECTIVE(tmp, ORTE_MAPPING_NO_OVERSUBSCRIBE);
ORTE_SET_MAPPING_DIRECTIVE(tmp, ORTE_MAPPING_SUBSCRIBE_GIVEN);
} else {
/* unrecognized modifier */
orte_show_help("help-orte-rmaps-base.txt", "unrecognized-modifier", true, "mapping", ck2[i]);
opal_argv_free(ck);
opal_argv_free(ck2);
return ORTE_ERR_SILENT;
}
}
opal_argv_free(ck2);
}
}
len = strlen(ck[0]);
if (0 == strncasecmp(ck[0], "slot", len)) {
ORTE_SET_MAPPING_POLICY(tmp, ORTE_MAPPING_BYSLOT);
} else if (0 == strncasecmp(ck[0], "node", len)) {
ORTE_SET_MAPPING_POLICY(tmp, ORTE_MAPPING_BYNODE);
#if OPAL_HAVE_HWLOC
} else if (0 == strncasecmp(ck[0], "core", len)) {
ORTE_SET_MAPPING_POLICY(tmp, ORTE_MAPPING_BYCORE);
} else if (0 == strncasecmp(ck[0], "l1cache", len)) {
ORTE_SET_MAPPING_POLICY(tmp, ORTE_MAPPING_BYL1CACHE);
} else if (0 == strncasecmp(ck[0], "l2cache", len)) {
ORTE_SET_MAPPING_POLICY(tmp, ORTE_MAPPING_BYL2CACHE);
} else if (0 == strncasecmp(ck[0], "l3cache", len)) {
ORTE_SET_MAPPING_POLICY(tmp, ORTE_MAPPING_BYL3CACHE);
} else if (0 == strncasecmp(ck[0], "socket", len)) {
ORTE_SET_MAPPING_POLICY(tmp, ORTE_MAPPING_BYSOCKET);
} else if (0 == strncasecmp(ck[0], "numa", len)) {
ORTE_SET_MAPPING_POLICY(tmp, ORTE_MAPPING_BYNUMA);
} else if (0 == strncasecmp(ck[0], "board", len)) {
ORTE_SET_MAPPING_POLICY(tmp, ORTE_MAPPING_BYBOARD);
} else if (0 == strncasecmp(ck[0], "hwthread", len)) {
ORTE_SET_MAPPING_POLICY(tmp, ORTE_MAPPING_BYHWTHREAD);
/* if we are mapping processes to individual hwthreads, then
* we need to treat those hwthreads as separate cpus
*/
opal_hwloc_use_hwthreads_as_cpus = true;
#endif
} else {
orte_show_help("help-orte-rmaps-base.txt", "unrecognized-policy", true, "mapping", rmaps_base_mapping_policy);
opal_argv_free(ck);
return ORTE_ERR_SILENT;
}
opal_argv_free(ck);
ORTE_SET_MAPPING_DIRECTIVE(tmp, ORTE_MAPPING_GIVEN);
}
#if OPAL_HAVE_HWLOC
setpolicy:
#endif
*policy = tmp;
return ORTE_SUCCESS;
}
static int ppr_mapper(orte_job_t *jdata)
{
int rc = ORTE_SUCCESS, j, n;
mca_base_component_t *c=&mca_rmaps_ppr_component.base_version;
orte_node_t *node;
orte_proc_t *proc;
orte_app_context_t *app;
orte_vpid_t total_procs, nprocs_mapped;
opal_hwloc_level_t start=OPAL_HWLOC_NODE_LEVEL;
#if OPAL_HAVE_HWLOC
hwloc_obj_t obj;
hwloc_obj_type_t lowest;
unsigned cache_level=0;
unsigned int nobjs, i;
bool pruning_reqd = false;
opal_hwloc_level_t level;
#endif
opal_list_t node_list;
opal_list_item_t *item;
orte_std_cntr_t num_slots;
orte_app_idx_t idx;
char **ppr_req, **ck;
size_t len;
bool initial_map=true;
/* only handle initial launch of loadbalanced
* or NPERxxx jobs - allow restarting of failed apps
*/
if (ORTE_FLAG_TEST(jdata, ORTE_JOB_FLAG_RESTART)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:ppr: job %s being restarted - ppr cannot map",
ORTE_JOBID_PRINT(jdata->jobid));
return ORTE_ERR_TAKE_NEXT_OPTION;
}
if (NULL != jdata->map->req_mapper &&
0 != strcasecmp(jdata->map->req_mapper, c->mca_component_name)) {
/* a mapper has been specified, and it isn't me */
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:ppr: job %s not using ppr mapper",
ORTE_JOBID_PRINT(jdata->jobid));
return ORTE_ERR_TAKE_NEXT_OPTION;
}
if (NULL == jdata->map->ppr ||
ORTE_MAPPING_PPR != ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
/* not for us */
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:ppr: job %s not using ppr mapper",
ORTE_JOBID_PRINT(jdata->jobid));
return ORTE_ERR_TAKE_NEXT_OPTION;
}
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:ppr: mapping job %s with ppr %s",
ORTE_JOBID_PRINT(jdata->jobid), jdata->map->ppr);
/* flag that I did the mapping */
if (NULL != jdata->map->last_mapper) {
free(jdata->map->last_mapper);
}
jdata->map->last_mapper = strdup(c->mca_component_name);
/* initialize */
memset(ppr, 0, OPAL_HWLOC_HWTHREAD_LEVEL * sizeof(opal_hwloc_level_t));
/* parse option */
n=0;
ppr_req = opal_argv_split(jdata->map->ppr, ',');
for (j=0; NULL != ppr_req[j]; j++) {
/* split on the colon */
ck = opal_argv_split(ppr_req[j], ':');
if (2 != opal_argv_count(ck)) {
/* must provide a specification */
orte_show_help("help-orte-rmaps-ppr.txt", "invalid-ppr", true, jdata->map->ppr);
opal_argv_free(ppr_req);
opal_argv_free(ck);
return ORTE_ERR_SILENT;
}
len = strlen(ck[1]);
if (0 == strncasecmp(ck[1], "node", len)) {
ppr[OPAL_HWLOC_NODE_LEVEL] = strtol(ck[0], NULL, 10);
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYNODE);
start = OPAL_HWLOC_NODE_LEVEL;
n++;
#if OPAL_HAVE_HWLOC
} else if (0 == strncasecmp(ck[1], "hwthread", len) ||
0 == strncasecmp(ck[1], "thread", len)) {
ppr[OPAL_HWLOC_HWTHREAD_LEVEL] = strtol(ck[0], NULL, 10);
start = OPAL_HWLOC_HWTHREAD_LEVEL;
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYHWTHREAD);
n++;
} else if (0 == strncasecmp(ck[1], "core", len)) {
ppr[OPAL_HWLOC_CORE_LEVEL] = strtol(ck[0], NULL, 10);
if (start < OPAL_HWLOC_CORE_LEVEL) {
start = OPAL_HWLOC_CORE_LEVEL;
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYCORE);
}
n++;
} else if (0 == strncasecmp(ck[1], "socket", len) ||
0 == strncasecmp(ck[1], "skt", len)) {
ppr[OPAL_HWLOC_SOCKET_LEVEL] = strtol(ck[0], NULL, 10);
if (start < OPAL_HWLOC_SOCKET_LEVEL) {
start = OPAL_HWLOC_SOCKET_LEVEL;
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYSOCKET);
}
n++;
} else if (0 == strncasecmp(ck[1], "l1cache", len)) {
ppr[OPAL_HWLOC_L1CACHE_LEVEL] = strtol(ck[0], NULL, 10);
if (start < OPAL_HWLOC_L1CACHE_LEVEL) {
start = OPAL_HWLOC_L1CACHE_LEVEL;
cache_level = 1;
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYL1CACHE);
}
n++;
} else if (0 == strncasecmp(ck[1], "l2cache", len)) {
ppr[OPAL_HWLOC_L2CACHE_LEVEL] = strtol(ck[0], NULL, 10);
if (start < OPAL_HWLOC_L2CACHE_LEVEL) {
start = OPAL_HWLOC_L2CACHE_LEVEL;
cache_level = 2;
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYL2CACHE);
}
n++;
} else if (0 == strncasecmp(ck[1], "l3cache", len)) {
ppr[OPAL_HWLOC_L3CACHE_LEVEL] = strtol(ck[0], NULL, 10);
if (start < OPAL_HWLOC_L3CACHE_LEVEL) {
start = OPAL_HWLOC_L3CACHE_LEVEL;
cache_level = 3;
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYL3CACHE);
}
n++;
} else if (0 == strncasecmp(ck[1], "numa", len)) {
ppr[OPAL_HWLOC_NUMA_LEVEL] = strtol(ck[0], NULL, 10);
if (start < OPAL_HWLOC_NUMA_LEVEL) {
start = OPAL_HWLOC_NUMA_LEVEL;
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYNUMA);
}
n++;
#endif
} else {
/* unknown spec */
orte_show_help("help-orte-rmaps-ppr.txt", "unrecognized-ppr-option", true, ck[1], jdata->map->ppr);
opal_argv_free(ppr_req);
opal_argv_free(ck);
return ORTE_ERR_SILENT;
}
opal_argv_free(ck);
}
opal_argv_free(ppr_req);
/* if nothing was given, that's an error */
if (0 == n) {
opal_output(0, "NOTHING GIVEN");
return ORTE_ERR_SILENT;
}
#if OPAL_HAVE_HWLOC
/* if more than one level was specified, then pruning will be reqd */
if (1 < n) {
pruning_reqd = true;
}
#endif
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:ppr: job %s assigned policy %s",
ORTE_JOBID_PRINT(jdata->jobid),
orte_rmaps_base_print_mapping(jdata->map->mapping));
#if OPAL_HAVE_HWLOC
/* convenience */
level = start;
lowest = opal_hwloc_levels[start];
#endif
for (idx=0; idx < (orte_app_idx_t)jdata->apps->size; idx++) {
if (NULL == (app = (orte_app_context_t*)opal_pointer_array_get_item(jdata->apps, idx))) {
continue;
}
/* if the number of total procs was given, set that
* limit - otherwise, set to max so we simply fill
* all the nodes with the pattern
*/
if (0 < app->num_procs) {
total_procs = app->num_procs;
} else {
total_procs = ORTE_VPID_MAX;
}
/* get the available nodes */
OBJ_CONSTRUCT(&node_list, opal_list_t);
if(ORTE_SUCCESS != (rc = orte_rmaps_base_get_target_nodes(&node_list, &num_slots, app,
jdata->map->mapping, initial_map, false))) {
ORTE_ERROR_LOG(rc);
goto error;
}
/* flag that all subsequent requests should not reset the node->mapped flag */
initial_map = false;
/* if a bookmark exists from some prior mapping, set us to start there */
jdata->bookmark = orte_rmaps_base_get_starting_point(&node_list, jdata);
/* cycle across the nodes */
nprocs_mapped = 0;
for (item = opal_list_get_first(&node_list);
item != opal_list_get_end(&node_list);
item = opal_list_get_next(item)) {
node = (orte_node_t*)item;
#if OPAL_HAVE_HWLOC
/* bozo check */
if (NULL == node->topology) {
orte_show_help("help-orte-rmaps-ppr.txt", "ppr-topo-missing",
true, node->name);
rc = ORTE_ERR_SILENT;
goto error;
}
#endif
/* add the node to the map, if needed */
if (!ORTE_FLAG_TEST(node, ORTE_NODE_FLAG_MAPPED)) {
if (ORTE_SUCCESS > (rc = opal_pointer_array_add(jdata->map->nodes, (void*)node))) {
ORTE_ERROR_LOG(rc);
goto error;
}
ORTE_FLAG_SET(node, ORTE_NODE_FLAG_MAPPED);
OBJ_RETAIN(node); /* maintain accounting on object */
jdata->map->num_nodes++;
}
/* if we are mapping solely at the node level, just put
* that many procs on this node
*/
if (OPAL_HWLOC_NODE_LEVEL == start) {
#if OPAL_HAVE_HWLOC
obj = hwloc_get_root_obj(node->topology);
#endif
for (j=0; j < ppr[start] && nprocs_mapped < total_procs; j++) {
if (NULL == (proc = orte_rmaps_base_setup_proc(jdata, node, idx))) {
rc = ORTE_ERR_OUT_OF_RESOURCE;
goto error;
}
nprocs_mapped++;
#if OPAL_HAVE_HWLOC
orte_set_attribute(&proc->attributes, ORTE_PROC_HWLOC_LOCALE, ORTE_ATTR_LOCAL, obj, OPAL_PTR);
#endif
}
#if OPAL_HAVE_HWLOC
} else {
/* get the number of lowest resources on this node */
nobjs = opal_hwloc_base_get_nbobjs_by_type(node->topology,
lowest, cache_level,
OPAL_HWLOC_AVAILABLE);
/* map the specified number of procs to each such resource on this node,
* recording the locale of each proc so we know its cpuset
*/
for (i=0; i < nobjs; i++) {
obj = opal_hwloc_base_get_obj_by_type(node->topology,
lowest, cache_level,
i, OPAL_HWLOC_AVAILABLE);
for (j=0; j < ppr[start] && nprocs_mapped < total_procs; j++) {
if (NULL == (proc = orte_rmaps_base_setup_proc(jdata, node, idx))) {
rc = ORTE_ERR_OUT_OF_RESOURCE;
goto error;
}
nprocs_mapped++;
orte_set_attribute(&proc->attributes, ORTE_PROC_HWLOC_LOCALE, ORTE_ATTR_LOCAL, obj, OPAL_PTR);
}
}
if (pruning_reqd) {
/* go up the ladder and prune the procs according to
* the specification, adjusting the count of procs on the
* node as we go
*/
level--;
prune(jdata->jobid, idx, node, &level, &nprocs_mapped);
}
#endif
}
/* set the total slots used */
if ((int)node->num_procs <= node->slots) {
node->slots_inuse = (int)node->num_procs;
} else {
node->slots_inuse = node->slots;
}
/* if no-oversubscribe was specified, check to see if
* we have violated the total slot specification - regardless,
* if slots_max was given, we are not allowed to violate it!
*/
if ((node->slots < (int)node->num_procs) ||
(0 < node->slots_max && node->slots_max < (int)node->num_procs)) {
if (ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, node->num_procs, app->app);
rc = ORTE_ERR_SILENT;
goto error;
}
/* flag the node as oversubscribed so that sched-yield gets
* properly set
*/
ORTE_FLAG_SET(node, ORTE_NODE_FLAG_OVERSUBSCRIBED);
}
/* if we haven't mapped all the procs, continue on to the
* next node
*/
if (total_procs == nprocs_mapped) {
break;
}
}
if (0 == app->num_procs) {
app->num_procs = nprocs_mapped;
}
if (ORTE_VPID_MAX != total_procs && nprocs_mapped < total_procs) {
/* couldn't map them all */
orte_show_help("help-orte-rmaps-ppr.txt", "ppr-too-many-procs",
true, app->app, app->num_procs, jdata->map->ppr);
rc = ORTE_ERR_SILENT;
goto error;
}
/* compute vpids and add proc objects to the job */
if (ORTE_SUCCESS != (rc = orte_rmaps_base_compute_vpids(jdata, app, &node_list))) {
ORTE_ERROR_LOG(rc);
goto error;
}
/* track the total number of processes we mapped - must update
* this AFTER we compute vpids so that computation is done
* correctly
*/
jdata->num_procs += app->num_procs;
while (NULL != (item = opal_list_remove_first(&node_list))) {
OBJ_RELEASE(item);
}
OBJ_DESTRUCT(&node_list);
}
return ORTE_SUCCESS;
error:
while (NULL != (item = opal_list_remove_first(&node_list))) {
OBJ_RELEASE(item);
}
OBJ_DESTRUCT(&node_list);
return rc;
}
int orte_rmaps_rr_byslot(orte_job_t *jdata,
orte_app_context_t *app,
opal_list_t *node_list,
orte_std_cntr_t num_slots,
orte_vpid_t num_procs)
{
int i, nprocs_mapped;
orte_node_t *node;
orte_proc_t *proc;
int num_procs_to_assign, extra_procs_to_assign=0, nxtra_nodes=0;
hwloc_obj_t obj=NULL;
float balance;
bool add_one=false;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr: mapping by slot for job %s slots %d num_procs %lu",
ORTE_JOBID_PRINT(jdata->jobid), (int)num_slots, (unsigned long)num_procs);
/* check to see if we can map all the procs */
if (num_slots < (int)app->num_procs) {
if (ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, app->num_procs, app->app, orte_process_info.nodename);
ORTE_UPDATE_EXIT_STATUS(ORTE_ERROR_DEFAULT_EXIT_CODE);
return ORTE_ERR_SILENT;
}
}
/* first pass: map the number of procs to each node until we
* map all specified procs or use all allocated slots
*/
nprocs_mapped = 0;
OPAL_LIST_FOREACH(node, node_list, orte_node_t) {
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot working node %s",
node->name);
/* get the root object as we are not assigning
* locale here except at the node level
*/
if (NULL != node->topology && NULL != node->topology->topo) {
obj = hwloc_get_root_obj(node->topology->topo);
}
if (node->slots <= node->slots_inuse) {
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot node %s is full - skipping",
node->name);
continue;
}
/* assign a number of procs equal to the number of available slots */
num_procs_to_assign = node->slots - node->slots_inuse;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot assigning %d procs to node %s",
(int)num_procs_to_assign, node->name);
for (i=0; i < num_procs_to_assign && nprocs_mapped < app->num_procs; i++) {
/* add this node to the map - do it only once */
if (!ORTE_FLAG_TEST(node, ORTE_NODE_FLAG_MAPPED)) {
ORTE_FLAG_SET(node, ORTE_NODE_FLAG_MAPPED);
++(jdata->map->num_nodes);
}
if (NULL == (proc = orte_rmaps_base_setup_proc(jdata, node, app->idx))) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
nprocs_mapped++;
orte_set_attribute(&proc->attributes, ORTE_PROC_HWLOC_LOCALE, ORTE_ATTR_LOCAL, obj, OPAL_PTR);
}
}
/**
* Function for finding and opening either all MCA components, or the one
* that was specifically requested via a MCA parameter.
*/
static int orte_rmaps_base_open(mca_base_open_flag_t flags)
{
int rc;
/* init the globals */
OBJ_CONSTRUCT(&orte_rmaps_base.selected_modules, opal_list_t);
orte_rmaps_base.ppr = NULL;
orte_rmaps_base.slot_list = NULL;
orte_rmaps_base.mapping = 0;
orte_rmaps_base.ranking = 0;
#if OPAL_HAVE_HWLOC
/* if a topology file was given, then set our topology
* from it. Even though our actual topology may differ,
* mpirun only needs to see the compute node topology
* for mapping purposes
*/
if (NULL != rmaps_base_topo_file) {
if (OPAL_SUCCESS != (rc = opal_hwloc_base_set_topology(rmaps_base_topo_file))) {
orte_show_help("help-orte-rmaps-base.txt", "topo-file", true, rmaps_base_topo_file);
return ORTE_ERR_SILENT;
}
}
#endif
if (ORTE_SUCCESS != (rc = orte_rmaps_base_set_mapping_policy(&orte_rmaps_base.mapping,
&orte_rmaps_base.device,
rmaps_base_mapping_policy))) {
return rc;
}
if (ORTE_SUCCESS != (rc = orte_rmaps_base_set_ranking_policy(&orte_rmaps_base.ranking,
orte_rmaps_base.mapping,
rmaps_base_ranking_policy))) {
return rc;
}
if (rmaps_base_byslot) {
/* set mapping policy to byslot - error if something else already set */
if ((ORTE_MAPPING_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) &&
ORTE_GET_MAPPING_POLICY(orte_rmaps_base.mapping) != ORTE_MAPPING_BYSLOT) {
/* error - cannot redefine the default mapping policy */
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "mapping",
"byslot", orte_rmaps_base_print_mapping(orte_rmaps_base.mapping));
return ORTE_ERR_SILENT;
}
ORTE_SET_MAPPING_POLICY(orte_rmaps_base.mapping, ORTE_MAPPING_BYSLOT);
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_GIVEN);
/* set ranking policy to byslot - error if something else already set */
if ((ORTE_RANKING_GIVEN & ORTE_GET_RANKING_DIRECTIVE(orte_rmaps_base.ranking)) &&
ORTE_GET_RANKING_POLICY(orte_rmaps_base.ranking) != ORTE_RANK_BY_SLOT) {
/* error - cannot redefine the default ranking policy */
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "ranking",
"byslot", orte_rmaps_base_print_ranking(orte_rmaps_base.ranking));
return ORTE_ERR_SILENT;
}
ORTE_SET_RANKING_POLICY(orte_rmaps_base.ranking, ORTE_RANK_BY_SLOT);
ORTE_SET_RANKING_DIRECTIVE(orte_rmaps_base.ranking, ORTE_RANKING_GIVEN);
}
if (rmaps_base_bynode) {
/* set mapping policy to bynode - error if something else already set */
if ((ORTE_MAPPING_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) &&
ORTE_GET_MAPPING_POLICY(orte_rmaps_base.mapping) != ORTE_MAPPING_BYNODE) {
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "mapping",
"bynode", orte_rmaps_base_print_mapping(orte_rmaps_base.mapping));
return ORTE_ERR_SILENT;
}
ORTE_SET_MAPPING_POLICY(orte_rmaps_base.mapping, ORTE_MAPPING_BYNODE);
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_GIVEN);
/* set ranking policy to bynode - error if something else already set */
if ((ORTE_RANKING_GIVEN & ORTE_GET_RANKING_DIRECTIVE(orte_rmaps_base.ranking)) &&
ORTE_GET_RANKING_POLICY(orte_rmaps_base.ranking) != ORTE_RANK_BY_NODE) {
/* error - cannot redefine the default ranking policy */
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "ranking",
"bynode", orte_rmaps_base_print_ranking(orte_rmaps_base.ranking));
return ORTE_ERR_SILENT;
}
ORTE_SET_RANKING_POLICY(orte_rmaps_base.ranking, ORTE_RANK_BY_NODE);
ORTE_SET_RANKING_DIRECTIVE(orte_rmaps_base.ranking, ORTE_RANKING_GIVEN);
}
/* Should we schedule on the local node or not? */
if (rmaps_base_no_schedule_local) {
orte_rmaps_base.mapping |= ORTE_MAPPING_NO_USE_LOCAL;
}
/* Should we oversubscribe or not? */
if (rmaps_base_no_oversubscribe) {
if ((ORTE_MAPPING_SUBSCRIBE_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) &&
!(ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping))) {
/* error - cannot redefine the default mapping policy */
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "mapping",
"no-oversubscribe", orte_rmaps_base_print_mapping(orte_rmaps_base.mapping));
return ORTE_ERR_SILENT;
}
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_NO_OVERSUBSCRIBE);
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_SUBSCRIBE_GIVEN);
}
/** force oversubscription permission */
if (rmaps_base_oversubscribe) {
if ((ORTE_MAPPING_SUBSCRIBE_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) &&
(ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping))) {
/* error - cannot redefine the default mapping policy */
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "mapping",
"oversubscribe", orte_rmaps_base_print_mapping(orte_rmaps_base.mapping));
return ORTE_ERR_SILENT;
}
ORTE_UNSET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_NO_OVERSUBSCRIBE);
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_SUBSCRIBE_GIVEN);
}
/* should we display a detailed (developer-quality) version of the map after determining it? */
if (rmaps_base_display_devel_map) {
orte_rmaps_base.display_map = true;
orte_devel_level_output = true;
}
/* should we display a diffable report of proc locations after determining it? */
if (rmaps_base_display_diffable_map) {
orte_rmaps_base.display_map = true;
orte_display_diffable_output = true;
}
/* Open up all available components */
rc = mca_base_framework_components_open(&orte_rmaps_base_framework, flags);
/* check to see if any component indicated a problem */
if (ORTE_MAPPING_CONFLICTED & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) {
/* the component would have already reported the error, so
* tell the rest of the chain to shut up
*/
return ORTE_ERR_SILENT;
}
/* All done */
return rc;
}
static int byobj_span(orte_job_t *jdata,
orte_app_context_t *app,
opal_list_t *node_list,
orte_std_cntr_t num_slots,
orte_vpid_t num_procs,
hwloc_obj_type_t target, unsigned cache_level)
{
int i, j, nprocs_mapped, lag, delta, navg;
orte_node_t *node;
orte_proc_t *proc;
opal_list_item_t *item;
int num_procs_to_assign, nperobj, nprocs, nxtra_objs=0;
int extra_procs_to_assign=0, nxtra_nodes=0, idx;
hwloc_obj_t obj=NULL;
unsigned int nobjs;
float balance;
bool add_one=false;
bool oversubscribed=false;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr: mapping span by %s for job %s slots %d num_procs %lu",
hwloc_obj_type_string(target),
ORTE_JOBID_PRINT(jdata->jobid),
(int)num_slots, (unsigned long)num_procs);
/* quick check to see if we can map all the procs - can't
* do more because we don't know how many total objects exist
* across all the nodes
*/
if (num_slots < (int)app->num_procs) {
if (ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, app->num_procs, app->app);
return ORTE_ERR_SILENT;
}
oversubscribed = true;
}
/* divide the procs evenly across all nodes - this is the
* average we have to maintain as we go, but we adjust
* the number on each node to reflect its available slots.
* Obviously, if all nodes have the same number of slots,
* then the avg is what we get on each node - this is
* the most common situation.
*/
navg = app->num_procs / opal_list_get_size(node_list);
if (0 == navg) {
/* if there are less procs than nodes, we have to
* place at least one/node
*/
navg = 1;
}
/* compute how many extra procs to put on each node */
balance = (float)((jdata->num_procs + app->num_procs) - (navg * opal_list_get_size(node_list))) / (float)opal_list_get_size(node_list);
extra_procs_to_assign = (int)balance;
if (0 < (balance - (float)extra_procs_to_assign)) {
/* compute how many nodes need an extra proc */
nxtra_nodes = (jdata->num_procs + app->num_procs) - ((navg + extra_procs_to_assign) * opal_list_get_size(node_list));
/* add one so that we add an extra proc to the first nodes
* until all procs are mapped
*/
extra_procs_to_assign++;
/* flag that we added one */
add_one = true;
}
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr: mapping by %s navg %d extra_procs %d extra_nodes %d",
hwloc_obj_type_string(target),
navg, extra_procs_to_assign, nxtra_nodes);
nprocs_mapped = 0;
lag = 0;
for (item = opal_list_get_first(node_list);
item != opal_list_get_end(node_list);
item = opal_list_get_next(item)) {
node = (orte_node_t*)item;
/* bozo check */
if (NULL == node->topology) {
orte_show_help("help-orte-rmaps-ppr.txt", "ppr-topo-missing",
true, node->name);
return ORTE_ERR_SILENT;
}
/* add this node to the map, if reqd */
if (!node->mapped) {
if (ORTE_SUCCESS > (idx = opal_pointer_array_add(jdata->map->nodes, (void*)node))) {
ORTE_ERROR_LOG(idx);
return idx;
}
node->mapped = true;
OBJ_RETAIN(node); /* maintain accounting on object */
++(jdata->map->num_nodes);
}
/* compute the number of procs to go on this node */
if (add_one) {
if (0 == nxtra_nodes) {
--extra_procs_to_assign;
add_one = false;
} else {
--nxtra_nodes;
}
}
if (oversubscribed) {
/* everybody just takes their share */
num_procs_to_assign = navg + extra_procs_to_assign;
} else {
/* if we are not oversubscribed, then there are enough
* slots to handle all the procs. However, not every
* node will have the same number of slots, so we
* have to track how many procs to "shift" elsewhere
* to make up the difference
*/
if (node->slots <= node->slots_inuse) {
/* if there are no extras to take, then we can
* safely remove this node as we don't need it
*/
if (0 == extra_procs_to_assign) {
opal_pointer_array_set_item(jdata->map->nodes, node->index, NULL);
OBJ_RELEASE(node);
--(jdata->map->num_nodes);
/* update how many we are lagging behind */
lag += navg;
continue;
}
/* everybody has to take at least the extras */
num_procs_to_assign = extra_procs_to_assign;
/* update how many we are lagging behind */
lag += navg;
} else {
/* if slots < avg, then take all */
if ((node->slots - node->slots_inuse) < navg) {
num_procs_to_assign = (node->slots - node->slots_inuse) + extra_procs_to_assign;
/* update how many we are lagging behind */
lag += navg - (node->slots - node->slots_inuse);
} else {
/* take the avg plus as much of the "lag" as we can */
delta = 0;
if (0 < lag) {
delta = (node->slots - node->slots_inuse) - navg;
if (lag < delta) {
delta = lag;
}
lag -= delta;
}
num_procs_to_assign = navg + delta + extra_procs_to_assign;
}
}
}
/* get the number of objects of this type on this node */
nobjs = opal_hwloc_base_get_nbobjs_by_type(node->topology, target, cache_level, OPAL_HWLOC_AVAILABLE);
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:byobj: found %d objs on node %s", nobjs, node->name);
/* compute the number of procs to go on each object */
nperobj = num_procs_to_assign / nobjs;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:byobj: placing %d procs on each object", nperobj);
if ((int)(nperobj * nobjs) < num_procs_to_assign) {
/* compute how many objs need an extra proc */
nxtra_objs = num_procs_to_assign - nperobj * nobjs;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:byobj: adding 1 extra proc to the first %d objects, if needed", nxtra_objs);
}
/* loop through the number of objects */
for (i=0; i < (int)nobjs && nprocs_mapped < (int)app->num_procs; i++) {
/* get the hwloc object */
if (NULL == (obj = opal_hwloc_base_get_obj_by_type(node->topology, target, cache_level, i, OPAL_HWLOC_AVAILABLE))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
return ORTE_ERR_NOT_FOUND;
}
/* map the reqd number of procs */
if (0 < nxtra_objs) {
nprocs = nperobj + 1;
--nxtra_objs;
} else {
nprocs = nperobj;
}
for (j=0; j < nprocs && nprocs_mapped < app->num_procs; j++) {
if (NULL == (proc = orte_rmaps_base_setup_proc(jdata, node, app->idx))) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
nprocs_mapped++;
proc->locale = obj;
}
/* keep track of the node we last used */
jdata->bookmark = node;
}
/* not all nodes are equal, so only set oversubscribed for
* this node if it is in that state
*/
if (node->slots < (int)node->num_procs) {
/* flag the node as oversubscribed so that sched-yield gets
* properly set
*/
node->oversubscribed = true;
}
if (nprocs_mapped == app->num_procs) {
/* we are done */
break;
}
}
return ORTE_SUCCESS;
}
/* mapping by hwloc object looks a lot like mapping by node,
* but has the added complication of possibly having different
* numbers of objects on each node
*/
int orte_rmaps_rr_byobj(orte_job_t *jdata,
orte_app_context_t *app,
opal_list_t *node_list,
orte_std_cntr_t num_slots,
orte_vpid_t num_procs,
hwloc_obj_type_t target, unsigned cache_level)
{
int i, j, nprocs_mapped;
orte_node_t *node;
orte_proc_t *proc;
opal_list_item_t *item;
int num_procs_to_assign, nperobj, nprocs, nxtra_objs=0;
int extra_procs_to_assign=0, nxtra_nodes=0, idx;
hwloc_obj_t obj=NULL;
unsigned int nobjs;
float balance;
bool add_one=false;
/* there are two modes for mapping by object: span and not-span. The
* span mode essentially operates as if there was just a single
* "super-node" in the system - i.e., it balances the load across
* all objects of the indicated type regardless of their location.
* In essence, it acts as if we placed one proc on each object, cycling
* across all objects on all nodes, and then wrapped around to place
* another proc on each object, doing so until all procs were placed.
*
* In contrast, the non-span mode operates similar to byslot mapping.
* All slots on each node are filled, assigning each proc to an object
* on that node in a balanced fashion, and then the mapper moves on
* to the next node. Thus, procs tend to be "front loaded" onto the
* list of nodes, as opposed to being "load balanced" in the span mode
*/
if (ORTE_MAPPING_SPAN & jdata->map->mapping) {
return byobj_span(jdata, app, node_list, num_slots,
num_procs, target, cache_level);
}
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr: mapping no-span by %s for job %s slots %d num_procs %lu",
hwloc_obj_type_string(target),
ORTE_JOBID_PRINT(jdata->jobid),
(int)num_slots, (unsigned long)num_procs);
/* quick check to see if we can map all the procs - can't
* do more because we don't know how many total objects exist
* across all the nodes
*/
if (num_slots < (app->num_procs * orte_rmaps_base.cpus_per_rank)) {
if (ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, app->num_procs, app->app);
return ORTE_ERR_SILENT;
}
/* compute how many extra procs to put on each node */
if (1 == opal_list_get_size(node_list)) {
/* if there is only one node, then they all have to go on it */
extra_procs_to_assign = app->num_procs;
} else {
balance = (float)(((jdata->num_procs + app->num_procs)*orte_rmaps_base.cpus_per_rank) - num_slots) / (float)opal_list_get_size(node_list);
extra_procs_to_assign = (int)balance;
if (0 < (balance - (float)extra_procs_to_assign)) {
/* compute how many nodes need an extra proc */
nxtra_nodes = ((jdata->num_procs + app->num_procs)*orte_rmaps_base.cpus_per_rank) - num_slots - (extra_procs_to_assign * opal_list_get_size(node_list));
/* add one so that we add an extra proc to the first nodes
* until all procs are mapped
*/
extra_procs_to_assign++;
/* flag that we added one */
add_one = true;
}
}
}
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr: mapping no-span by %s extra_procs %d extra_nodes %d",
hwloc_obj_type_string(target),
extra_procs_to_assign, nxtra_nodes);
nprocs_mapped = 0;
for (item = opal_list_get_first(node_list);
item != opal_list_get_end(node_list);
item = opal_list_get_next(item)) {
node = (orte_node_t*)item;
/* bozo check */
if (NULL == node->topology) {
orte_show_help("help-orte-rmaps-ppr.txt", "ppr-topo-missing",
true, node->name);
return ORTE_ERR_SILENT;
}
/* add this node to the map, if reqd */
if (!node->mapped) {
if (ORTE_SUCCESS > (idx = opal_pointer_array_add(jdata->map->nodes, (void*)node))) {
ORTE_ERROR_LOG(idx);
return idx;
}
node->mapped = true;
OBJ_RETAIN(node); /* maintain accounting on object */
++(jdata->map->num_nodes);
}
/* compute the number of procs to go on this node */
if (add_one) {
if (0 == nxtra_nodes) {
--extra_procs_to_assign;
add_one = false;
} else {
--nxtra_nodes;
}
}
if (node->slots <= node->slots_inuse) {
/* everybody takes at least the extras */
num_procs_to_assign = extra_procs_to_assign;
} else {
num_procs_to_assign = (node->slots - node->slots_inuse)/orte_rmaps_base.cpus_per_rank + extra_procs_to_assign;
if (app->num_procs < num_procs_to_assign) {
/* might have more slots than procs */
num_procs_to_assign = app->num_procs;
}
}
/* get the number of objects of this type on this node */
nobjs = opal_hwloc_base_get_nbobjs_by_type(node->topology, target, cache_level, OPAL_HWLOC_AVAILABLE);
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:byobj: nprocs-to-assign %d for %d objs on node %s", num_procs_to_assign, nobjs, node->name);
/* if there are no objects of this type, then report the error
* and abort - this can happen, for example, on systems that
* don't report "sockets" as an independent object. However, IF
* this object is the default one - i.e., not specified by the
* user - then we can fall back to mapping by slot
*/
if (0 == nobjs) {
if (ORTE_MAPPING_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:no-objects",
true, hwloc_obj_type_string(target), node->name);
return ORTE_ERR_SILENT;
} else {
/* this was the default mapping policy, so clear the map
* of any prior work and indicate that map-by slot is reqd
*/
for (i=0; i < jdata->map->nodes->size; i++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(jdata->map->nodes, i))) {
continue;
}
for (idx=0; idx < node->procs->size; idx++) {
if (NULL == (proc = (orte_proc_t*)opal_pointer_array_get_item(node->procs, idx))) {
continue;
}
if (proc->name.jobid != jdata->jobid) {
continue;
}
--node->num_procs;
OBJ_RELEASE(proc);
opal_pointer_array_set_item(node->procs, idx, NULL);
}
if (0 == node->num_procs) {
node->mapped = false;
OBJ_RELEASE(node);
opal_pointer_array_set_item(jdata->map->nodes, i, NULL);
}
}
return ORTE_ERR_NOT_SUPPORTED;
}
}
/* compute the number of procs to go on each object */
nperobj = num_procs_to_assign / nobjs;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:byobj: placing %d procs on each object", nperobj);
if ((int)(nperobj * nobjs) < num_procs_to_assign) {
/* compute how many objs need an extra proc */
nxtra_objs = num_procs_to_assign - nperobj * nobjs;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:byobj: adding 1 extra proc to the first %d objects, if needed", nxtra_objs);
}
/* loop through the number of objects */
for (i=0; i < (int)nobjs && nprocs_mapped < (int)app->num_procs; i++) {
/* get the hwloc object */
if (NULL == (obj = opal_hwloc_base_get_obj_by_type(node->topology, target, cache_level, i, OPAL_HWLOC_AVAILABLE))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
return ORTE_ERR_NOT_FOUND;
}
/* map the reqd number of procs */
if (0 < nxtra_objs) {
nprocs = nperobj + 1;
--nxtra_objs;
} else {
nprocs = nperobj;
}
for (j=0; j < nprocs && nprocs_mapped < app->num_procs; j++) {
if (NULL == (proc = orte_rmaps_base_setup_proc(jdata, node, app->idx))) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
nprocs_mapped++;
proc->locale = obj;
}
}
/* not all nodes are equal, so only set oversubscribed for
* this node if it is in that state
*/
if (node->slots < (int)node->num_procs) {
/* flag the node as oversubscribed so that sched-yield gets
* properly set
*/
node->oversubscribed = true;
}
if (nprocs_mapped == app->num_procs) {
/* we are done */
break;
}
}
return ORTE_SUCCESS;
}
int orte_rmaps_rr_byslot(orte_job_t *jdata,
orte_app_context_t *app,
opal_list_t *node_list,
orte_std_cntr_t num_slots,
orte_vpid_t num_procs)
{
int rc, i, nprocs_mapped;
orte_node_t *node;
orte_proc_t *proc;
opal_list_item_t *item;
int num_procs_to_assign, extra_procs_to_assign=0, nxtra_nodes=0;
#if OPAL_HAVE_HWLOC
hwloc_obj_t obj=NULL;
#endif
float balance;
bool add_one=false;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr: mapping by slot for job %s slots %d num_procs %lu",
ORTE_JOBID_PRINT(jdata->jobid), (int)num_slots, (unsigned long)num_procs);
/* check to see if we can map all the procs */
if (num_slots < ((int)app->num_procs * orte_rmaps_base.cpus_per_rank)) {
if (ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, app->num_procs, app->app);
return ORTE_ERR_SILENT;
}
}
/* first pass: map the number of procs to each node until we
* map all specified procs or use all allocated slots
*/
nprocs_mapped = 0;
for (item = opal_list_get_first(node_list);
item != opal_list_get_end(node_list);
item = opal_list_get_next(item)) {
node = (orte_node_t*)item;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot working node %s",
node->name);
#if OPAL_HAVE_HWLOC
/* get the root object as we are not assigning
* locale here except at the node level
*/
if (NULL != node->topology) {
obj = hwloc_get_root_obj(node->topology);
}
#endif
if (node->slots <= node->slots_inuse) {
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot node %s is full - skipping",
node->name);
continue;
}
/* assign a number of procs equal to the number of available
* slots divided by the number of cpus/rank the user
* requested
*/
num_procs_to_assign = (node->slots - node->slots_inuse) / orte_rmaps_base.cpus_per_rank;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot assigning %d procs to node %s",
(int)num_procs_to_assign, node->name);
for (i=0; i < num_procs_to_assign && nprocs_mapped < app->num_procs; i++) {
/* add this node to the map - do it only once */
if (!node->mapped) {
if (ORTE_SUCCESS > (rc = opal_pointer_array_add(jdata->map->nodes, (void*)node))) {
ORTE_ERROR_LOG(rc);
return rc;
}
node->mapped = true;
OBJ_RETAIN(node); /* maintain accounting on object */
++(jdata->map->num_nodes);
}
if (NULL == (proc = orte_rmaps_base_setup_proc(jdata, node, app->idx))) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
nprocs_mapped++;
#if OPAL_HAVE_HWLOC
proc->locale = obj;
#endif
}
}
if (nprocs_mapped == app->num_procs) {
/* we are done */
return ORTE_SUCCESS;
}
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot job %s is oversubscribed - performing second pass",
ORTE_JOBID_PRINT(jdata->jobid));
/* second pass: if we haven't mapped everyone yet, it is
* because we are oversubscribed. Figure out how many procs
* to add
*/
balance = (float)(app->num_procs - nprocs_mapped) / (float)opal_list_get_size(node_list);
extra_procs_to_assign = (int)balance;
if (0 < (balance - (float)extra_procs_to_assign)) {
/* compute how many nodes need an extra proc */
nxtra_nodes = app->num_procs - nprocs_mapped - (extra_procs_to_assign * opal_list_get_size(node_list));
/* add one so that we add an extra proc to the first nodes
* until all procs are mapped
*/
extra_procs_to_assign++;
/* flag that we added one */
add_one = true;
}
for (item = opal_list_get_first(node_list);
item != opal_list_get_end(node_list);
item = opal_list_get_next(item)) {
node = (orte_node_t*)item;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot working node %s",
node->name);
#if OPAL_HAVE_HWLOC
/* get the root object as we are not assigning
* locale except at the node level
*/
if (NULL != node->topology) {
obj = hwloc_get_root_obj(node->topology);
}
#endif
/* add this node to the map - do it only once */
if (!node->mapped) {
if (ORTE_SUCCESS > (rc = opal_pointer_array_add(jdata->map->nodes, (void*)node))) {
ORTE_ERROR_LOG(rc);
return rc;
}
node->mapped = true;
OBJ_RETAIN(node); /* maintain accounting on object */
++(jdata->map->num_nodes);
}
if (add_one) {
if (0 == nxtra_nodes) {
--extra_procs_to_assign;
add_one = false;
} else {
--nxtra_nodes;
}
}
num_procs_to_assign = ((node->slots - node->slots_inuse)/orte_rmaps_base.cpus_per_rank) + extra_procs_to_assign;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot adding up to %d procs to node %s",
num_procs_to_assign, node->name);
for (i=0; i < num_procs_to_assign && nprocs_mapped < app->num_procs; i++) {
if (NULL == (proc = orte_rmaps_base_setup_proc(jdata, node, app->idx))) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
nprocs_mapped++;
#if OPAL_HAVE_HWLOC
proc->locale = obj;
#endif
}
/* not all nodes are equal, so only set oversubscribed for
* this node if it is in that state
*/
if (node->slots < (int)node->num_procs) {
/* flag the node as oversubscribed so that sched-yield gets
* properly set
*/
node->oversubscribed = true;
}
}
return ORTE_SUCCESS;
}
/**
* Function for finding and opening either all MCA components, or the one
* that was specifically requested via a MCA parameter.
*/
static int orte_rmaps_base_open(mca_base_open_flag_t flags)
{
int rc;
/* init the globals */
OBJ_CONSTRUCT(&orte_rmaps_base.selected_modules, opal_list_t);
orte_rmaps_base.slot_list = NULL;
orte_rmaps_base.mapping = 0;
orte_rmaps_base.ranking = 0;
orte_rmaps_base.device = NULL;
/* if a topology file was given, then set our topology
* from it. Even though our actual topology may differ,
* mpirun only needs to see the compute node topology
* for mapping purposes
*/
if (NULL != rmaps_base_topo_file) {
if (OPAL_SUCCESS != (rc = opal_hwloc_base_set_topology(rmaps_base_topo_file))) {
orte_show_help("help-orte-rmaps-base.txt", "topo-file", true, rmaps_base_topo_file);
return ORTE_ERR_SILENT;
}
}
/* check for violations that has to be detected before we parse the mapping option */
if (NULL != orte_rmaps_base.ppr) {
orte_show_help("help-orte-rmaps-base.txt", "deprecated", true,
"--ppr, -ppr", "--map-by ppr:<pattern>",
"rmaps_base_pattern, rmaps_ppr_pattern",
"rmaps_base_mapping_policy=ppr:<pattern>");
/* if the mapping policy is NULL, then we can proceed */
if (NULL == rmaps_base_mapping_policy) {
asprintf(&rmaps_base_mapping_policy, "ppr:%s", orte_rmaps_base.ppr);
} else {
return ORTE_ERR_SILENT;
}
}
if (1 < orte_rmaps_base.cpus_per_rank) {
orte_show_help("help-orte-rmaps-base.txt", "deprecated", true,
"--cpus-per-proc, -cpus-per-proc, --cpus-per-rank, -cpus-per-rank",
"--map-by <obj>:PE=N, default <obj>=NUMA",
"rmaps_base_cpus_per_proc", "rmaps_base_mapping_policy=<obj>:PE=N, default <obj>=NUMA");
}
if (ORTE_SUCCESS != (rc = orte_rmaps_base_set_mapping_policy(&orte_rmaps_base.mapping,
&orte_rmaps_base.device,
rmaps_base_mapping_policy))) {
return rc;
}
if (ORTE_SUCCESS != (rc = orte_rmaps_base_set_ranking_policy(&orte_rmaps_base.ranking,
orte_rmaps_base.mapping,
rmaps_base_ranking_policy))) {
return rc;
}
if (rmaps_base_bycore) {
orte_show_help("help-orte-rmaps-base.txt", "deprecated", true,
"--bycore, -bycore", "--map-by core",
"rmaps_base_bycore", "rmaps_base_mapping_policy=core");
/* set mapping policy to bycore - error if something else already set */
if ((ORTE_MAPPING_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) &&
ORTE_GET_MAPPING_POLICY(orte_rmaps_base.mapping) != ORTE_MAPPING_BYCORE) {
/* error - cannot redefine the default mapping policy */
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "mapping",
"bycore", orte_rmaps_base_print_mapping(orte_rmaps_base.mapping));
return ORTE_ERR_SILENT;
}
ORTE_SET_MAPPING_POLICY(orte_rmaps_base.mapping, ORTE_MAPPING_BYCORE);
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_GIVEN);
/* set ranking policy to bycore - error if something else already set */
if ((ORTE_RANKING_GIVEN & ORTE_GET_RANKING_DIRECTIVE(orte_rmaps_base.ranking)) &&
ORTE_GET_RANKING_POLICY(orte_rmaps_base.ranking) != ORTE_RANK_BY_CORE) {
/* error - cannot redefine the default ranking policy */
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "ranking",
"bycore", orte_rmaps_base_print_ranking(orte_rmaps_base.ranking));
return ORTE_ERR_SILENT;
}
ORTE_SET_RANKING_POLICY(orte_rmaps_base.ranking, ORTE_RANK_BY_CORE);
ORTE_SET_RANKING_DIRECTIVE(orte_rmaps_base.ranking, ORTE_RANKING_GIVEN);
}
if (rmaps_base_byslot) {
orte_show_help("help-orte-rmaps-base.txt", "deprecated", true,
"--byslot, -byslot", "--map-by slot",
"rmaps_base_byslot", "rmaps_base_mapping_policy=slot");
/* set mapping policy to byslot - error if something else already set */
if ((ORTE_MAPPING_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) &&
ORTE_GET_MAPPING_POLICY(orte_rmaps_base.mapping) != ORTE_MAPPING_BYSLOT) {
/* error - cannot redefine the default mapping policy */
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "mapping",
"byslot", orte_rmaps_base_print_mapping(orte_rmaps_base.mapping));
return ORTE_ERR_SILENT;
}
ORTE_SET_MAPPING_POLICY(orte_rmaps_base.mapping, ORTE_MAPPING_BYSLOT);
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_GIVEN);
/* set ranking policy to byslot - error if something else already set */
if ((ORTE_RANKING_GIVEN & ORTE_GET_RANKING_DIRECTIVE(orte_rmaps_base.ranking)) &&
ORTE_GET_RANKING_POLICY(orte_rmaps_base.ranking) != ORTE_RANK_BY_SLOT) {
/* error - cannot redefine the default ranking policy */
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "ranking",
"byslot", orte_rmaps_base_print_ranking(orte_rmaps_base.ranking));
return ORTE_ERR_SILENT;
}
ORTE_SET_RANKING_POLICY(orte_rmaps_base.ranking, ORTE_RANK_BY_SLOT);
ORTE_SET_RANKING_DIRECTIVE(orte_rmaps_base.ranking, ORTE_RANKING_GIVEN);
}
if (rmaps_base_bynode) {
orte_show_help("help-orte-rmaps-base.txt", "deprecated", true,
"--bynode, -bynode", "--map-by node",
"rmaps_base_bynode", "rmaps_base_mapping_policy=node");
/* set mapping policy to bynode - error if something else already set */
if ((ORTE_MAPPING_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) &&
ORTE_GET_MAPPING_POLICY(orte_rmaps_base.mapping) != ORTE_MAPPING_BYNODE) {
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "mapping",
"bynode", orte_rmaps_base_print_mapping(orte_rmaps_base.mapping));
return ORTE_ERR_SILENT;
}
ORTE_SET_MAPPING_POLICY(orte_rmaps_base.mapping, ORTE_MAPPING_BYNODE);
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_GIVEN);
/* set ranking policy to bynode - error if something else already set */
if ((ORTE_RANKING_GIVEN & ORTE_GET_RANKING_DIRECTIVE(orte_rmaps_base.ranking)) &&
ORTE_GET_RANKING_POLICY(orte_rmaps_base.ranking) != ORTE_RANK_BY_NODE) {
/* error - cannot redefine the default ranking policy */
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "ranking",
"bynode", orte_rmaps_base_print_ranking(orte_rmaps_base.ranking));
return ORTE_ERR_SILENT;
}
ORTE_SET_RANKING_POLICY(orte_rmaps_base.ranking, ORTE_RANK_BY_NODE);
ORTE_SET_RANKING_DIRECTIVE(orte_rmaps_base.ranking, ORTE_RANKING_GIVEN);
}
if (1 < orte_rmaps_base.cpus_per_rank) {
/* if we were asked for multiple cpus/proc, then we have to
* bind to those cpus - any other binding policy is an
* error
*/
if (OPAL_BINDING_POLICY_IS_SET(opal_hwloc_binding_policy)) {
if (opal_hwloc_use_hwthreads_as_cpus) {
if (OPAL_BIND_TO_HWTHREAD != OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy) &&
OPAL_BIND_TO_NONE != OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
orte_show_help("help-orte-rmaps-base.txt", "mismatch-binding", true,
orte_rmaps_base.cpus_per_rank, "use-hwthreads-as-cpus",
opal_hwloc_base_print_binding(opal_hwloc_binding_policy),
"bind-to hwthread");
return ORTE_ERR_SILENT;
}
} else if (OPAL_BIND_TO_CORE != OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy) &&
OPAL_BIND_TO_NONE != OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
orte_show_help("help-orte-rmaps-base.txt", "mismatch-binding", true,
orte_rmaps_base.cpus_per_rank, "cores as cpus",
opal_hwloc_base_print_binding(opal_hwloc_binding_policy),
"bind-to core");
return ORTE_ERR_SILENT;
}
} else {
if (opal_hwloc_use_hwthreads_as_cpus) {
OPAL_SET_BINDING_POLICY(opal_hwloc_binding_policy, OPAL_BIND_TO_HWTHREAD);
} else {
OPAL_SET_BINDING_POLICY(opal_hwloc_binding_policy, OPAL_BIND_TO_CORE);
}
}
/* we also need to ensure we are mapping to a high-enough level to have
* multiple cpus beneath it - by default, we'll go to the NUMA level */
if (ORTE_MAPPING_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) {
if (ORTE_GET_MAPPING_POLICY(orte_rmaps_base.mapping) == ORTE_MAPPING_BYHWTHREAD ||
(ORTE_GET_MAPPING_POLICY(orte_rmaps_base.mapping) == ORTE_MAPPING_BYCORE &&
!opal_hwloc_use_hwthreads_as_cpus)) {
orte_show_help("help-orte-rmaps-base.txt", "mapping-too-low-init", true);
return ORTE_ERR_SILENT;
}
} else {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s rmaps:base pe/rank set - setting mapping to BYNUMA",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
ORTE_SET_MAPPING_POLICY(orte_rmaps_base.mapping, ORTE_MAPPING_BYNUMA);
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_GIVEN);
}
}
if (orte_rmaps_base_pernode) {
/* there is no way to resolve this conflict, so if something else was
* given, we have no choice but to error out
*/
if (ORTE_MAPPING_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "mapping",
"bynode", orte_rmaps_base_print_mapping(orte_rmaps_base.mapping));
return ORTE_ERR_SILENT;
}
/* ensure we set the mapping policy to ppr */
ORTE_SET_MAPPING_POLICY(orte_rmaps_base.mapping, ORTE_MAPPING_PPR);
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_GIVEN);
/* define the ppr */
orte_rmaps_base.ppr = strdup("1:node");
}
if (0 < orte_rmaps_base_n_pernode) {
/* there is no way to resolve this conflict, so if something else was
* given, we have no choice but to error out
*/
if (ORTE_MAPPING_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "mapping",
"bynode", orte_rmaps_base_print_mapping(orte_rmaps_base.mapping));
return ORTE_ERR_SILENT;
}
/* ensure we set the mapping policy to ppr */
ORTE_SET_MAPPING_POLICY(orte_rmaps_base.mapping, ORTE_MAPPING_PPR);
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_GIVEN);
/* define the ppr */
asprintf(&orte_rmaps_base.ppr, "%d:node", orte_rmaps_base_n_pernode);
}
if (0 < orte_rmaps_base_n_persocket) {
/* there is no way to resolve this conflict, so if something else was
* given, we have no choice but to error out
*/
if (ORTE_MAPPING_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "mapping",
"bynode", orte_rmaps_base_print_mapping(orte_rmaps_base.mapping));
return ORTE_ERR_SILENT;
}
/* ensure we set the mapping policy to ppr */
ORTE_SET_MAPPING_POLICY(orte_rmaps_base.mapping, ORTE_MAPPING_PPR);
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_GIVEN);
/* define the ppr */
asprintf(&orte_rmaps_base.ppr, "%d:socket", orte_rmaps_base_n_persocket);
}
/* Should we schedule on the local node or not? */
if (rmaps_base_no_schedule_local) {
orte_rmaps_base.mapping |= ORTE_MAPPING_NO_USE_LOCAL;
}
/* Should we oversubscribe or not? */
if (rmaps_base_no_oversubscribe) {
if ((ORTE_MAPPING_SUBSCRIBE_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) &&
!(ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping))) {
/* error - cannot redefine the default mapping policy */
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "mapping",
"no-oversubscribe", orte_rmaps_base_print_mapping(orte_rmaps_base.mapping));
return ORTE_ERR_SILENT;
}
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_NO_OVERSUBSCRIBE);
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_SUBSCRIBE_GIVEN);
}
/** force oversubscription permission */
if (rmaps_base_oversubscribe) {
if ((ORTE_MAPPING_SUBSCRIBE_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) &&
(ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping))) {
/* error - cannot redefine the default mapping policy */
orte_show_help("help-orte-rmaps-base.txt", "redefining-policy", true, "mapping",
"oversubscribe", orte_rmaps_base_print_mapping(orte_rmaps_base.mapping));
return ORTE_ERR_SILENT;
}
ORTE_UNSET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_NO_OVERSUBSCRIBE);
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_SUBSCRIBE_GIVEN);
/* also set the overload allowed flag */
opal_hwloc_binding_policy |= OPAL_BIND_ALLOW_OVERLOAD;
}
/* should we display a detailed (developer-quality) version of the map after determining it? */
if (rmaps_base_display_devel_map) {
orte_rmaps_base.display_map = true;
orte_devel_level_output = true;
}
/* should we display a diffable report of proc locations after determining it? */
if (rmaps_base_display_diffable_map) {
orte_rmaps_base.display_map = true;
orte_display_diffable_output = true;
}
/* Open up all available components */
rc = mca_base_framework_components_open(&orte_rmaps_base_framework, flags);
/* check to see if any component indicated a problem */
if (ORTE_MAPPING_CONFLICTED & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) {
/* the component would have already reported the error, so
* tell the rest of the chain to shut up
*/
return ORTE_ERR_SILENT;
}
/* All done */
return rc;
}
int orte_rmaps_rr_byslot(orte_job_t *jdata,
orte_app_context_t *app,
opal_list_t *node_list,
orte_std_cntr_t num_slots,
orte_vpid_t num_procs)
{
int rc, i, nprocs_mapped;
orte_node_t *node;
orte_proc_t *proc;
int num_procs_to_assign, extra_procs_to_assign=0, nxtra_nodes=0;
#if OPAL_HAVE_HWLOC
hwloc_obj_t obj=NULL;
#endif
float balance;
bool add_one=false;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr: mapping by slot for job %s slots %d num_procs %lu",
ORTE_JOBID_PRINT(jdata->jobid), (int)num_slots, (unsigned long)num_procs);
/* check to see if we can map all the procs */
if (num_slots < ((int)app->num_procs * orte_rmaps_base.cpus_per_rank)) {
if (ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, app->num_procs, app->app);
return ORTE_ERR_SILENT;
}
}
/* first pass: map the number of procs to each node until we
* map all specified procs or use all allocated slots
*/
nprocs_mapped = 0;
OPAL_LIST_FOREACH(node, node_list, orte_node_t) {
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot working node %s",
node->name);
#if OPAL_HAVE_HWLOC
/* get the root object as we are not assigning
* locale here except at the node level
*/
if (NULL != node->topology) {
obj = hwloc_get_root_obj(node->topology);
}
#endif
if (node->slots <= node->slots_inuse) {
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot node %s is full - skipping",
node->name);
continue;
}
/* assign a number of procs equal to the number of available
* slots divided by the number of cpus/rank the user
* requested
*/
num_procs_to_assign = (node->slots - node->slots_inuse) / orte_rmaps_base.cpus_per_rank;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:slot assigning %d procs to node %s",
(int)num_procs_to_assign, node->name);
for (i=0; i < num_procs_to_assign && nprocs_mapped < app->num_procs; i++) {
/* add this node to the map - do it only once */
if (!node->mapped) {
if (ORTE_SUCCESS > (rc = opal_pointer_array_add(jdata->map->nodes, (void*)node))) {
ORTE_ERROR_LOG(rc);
return rc;
}
node->mapped = true;
OBJ_RETAIN(node); /* maintain accounting on object */
++(jdata->map->num_nodes);
}
if (NULL == (proc = orte_rmaps_base_setup_proc(jdata, node, app->idx))) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
nprocs_mapped++;
#if OPAL_HAVE_HWLOC
proc->locale = obj;
#endif
}
}
int orte_rmaps_base_set_ranking_policy(orte_ranking_policy_t *policy,
orte_mapping_policy_t mapping,
char *spec)
{
orte_mapping_policy_t map;
orte_ranking_policy_t tmp;
char **ck;
size_t len;
/* set default */
tmp = 0;
if (NULL == spec) {
/* check for map-by object directives - we set the
* ranking to match if one was given
*/
if (ORTE_MAPPING_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(mapping)) {
map = ORTE_GET_MAPPING_POLICY(mapping);
switch (map) {
case ORTE_MAPPING_BYSLOT:
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_SLOT);
break;
case ORTE_MAPPING_BYNODE:
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_NODE);
break;
case ORTE_MAPPING_BYCORE:
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_CORE);
break;
case ORTE_MAPPING_BYL1CACHE:
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_L1CACHE);
break;
case ORTE_MAPPING_BYL2CACHE:
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_L2CACHE);
break;
case ORTE_MAPPING_BYL3CACHE:
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_L3CACHE);
break;
case ORTE_MAPPING_BYSOCKET:
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_SOCKET);
break;
case ORTE_MAPPING_BYNUMA:
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_NUMA);
break;
case ORTE_MAPPING_BYBOARD:
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_BOARD);
break;
case ORTE_MAPPING_BYHWTHREAD:
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_HWTHREAD);
break;
default:
/* anything not tied to a specific hw obj can rank by slot */
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_SLOT);
break;
}
} else {
/* if no map-by was given, default to by-slot */
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_SLOT);
}
} else {
ck = opal_argv_split(spec, ':');
if (2 < opal_argv_count(ck)) {
/* incorrect format */
orte_show_help("help-orte-rmaps-base.txt", "unrecognized-policy", true, "ranking", policy);
opal_argv_free(ck);
return ORTE_ERR_SILENT;
}
if (2 == opal_argv_count(ck)) {
if (0 == strncasecmp(ck[1], "span", strlen(ck[1]))) {
ORTE_SET_RANKING_DIRECTIVE(tmp, ORTE_RANKING_SPAN);
} else if (0 == strncasecmp(ck[1], "fill", strlen(ck[1]))) {
ORTE_SET_RANKING_DIRECTIVE(tmp, ORTE_RANKING_FILL);
} else {
/* unrecognized modifier */
orte_show_help("help-orte-rmaps-base.txt", "unrecognized-modifier", true, ck[1]);
opal_argv_free(ck);
return ORTE_ERR_SILENT;
}
}
len = strlen(ck[0]);
if (0 == strncasecmp(ck[0], "slot", len)) {
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_SLOT);
} else if (0 == strncasecmp(ck[0], "node", len)) {
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_NODE);
} else if (0 == strncasecmp(ck[0], "hwthread", len)) {
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_HWTHREAD);
} else if (0 == strncasecmp(ck[0], "core", len)) {
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_CORE);
} else if (0 == strncasecmp(ck[0], "l1cache", len)) {
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_L1CACHE);
} else if (0 == strncasecmp(ck[0], "l2cache", len)) {
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_L2CACHE);
} else if (0 == strncasecmp(ck[0], "l3cache", len)) {
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_L3CACHE);
} else if (0 == strncasecmp(ck[0], "socket", len)) {
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_SOCKET);
} else if (0 == strncasecmp(ck[0], "numa", len)) {
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_NUMA);
} else if (0 == strncasecmp(ck[0], "board", len)) {
ORTE_SET_RANKING_POLICY(tmp, ORTE_RANK_BY_BOARD);
} else {
orte_show_help("help-orte-rmaps-base.txt", "unrecognized-policy", true, "ranking", rmaps_base_ranking_policy);
opal_argv_free(ck);
return ORTE_ERR_SILENT;
}
opal_argv_free(ck);
ORTE_SET_RANKING_DIRECTIVE(tmp, ORTE_RANKING_GIVEN);
}
*policy = tmp;
return ORTE_SUCCESS;
}
static void recv_data(int fd, short args, void *cbdata)
{
bool found;
int i, rc;
orte_node_t *nd, *nd2;
opal_list_t nds, ndtmp;
opal_list_item_t *item, *itm;
char recv_msg[8192];
int nbytes, idx, sjob;
char **alloc, *nodelist, *tpn;
local_jobtracker_t *ptr, *jtrk;
local_apptracker_t *aptrk;
orte_app_context_t *app;
orte_jobid_t jobid;
orte_job_t *jdata;
char **dash_host = NULL;
opal_output_verbose(2, orte_ras_base_framework.framework_output,
"%s ras:slurm: dynamic allocation - data recvd",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
/* read the data from the socket and put it in the
* nodes field of op
*/
memset(recv_msg, 0, sizeof(recv_msg));
nbytes = read(fd, recv_msg, sizeof(recv_msg) - 1);
opal_output_verbose(2, orte_ras_base_framework.framework_output,
"%s ras:slurm: dynamic allocation msg: %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), recv_msg);
/* check if we got something */
if (0 == nbytes || 0 == strlen(recv_msg) || strstr(recv_msg, "failure") != NULL) {
/* show an error here - basically, a "nothing was available"
* message
*/
orte_show_help("help-ras-slurm.txt", "slurm-dyn-alloc-failed", true,
(0 == strlen(recv_msg)) ? "NO MSG" : recv_msg);
ORTE_ACTIVATE_JOB_STATE(NULL, ORTE_JOB_STATE_ALLOC_FAILED);
return;
}
/* break the message into its component parts, separated by colons */
alloc = opal_argv_split(recv_msg, ':');
/* the first section contains the ORTE jobid for this allocation */
tpn = strchr(alloc[0], '=');
orte_util_convert_string_to_jobid(&jobid, tpn+1);
/* get the corresponding job object */
jdata = orte_get_job_data_object(jobid);
jtrk = NULL;
/* find the associated tracking object */
for (item = opal_list_get_first(&jobs);
item != opal_list_get_end(&jobs);
item = opal_list_get_next(item)) {
ptr = (local_jobtracker_t*)item;
if (ptr->jobid == jobid) {
jtrk = ptr;
break;
}
}
if (NULL == jtrk) {
orte_show_help("help-ras-slurm.txt", "slurm-dyn-alloc-failed", true, "NO JOB TRACKER");
ORTE_ACTIVATE_JOB_STATE(NULL, ORTE_JOB_STATE_ALLOC_FAILED);
opal_argv_free(alloc);
return;
}
/* stop the timeout event */
opal_event_del(&jtrk->timeout_ev);
/* cycle across all the remaining parts - each is the allocation for
* an app in this job
*/
OBJ_CONSTRUCT(&nds, opal_list_t);
OBJ_CONSTRUCT(&ndtmp, opal_list_t);
idx = -1;
sjob = -1;
nodelist = NULL;
tpn = NULL;
for (i=1; NULL != alloc[i]; i++) {
if (ORTE_SUCCESS != parse_alloc_msg(alloc[i], &idx, &sjob, &nodelist, &tpn)) {
orte_show_help("help-ras-slurm.txt", "slurm-dyn-alloc-failed", true, jtrk->cmd);
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_ALLOC_FAILED);
opal_argv_free(alloc);
if (NULL != nodelist) {
free(nodelist);
}
if (NULL != tpn) {
free(tpn);
}
return;
}
if (idx < 0) {
orte_show_help("help-ras-slurm.txt", "slurm-dyn-alloc-failed", true, jtrk->cmd);
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_ALLOC_FAILED);
opal_argv_free(alloc);
free(nodelist);
free(tpn);
return;
}
if (NULL == (app = (orte_app_context_t*)opal_pointer_array_get_item(jdata->apps, idx))) {
orte_show_help("help-ras-slurm.txt", "slurm-dyn-alloc-failed", true, jtrk->cmd);
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_ALLOC_FAILED);
opal_argv_free(alloc);
free(nodelist);
free(tpn);
return;
}
/* release the current dash_host as that contained the *desired* allocation */
orte_remove_attribute(&app->attributes, ORTE_APP_DASH_HOST);
/* track the Slurm jobid */
if (NULL == (aptrk = (local_apptracker_t*)opal_pointer_array_get_item(&jtrk->apps, idx))) {
aptrk = OBJ_NEW(local_apptracker_t);
opal_pointer_array_set_item(&jtrk->apps, idx, aptrk);
}
aptrk->sjob = sjob;
/* since the nodelist/tpn may contain regular expressions, parse them */
if (ORTE_SUCCESS != (rc = orte_ras_slurm_discover(nodelist, tpn, &ndtmp))) {
ORTE_ERROR_LOG(rc);
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_ALLOC_FAILED);
opal_argv_free(alloc);
free(nodelist);
free(tpn);
return;
}
/* transfer the discovered nodes to our node list, and construct
* the new dash_host entry to match what was allocated
*/
while (NULL != (item = opal_list_remove_first(&ndtmp))) {
nd = (orte_node_t*)item;
opal_argv_append_nosize(&dash_host, nd->name);
/* check for duplicates */
found = false;
for (itm = opal_list_get_first(&nds);
itm != opal_list_get_end(&nds);
itm = opal_list_get_next(itm)) {
nd2 = (orte_node_t*)itm;
if (0 == strcmp(nd->name, nd2->name)) {
found = true;
nd2->slots += nd->slots;
OBJ_RELEASE(item);
break;
}
}
if (!found) {
/* append the new node to our list */
opal_list_append(&nds, item);
}
}
/* cleanup */
free(nodelist);
free(tpn);
}
/* cleanup */
opal_argv_free(alloc);
OBJ_DESTRUCT(&ndtmp);
if (NULL != dash_host) {
tpn = opal_argv_join(dash_host, ',');
orte_set_attribute(&app->attributes, ORTE_APP_DASH_HOST, ORTE_ATTR_LOCAL, (void*)tpn, OPAL_STRING);
opal_argv_free(dash_host);
free(tpn);
}
if (opal_list_is_empty(&nds)) {
/* if we get here, then we were able to contact slurm,
* which means we are in an actively managed cluster.
* However, slurm indicated that nothing is currently
* available that meets our requirements. This is a fatal
* situation - we do NOT have the option of running on
* user-specified hosts as the cluster is managed.
*/
OBJ_DESTRUCT(&nds);
orte_show_help("help-ras-base.txt", "ras-base:no-allocation", true);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
}
/* store the found nodes */
if (ORTE_SUCCESS != (rc = orte_ras_base_node_insert(&nds, jdata))) {
ORTE_ERROR_LOG(rc);
OBJ_DESTRUCT(&nds);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
return;
}
OBJ_DESTRUCT(&nds);
/* default to no-oversubscribe-allowed for managed systems */
if (!(ORTE_MAPPING_SUBSCRIBE_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping))) {
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_NO_OVERSUBSCRIBE);
}
/* flag that the allocation is managed */
orte_managed_allocation = true;
/* move the job along */
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_ALLOCATION_COMPLETE);
/* all done */
return;
}
/*
* Sequentially map the ranks according to the placement in the
* specified hostfile
*/
static int orte_rmaps_seq_map(orte_job_t *jdata)
{
orte_job_map_t *map;
orte_app_context_t *app;
int i, n;
orte_std_cntr_t j;
opal_list_item_t *item;
orte_node_t *node, *nd;
seq_node_t *sq, *save=NULL, *seq;;
orte_vpid_t vpid;
orte_std_cntr_t num_nodes;
int rc;
opal_list_t default_seq_list;
opal_list_t node_list, *seq_list, sq_list;
orte_proc_t *proc;
mca_base_component_t *c = &mca_rmaps_seq_component.base_version;
char *hosts = NULL, *sep, *eptr;
FILE *fp;
opal_hwloc_resource_type_t rtype;
OPAL_OUTPUT_VERBOSE((1, orte_rmaps_base_framework.framework_output,
"%s rmaps:seq called on job %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_JOBID_PRINT(jdata->jobid)));
/* this mapper can only handle initial launch
* when seq mapping is desired - allow
* restarting of failed apps
*/
if (ORTE_FLAG_TEST(jdata, ORTE_JOB_FLAG_RESTART)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: job %s is being restarted - seq cannot map",
ORTE_JOBID_PRINT(jdata->jobid));
return ORTE_ERR_TAKE_NEXT_OPTION;
}
if (NULL != jdata->map->req_mapper) {
if (0 != strcasecmp(jdata->map->req_mapper, c->mca_component_name)) {
/* a mapper has been specified, and it isn't me */
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: job %s not using sequential mapper",
ORTE_JOBID_PRINT(jdata->jobid));
return ORTE_ERR_TAKE_NEXT_OPTION;
}
/* we need to process it */
goto process;
}
if (ORTE_MAPPING_SEQ != ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
/* I don't know how to do these - defer */
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: job %s not using seq mapper",
ORTE_JOBID_PRINT(jdata->jobid));
return ORTE_ERR_TAKE_NEXT_OPTION;
}
process:
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: mapping job %s",
ORTE_JOBID_PRINT(jdata->jobid));
/* flag that I did the mapping */
if (NULL != jdata->map->last_mapper) {
free(jdata->map->last_mapper);
}
jdata->map->last_mapper = strdup(c->mca_component_name);
/* convenience def */
map = jdata->map;
/* if there is a default hostfile, go and get its ordered list of nodes */
OBJ_CONSTRUCT(&default_seq_list, opal_list_t);
if (NULL != orte_default_hostfile) {
char *hstname = NULL;
/* open the file */
fp = fopen(orte_default_hostfile, "r");
if (NULL == fp) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
rc = ORTE_ERR_NOT_FOUND;
goto error;
}
while (NULL != (hstname = orte_getline(fp))) {
if (0 == strlen(hstname)) {
free(hstname);
/* blank line - ignore */
continue;
}
if( '#' == hstname[0] ) {
free(hstname);
/* Comment line - ignore */
continue;
}
sq = OBJ_NEW(seq_node_t);
if (NULL != (sep = strchr(hstname, ' '))) {
*sep = '\0';
sep++;
/* remove any trailing space */
eptr = sep + strlen(sep) - 1;
while (eptr > sep && isspace(*eptr)) {
eptr--;
}
*(eptr+1) = 0;
sq->cpuset = strdup(sep);
}
// Strip off the FQDN if present, ignore IP addresses
if( !orte_keep_fqdn_hostnames && !opal_net_isaddr(hstname) ) {
char *ptr;
if (NULL != (ptr = strchr(hstname, '.'))) {
*ptr = '\0';
}
}
sq->hostname = hstname;
opal_list_append(&default_seq_list, &sq->super);
}
fclose(fp);
}
/* start at the beginning... */
vpid = 0;
jdata->num_procs = 0;
if (0 < opal_list_get_size(&default_seq_list)) {
save = (seq_node_t*)opal_list_get_first(&default_seq_list);
}
/* default to LOGICAL processors */
if (orte_get_attribute(&jdata->attributes, ORTE_JOB_PHYSICAL_CPUIDS, NULL, OPAL_BOOL)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: using PHYSICAL processors");
rtype = OPAL_HWLOC_PHYSICAL;
} else {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: using LOGICAL processors");
rtype = OPAL_HWLOC_LOGICAL;
}
/* initialize all the nodes as not included in this job map */
for (j=0; j < orte_node_pool->size; j++) {
if (NULL != (node = (orte_node_t*)opal_pointer_array_get_item(orte_node_pool, j))) {
ORTE_FLAG_UNSET(node, ORTE_NODE_FLAG_MAPPED);
}
}
/* cycle through the app_contexts, mapping them sequentially */
for(i=0; i < jdata->apps->size; i++) {
if (NULL == (app = (orte_app_context_t*)opal_pointer_array_get_item(jdata->apps, i))) {
continue;
}
/* dash-host trumps hostfile */
if (orte_get_attribute(&app->attributes, ORTE_APP_DASH_HOST, (void**)&hosts, OPAL_STRING)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: using dash-host nodes on app %s", app->app);
OBJ_CONSTRUCT(&node_list, opal_list_t);
/* dash host entries cannot specify cpusets, so used the std function to retrieve the list */
if (ORTE_SUCCESS != (rc = orte_util_get_ordered_dash_host_list(&node_list, hosts))) {
ORTE_ERROR_LOG(rc);
free(hosts);
goto error;
}
free(hosts);
/* transfer the list to a seq_node_t list */
OBJ_CONSTRUCT(&sq_list, opal_list_t);
while (NULL != (nd = (orte_node_t*)opal_list_remove_first(&node_list))) {
sq = OBJ_NEW(seq_node_t);
sq->hostname = strdup(nd->name);
opal_list_append(&sq_list, &sq->super);
OBJ_RELEASE(nd);
}
OBJ_DESTRUCT(&node_list);
seq_list = &sq_list;
} else if (orte_get_attribute(&app->attributes, ORTE_APP_HOSTFILE, (void**)&hosts, OPAL_STRING)) {
char *hstname;
if (NULL == hosts) {
rc = ORTE_ERR_NOT_FOUND;
goto error;
}
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: using hostfile %s nodes on app %s", hosts, app->app);
OBJ_CONSTRUCT(&sq_list, opal_list_t);
/* open the file */
fp = fopen(hosts, "r");
if (NULL == fp) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
rc = ORTE_ERR_NOT_FOUND;
OBJ_DESTRUCT(&sq_list);
goto error;
}
while (NULL != (hstname = orte_getline(fp))) {
if (0 == strlen(hstname)) {
free(hstname);
/* blank line - ignore */
continue;
}
if( '#' == hstname[0] ) {
free(hstname);
/* Comment line - ignore */
continue;
}
sq = OBJ_NEW(seq_node_t);
if (NULL != (sep = strchr(hstname, ' '))) {
*sep = '\0';
sep++;
/* remove any trailing space */
eptr = sep + strlen(sep) - 1;
while (eptr > sep && isspace(*eptr)) {
eptr--;
}
*(eptr+1) = 0;
sq->cpuset = strdup(sep);
}
// Strip off the FQDN if present, ignore IP addresses
if( !orte_keep_fqdn_hostnames && !opal_net_isaddr(hstname) ) {
char *ptr;
if (NULL != (ptr = strchr(hstname, '.'))) {
(*ptr) = '\0';
}
}
sq->hostname = hstname;
opal_list_append(&sq_list, &sq->super);
}
fclose(fp);
free(hosts);
seq_list = &sq_list;
} else if (0 < opal_list_get_size(&default_seq_list)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: using default hostfile nodes on app %s", app->app);
seq_list = &default_seq_list;
} else {
/* can't do anything - no nodes available! */
orte_show_help("help-orte-rmaps-base.txt",
"orte-rmaps-base:no-available-resources",
true);
return ORTE_ERR_SILENT;
}
/* check for nolocal and remove the head node, if required */
if (map->mapping & ORTE_MAPPING_NO_USE_LOCAL) {
for (item = opal_list_get_first(seq_list);
item != opal_list_get_end(seq_list);
item = opal_list_get_next(item) ) {
seq = (seq_node_t*)item;
/* need to check ifislocal because the name in the
* hostfile may not have been FQDN, while name returned
* by gethostname may have been (or vice versa)
*/
if (orte_ifislocal(seq->hostname)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: removing head node %s", seq->hostname);
opal_list_remove_item(seq_list, item);
OBJ_RELEASE(item); /* "un-retain" it */
}
}
}
if (NULL == seq_list || 0 == (num_nodes = (orte_std_cntr_t)opal_list_get_size(seq_list))) {
orte_show_help("help-orte-rmaps-base.txt",
"orte-rmaps-base:no-available-resources",
true);
return ORTE_ERR_SILENT;
}
/* if num_procs wasn't specified, set it now */
if (0 == app->num_procs) {
app->num_procs = num_nodes;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: setting num procs to %s for app %s",
ORTE_VPID_PRINT(app->num_procs), app->app);
} else if (num_nodes < app->num_procs) {
orte_show_help("help-orte-rmaps-base.txt", "seq:not-enough-resources", true,
app->num_procs, num_nodes);
return ORTE_ERR_SILENT;
}
if (seq_list == &default_seq_list) {
sq = save;
} else {
sq = (seq_node_t*)opal_list_get_first(seq_list);
}
for (n=0; n < app->num_procs; n++) {
/* find this node on the global array - this is necessary so
* that our mapping gets saved on that array as the objects
* returned by the hostfile function are -not- on the array
*/
node = NULL;
for (j=0; j < orte_node_pool->size; j++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(orte_node_pool, j))) {
continue;
}
if (0 == strcmp(sq->hostname, node->name)) {
break;
}
}
if (NULL == node) {
/* wasn't found - that is an error */
orte_show_help("help-orte-rmaps-seq.txt",
"orte-rmaps-seq:resource-not-found",
true, sq->hostname);
rc = ORTE_ERR_SILENT;
goto error;
}
/* ensure the node is in the map */
if (!ORTE_FLAG_TEST(node, ORTE_NODE_FLAG_MAPPED)) {
OBJ_RETAIN(node);
opal_pointer_array_add(map->nodes, node);
jdata->map->num_nodes++;
ORTE_FLAG_SET(node, ORTE_NODE_FLAG_MAPPED);
}
proc = orte_rmaps_base_setup_proc(jdata, node, i);
if ((node->slots < (int)node->num_procs) ||
(0 < node->slots_max && node->slots_max < (int)node->num_procs)) {
if (ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, node->num_procs, app->app);
ORTE_UPDATE_EXIT_STATUS(ORTE_ERROR_DEFAULT_EXIT_CODE);
rc = ORTE_ERR_SILENT;
goto error;
}
/* flag the node as oversubscribed so that sched-yield gets
* properly set
*/
ORTE_FLAG_SET(node, ORTE_NODE_FLAG_OVERSUBSCRIBED);
ORTE_FLAG_SET(jdata, ORTE_JOB_FLAG_OVERSUBSCRIBED);
/* check for permission */
if (ORTE_FLAG_TEST(node, ORTE_NODE_FLAG_SLOTS_GIVEN)) {
/* if we weren't given a directive either way, then we will error out
* as the #slots were specifically given, either by the host RM or
* via hostfile/dash-host */
if (!(ORTE_MAPPING_SUBSCRIBE_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping))) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, app->num_procs, app->app);
ORTE_UPDATE_EXIT_STATUS(ORTE_ERROR_DEFAULT_EXIT_CODE);
return ORTE_ERR_SILENT;
} else if (ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping)) {
/* if we were explicitly told not to oversubscribe, then don't */
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:alloc-error",
true, app->num_procs, app->app);
ORTE_UPDATE_EXIT_STATUS(ORTE_ERROR_DEFAULT_EXIT_CODE);
return ORTE_ERR_SILENT;
}
}
}
/* assign the vpid */
proc->name.vpid = vpid++;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: assign proc %s to node %s for app %s",
ORTE_VPID_PRINT(proc->name.vpid), sq->hostname, app->app);
/* record the cpuset, if given */
if (NULL != sq->cpuset) {
hwloc_cpuset_t bitmap;
char *cpu_bitmap;
if (NULL == node->topology) {
/* not allowed - for sequential cpusets, we must have
* the topology info
*/
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-topology", true, node->name);
rc = ORTE_ERR_SILENT;
goto error;
}
/* if we are using hwthreads as cpus and binding to hwthreads, then
* we can just copy the cpuset across as it already specifies things
* at that level */
if (opal_hwloc_use_hwthreads_as_cpus &&
OPAL_BIND_TO_HWTHREAD == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
cpu_bitmap = strdup(sq->cpuset);
} else {
/* setup the bitmap */
bitmap = hwloc_bitmap_alloc();
/* parse the slot_list to find the socket and core */
if (ORTE_SUCCESS != (rc = opal_hwloc_base_slot_list_parse(sq->cpuset, node->topology, rtype, bitmap))) {
ORTE_ERROR_LOG(rc);
hwloc_bitmap_free(bitmap);
goto error;
}
/* note that we cannot set the proc locale to any specific object
* as the slot list may have assigned it to more than one - so
* leave that field NULL
*/
/* set the proc to the specified map */
hwloc_bitmap_list_asprintf(&cpu_bitmap, bitmap);
hwloc_bitmap_free(bitmap);
}
orte_set_attribute(&proc->attributes, ORTE_PROC_CPU_BITMAP, ORTE_ATTR_GLOBAL, cpu_bitmap, OPAL_STRING);
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:seq: binding proc %s to cpuset %s bitmap %s",
ORTE_VPID_PRINT(proc->name.vpid), sq->cpuset, cpu_bitmap);
/* we are going to bind to cpuset since the user is specifying the cpus */
OPAL_SET_BINDING_POLICY(jdata->map->binding, OPAL_BIND_TO_CPUSET);
/* note that the user specified the mapping */
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_BYUSER);
ORTE_SET_MAPPING_DIRECTIVE(jdata->map->mapping, ORTE_MAPPING_GIVEN);
/* cleanup */
free(cpu_bitmap);
} else {
hwloc_obj_t locale;
/* assign the locale - okay for the topo to be null as
* it just means it wasn't returned
*/
if (NULL != node->topology) {
locale = hwloc_get_root_obj(node->topology);
orte_set_attribute(&proc->attributes, ORTE_PROC_HWLOC_LOCALE,
ORTE_ATTR_LOCAL, locale, OPAL_PTR);
}
}
/* add to the jdata proc array */
if (ORTE_SUCCESS != (rc = opal_pointer_array_set_item(jdata->procs, proc->name.vpid, proc))) {
ORTE_ERROR_LOG(rc);
goto error;
}
/* move to next node */
sq = (seq_node_t*)opal_list_get_next(&sq->super);
}
/** track the total number of processes we mapped */
jdata->num_procs += app->num_procs;
/* cleanup the node list if it came from this app_context */
if (seq_list != &default_seq_list) {
OPAL_LIST_DESTRUCT(seq_list);
} else {
save = sq;
}
}
return ORTE_SUCCESS;
error:
OPAL_LIST_DESTRUCT(&default_seq_list);
return rc;
}
char* orte_rmaps_base_print_mapping(orte_mapping_policy_t mapping)
{
char *ret, *map, *mymap, *tmp;
orte_rmaps_print_buffers_t *ptr;
if (ORTE_MAPPING_CONFLICTED & ORTE_GET_MAPPING_DIRECTIVE(mapping)) {
return "CONFLICTED";
}
ptr = get_print_buffer();
if (NULL == ptr) {
ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
return orte_rmaps_print_null;
}
/* cycle around the ring */
if (ORTE_RMAPS_PRINT_NUM_BUFS == ptr->cntr) {
ptr->cntr = 0;
}
switch(ORTE_GET_MAPPING_POLICY(mapping)) {
case ORTE_MAPPING_BYNODE:
map = "BYNODE";
break;
case ORTE_MAPPING_BYBOARD:
map = "BYBOARD";
break;
case ORTE_MAPPING_BYNUMA:
map = "BYNUMA";
break;
case ORTE_MAPPING_BYSOCKET:
map = "BYSOCKET";
break;
case ORTE_MAPPING_BYL3CACHE:
map = "BYL3CACHE";
break;
case ORTE_MAPPING_BYL2CACHE:
map = "BYL2CACHE";
break;
case ORTE_MAPPING_BYL1CACHE:
map = "BYL1CACHE";
break;
case ORTE_MAPPING_BYCORE:
map = "BYCORE";
break;
case ORTE_MAPPING_BYHWTHREAD:
map = "BYHWTHREAD";
break;
case ORTE_MAPPING_BYSLOT:
map = "BYSLOT";
break;
case ORTE_MAPPING_SEQ:
map = "SEQUENTIAL";
break;
case ORTE_MAPPING_BYUSER:
map = "BYUSER";
break;
case ORTE_MAPPING_BYDIST:
map = "MINDIST";
break;
default:
if (ORTE_MAPPING_PPR & ORTE_GET_MAPPING_DIRECTIVE(mapping)) {
map = "PPR";
} else {
map = "UNKNOWN";
}
}
if (0 != strcmp(map, "PPR") && (ORTE_MAPPING_PPR & ORTE_GET_MAPPING_DIRECTIVE(mapping))) {
asprintf(&mymap, "%s[PPR]:", map);
} else {
asprintf(&mymap, "%s:", map);
}
if (ORTE_MAPPING_NO_USE_LOCAL & ORTE_GET_MAPPING_DIRECTIVE(mapping)) {
asprintf(&tmp, "%sNO_USE_LOCAL,", mymap);
free(mymap);
mymap = tmp;
}
if (ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(mapping)) {
asprintf(&tmp, "%sNOOVERSUBSCRIBE,", mymap);
free(mymap);
mymap = tmp;
} else if (ORTE_MAPPING_SUBSCRIBE_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(mapping)) {
asprintf(&tmp, "%sOVERSUBSCRIBE,", mymap);
free(mymap);
mymap = tmp;
}
if (ORTE_MAPPING_SPAN & ORTE_GET_MAPPING_DIRECTIVE(mapping)) {
asprintf(&tmp, "%sSPAN,", mymap);
free(mymap);
mymap = tmp;
}
/* remove the trailing mark */
mymap[strlen(mymap)-1] = '\0';
snprintf(ptr->buffers[ptr->cntr], ORTE_RMAPS_PRINT_MAX_SIZE, "%s", mymap);
free(mymap);
ret = ptr->buffers[ptr->cntr];
ptr->cntr++;
return ret;
}
/*
* Function for selecting one component from all those that are
* available.
*/
void orte_ras_base_allocate(int fd, short args, void *cbdata)
{
int rc;
orte_job_t *jdata;
opal_list_t nodes;
orte_node_t *node;
orte_std_cntr_t i;
orte_app_context_t *app;
orte_state_caddy_t *caddy = (orte_state_caddy_t*)cbdata;
OPAL_OUTPUT_VERBOSE((5, orte_ras_base_framework.framework_output,
"%s ras:base:allocate",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* convenience */
jdata = caddy->jdata;
/* if we already did this, don't do it again - the pool of
* global resources is set.
*/
if (orte_ras_base.allocation_read) {
OPAL_OUTPUT_VERBOSE((5, orte_ras_base_framework.framework_output,
"%s ras:base:allocate allocation already read",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
goto next_state;
}
orte_ras_base.allocation_read = true;
/* Otherwise, we have to create
* the initial set of resources that will delineate all
* further operations serviced by this HNP. This list will
* contain ALL nodes that can be used by any subsequent job.
*
* In other words, if a node isn't found in this step, then
* no job launched by this HNP will be able to utilize it.
*/
/* construct a list to hold the results */
OBJ_CONSTRUCT(&nodes, opal_list_t);
/* if a component was selected, then we know we are in a managed
* environment. - the active module will return a list of what it found
*/
if (NULL != orte_ras_base.active_module) {
/* read the allocation */
if (ORTE_SUCCESS != (rc = orte_ras_base.active_module->allocate(jdata, &nodes))) {
if (ORTE_ERR_ALLOCATION_PENDING == rc) {
/* an allocation request is underway, so just do nothing */
OBJ_DESTRUCT(&nodes);
OBJ_RELEASE(caddy);
return;
}
if (ORTE_ERR_SYSTEM_WILL_BOOTSTRAP == rc) {
/* this module indicates that nodes will be discovered
* on a bootstrap basis, so all we do here is add our
* own node to the list
*/
goto addlocal;
}
if (ORTE_ERR_TAKE_NEXT_OPTION == rc) {
/* we have an active module, but it is unable to
* allocate anything for this job - this indicates
* that it isn't a fatal error, but could be if
* an allocation is required
*/
if (orte_allocation_required) {
/* an allocation is required, so this is fatal */
OBJ_DESTRUCT(&nodes);
orte_show_help("help-ras-base.txt", "ras-base:no-allocation", true);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
} else {
/* an allocation is not required, so we can just
* run on the local node - go add it
*/
goto addlocal;
}
}
ORTE_ERROR_LOG(rc);
OBJ_DESTRUCT(&nodes);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
}
/* If something came back, save it and we are done */
if (!opal_list_is_empty(&nodes)) {
/* store the results in the global resource pool - this removes the
* list items
*/
if (ORTE_SUCCESS != (rc = orte_ras_base_node_insert(&nodes, jdata))) {
ORTE_ERROR_LOG(rc);
OBJ_DESTRUCT(&nodes);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
OBJ_DESTRUCT(&nodes);
/* default to no-oversubscribe-allowed for managed systems */
if (!(ORTE_MAPPING_SUBSCRIBE_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping))) {
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_NO_OVERSUBSCRIBE);
}
/* flag that the allocation is managed */
orte_managed_allocation = true;
goto DISPLAY;
} else if (orte_allocation_required) {
/* if nothing was found, and an allocation is
* required, then error out
*/
OBJ_DESTRUCT(&nodes);
orte_show_help("help-ras-base.txt", "ras-base:no-allocation", true);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
OPAL_OUTPUT_VERBOSE((5, orte_ras_base_framework.framework_output,
"%s ras:base:allocate nothing found in module - proceeding to hostfile",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* nothing was found, or no active module was alive. Our next
* option is to look for a hostfile and assign our global
* pool from there.
*
* Individual hostfile names, if given, are included
* in the app_contexts for this job. We therefore need to
* retrieve the app_contexts for the job, and then cycle
* through them to see if anything is there. The parser will
* add the nodes found in each hostfile to our list - i.e.,
* the resulting list contains the UNION of all nodes specified
* in hostfiles from across all app_contexts
*
* We then continue to add any hosts provided by dash-host and
* the default hostfile, if we have it. We will then filter out
* all the non-desired hosts (i.e., those not specified by
* -host and/or -hostfile) when we start the mapping process
*
* Note that any relative node syntax found in the hostfiles will
* generate an error in this scenario, so only non-relative syntax
* can be present
*/
if (NULL != orte_default_hostfile) {
OPAL_OUTPUT_VERBOSE((5, orte_ras_base_framework.framework_output,
"%s ras:base:allocate parsing default hostfile %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
orte_default_hostfile));
/* a default hostfile was provided - parse it */
if (ORTE_SUCCESS != (rc = orte_util_add_hostfile_nodes(&nodes,
orte_default_hostfile))) {
ORTE_ERROR_LOG(rc);
OBJ_DESTRUCT(&nodes);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
}
for (i=0; i < jdata->apps->size; i++) {
if (NULL == (app = (orte_app_context_t*)opal_pointer_array_get_item(jdata->apps, i))) {
continue;
}
if (NULL != app->hostfile) {
OPAL_OUTPUT_VERBOSE((5, orte_ras_base_framework.framework_output,
"%s ras:base:allocate adding hostfile %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
app->hostfile));
/* hostfile was specified - parse it and add it to the list */
if (ORTE_SUCCESS != (rc = orte_util_add_hostfile_nodes(&nodes,
app->hostfile))) {
ORTE_ERROR_LOG(rc);
OBJ_DESTRUCT(&nodes);
/* set an error event */
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
} else if (!orte_soft_locations && NULL != app->dash_host) {
/* if we are using soft locations, then any dash-host would
* just include desired nodes and not required. We don't want
* to pick them up here as this would mean the request was
* always satisfied - instead, we want to allow the request
* to fail later on and use whatever nodes are actually
* available
*/
OPAL_OUTPUT_VERBOSE((5, orte_ras_base_framework.framework_output,
"%s ras:base:allocate adding dash_hosts",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
if (ORTE_SUCCESS != (rc = orte_util_add_dash_host_nodes(&nodes,
app->dash_host))) {
ORTE_ERROR_LOG(rc);
OBJ_DESTRUCT(&nodes);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
}
}
/* if something was found in the hostfile(s), we use that as our global
* pool - set it and we are done
*/
if (!opal_list_is_empty(&nodes)) {
/* store the results in the global resource pool - this removes the
* list items
*/
if (ORTE_SUCCESS != (rc = orte_ras_base_node_insert(&nodes, jdata))) {
ORTE_ERROR_LOG(rc);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
/* cleanup */
OBJ_DESTRUCT(&nodes);
goto DISPLAY;
}
OPAL_OUTPUT_VERBOSE((5, orte_ras_base_framework.framework_output,
"%s ras:base:allocate nothing found in hostfiles - checking for rankfile",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* Our next option is to look for a rankfile - if one was provided, we
* will use its nodes to create a default allocation pool
*/
if (NULL != orte_rankfile) {
/* check the rankfile for node information */
if (ORTE_SUCCESS != (rc = orte_util_add_hostfile_nodes(&nodes,
orte_rankfile))) {
ORTE_ERROR_LOG(rc);
OBJ_DESTRUCT(&nodes);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return ;
}
}
/* if something was found in rankfile, we use that as our global
* pool - set it and we are done
*/
if (!opal_list_is_empty(&nodes)) {
/* store the results in the global resource pool - this removes the
* list items
*/
if (ORTE_SUCCESS != (rc = orte_ras_base_node_insert(&nodes, jdata))) {
ORTE_ERROR_LOG(rc);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
/* rankfile is considered equivalent to an RM allocation */
if (!(ORTE_MAPPING_SUBSCRIBE_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping))) {
ORTE_SET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping, ORTE_MAPPING_NO_OVERSUBSCRIBE);
}
/* cleanup */
OBJ_DESTRUCT(&nodes);
goto DISPLAY;
}
OPAL_OUTPUT_VERBOSE((5, orte_ras_base_framework.framework_output,
"%s ras:base:allocate nothing found in rankfile - inserting current node",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
addlocal:
/* if nothing was found by any of the above methods, then we have no
* earthly idea what to do - so just add the local host
*/
node = OBJ_NEW(orte_node_t);
if (NULL == node) {
ORTE_ERROR_LOG(ORTE_ERR_OUT_OF_RESOURCE);
OBJ_DESTRUCT(&nodes);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
/* use the same name we got in orte_process_info so we avoid confusion in
* the session directories
*/
node->name = strdup(orte_process_info.nodename);
node->state = ORTE_NODE_STATE_UP;
node->slots_inuse = 0;
node->slots_max = 0;
node->slots = 1;
opal_list_append(&nodes, &node->super);
/* store the results in the global resource pool - this removes the
* list items
*/
if (ORTE_SUCCESS != (rc = orte_ras_base_node_insert(&nodes, jdata))) {
ORTE_ERROR_LOG(rc);
OBJ_DESTRUCT(&nodes);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
return;
}
OBJ_DESTRUCT(&nodes);
DISPLAY:
/* shall we display the results? */
if (4 < opal_output_get_verbosity(orte_ras_base_framework.framework_output)) {
orte_ras_base_display_alloc();
}
next_state:
/* are we to report this event? */
if (orte_report_events) {
if (ORTE_SUCCESS != (rc = orte_util_comm_report_event(ORTE_COMM_EVENT_ALLOCATE))) {
ORTE_ERROR_LOG(rc);
ORTE_FORCED_TERMINATE(ORTE_ERROR_DEFAULT_EXIT_CODE);
OBJ_RELEASE(caddy);
}
}
/* set total slots alloc */
jdata->total_slots_alloc = orte_ras_base.total_slots_alloc;
/* set the job state to the next position */
ORTE_ACTIVATE_JOB_STATE(jdata, ORTE_JOB_STATE_ALLOCATION_COMPLETE);
/* cleanup */
OBJ_RELEASE(caddy);
}
/*********************************
* Parsing Functions
*********************************/
int rmaps_lama_process_alias_params(orte_job_t *jdata)
{
int exit_status = ORTE_SUCCESS;
/*
* Mapping options
* Note: L1, L2, L3 are not exposed in orterun to the user, so
* there is no need to specify them here.
*/
if( NULL == rmaps_lama_cmd_map ) {
/* orte_rmaps_base.mapping */
switch( ORTE_GET_MAPPING_POLICY(jdata->map->mapping) ) {
case ORTE_MAPPING_BYNODE:
/* rmaps_lama_cmd_map = strdup("nbNsL3L2L1ch"); */
rmaps_lama_cmd_map = strdup("nbsch");
break;
case ORTE_MAPPING_BYBOARD:
/* rmaps_lama_cmd_map = strdup("bnNsL3L2L1ch"); */
orte_show_help("help-orte-rmaps-lama.txt",
"invalid mapping option",
true,
"by board", "mapping by board not supported by LAMA");
exit_status = ORTE_ERR_NOT_SUPPORTED;
goto cleanup;
break;
case ORTE_MAPPING_BYNUMA:
/* rmaps_lama_cmd_map = strdup("NbnsL3L2L1ch"); */
rmaps_lama_cmd_map = strdup("Nbnsch");
break;
case ORTE_MAPPING_BYSOCKET:
/* rmaps_lama_cmd_map = strdup("sNbnL3L2L1ch"); */
rmaps_lama_cmd_map = strdup("sbnch");
break;
case ORTE_MAPPING_BYL3CACHE:
rmaps_lama_cmd_map = strdup("L3sNbnL2L1ch");
break;
case ORTE_MAPPING_BYL2CACHE:
rmaps_lama_cmd_map = strdup("L2sNbnL1ch");
break;
case ORTE_MAPPING_BYL1CACHE:
rmaps_lama_cmd_map = strdup("L1sNbnch");
break;
case ORTE_MAPPING_BYCORE:
case ORTE_MAPPING_BYSLOT:
/* rmaps_lama_cmd_map = strdup("cL1L2L3sNbnh"); */
rmaps_lama_cmd_map = strdup("csbnh");
break;
case ORTE_MAPPING_BYHWTHREAD:
/* rmaps_lama_cmd_map = strdup("hcL1L2L3sNbn"); */
rmaps_lama_cmd_map = strdup("hcsbn");
break;
case ORTE_MAPPING_RR:
orte_show_help("help-orte-rmaps-lama.txt",
"invalid mapping option",
true,
"round robin", "mapping by round robin not supported by LAMA");
exit_status = ORTE_ERR_NOT_SUPPORTED;
goto cleanup;
case ORTE_MAPPING_SEQ:
orte_show_help("help-orte-rmaps-lama.txt",
"invalid mapping option",
true,
"sequential", "mapping by sequential not supported by LAMA");
exit_status = ORTE_ERR_NOT_SUPPORTED;
goto cleanup;
case ORTE_MAPPING_BYUSER:
orte_show_help("help-orte-rmaps-lama.txt",
"invalid mapping option",
true,
"by user", "mapping by user not supported by LAMA");
exit_status = ORTE_ERR_NOT_SUPPORTED;
goto cleanup;
default:
/*
* Default is map-by core
*/
rmaps_lama_cmd_map = strdup("cL1L2L3sNbnh");
break;
}
}
/*
* Binding Options
*/
if( NULL == rmaps_lama_cmd_bind ) {
/*
* No binding specified, use default
*/
if( !OPAL_BINDING_POLICY_IS_SET(jdata->map->binding) ||
!OPAL_BINDING_REQUIRED(opal_hwloc_binding_policy) ||
OPAL_BIND_TO_NONE == OPAL_GET_BINDING_POLICY(jdata->map->binding) ) {
rmaps_lama_cmd_bind = NULL;
}
switch( OPAL_GET_BINDING_POLICY(jdata->map->binding) ) {
case OPAL_BIND_TO_BOARD:
/* rmaps_lama_cmd_bind = strdup("1b"); */
orte_show_help("help-orte-rmaps-lama.txt",
"invalid binding option",
true,
"by board", "binding to board not supported by LAMA");
exit_status = ORTE_ERR_NOT_SUPPORTED;
goto cleanup;
break;
case OPAL_BIND_TO_NUMA:
rmaps_lama_cmd_bind = strdup("1N");
break;
case OPAL_BIND_TO_SOCKET:
rmaps_lama_cmd_bind = strdup("1s");
break;
case OPAL_BIND_TO_L3CACHE:
rmaps_lama_cmd_bind = strdup("1L3");
break;
case OPAL_BIND_TO_L2CACHE:
rmaps_lama_cmd_bind = strdup("1L2");
break;
case OPAL_BIND_TO_L1CACHE:
rmaps_lama_cmd_bind = strdup("1L1");
break;
case OPAL_BIND_TO_CORE:
rmaps_lama_cmd_bind = strdup("1c");
break;
case OPAL_BIND_TO_HWTHREAD:
rmaps_lama_cmd_bind = strdup("1h");
break;
case OPAL_BIND_TO_CPUSET:
orte_show_help("help-orte-rmaps-lama.txt",
"invalid binding option",
true,
"by CPU set", "binding to CPU set not supported by LAMA");
exit_status = ORTE_ERR_NOT_SUPPORTED;
goto cleanup;
break;
default:
rmaps_lama_cmd_bind = NULL;
break;
}
}
/*
* Ordering (a.k.a. Ranking) Options
*/
if( NULL == rmaps_lama_cmd_ordering ) {
/* orte_rmaps_base.ranking */
switch( ORTE_GET_RANKING_POLICY(jdata->map->ranking) ) {
case ORTE_RANK_BY_SLOT:
rmaps_lama_cmd_ordering = strdup("s");
break;
case ORTE_RANK_BY_NODE:
case ORTE_RANK_BY_NUMA:
case ORTE_RANK_BY_SOCKET:
case ORTE_RANK_BY_L3CACHE:
case ORTE_RANK_BY_L2CACHE:
case ORTE_RANK_BY_L1CACHE:
case ORTE_RANK_BY_CORE:
case ORTE_RANK_BY_HWTHREAD:
rmaps_lama_cmd_ordering = strdup("n");
break;
case ORTE_RANK_BY_BOARD:
/* rmaps_lama_cmd_ordering = strdup("n"); */
orte_show_help("help-orte-rmaps-lama.txt",
"invalid ordering option",
true,
"by board", "ordering by board not supported by LAMA");
exit_status = ORTE_ERR_NOT_SUPPORTED;
goto cleanup;
break;
default:
rmaps_lama_cmd_ordering = strdup("n");
break;
}
}
/*
* MPPR
*/
if( NULL == rmaps_lama_cmd_mppr ) {
/*
* The ppr is given in the map
*/
if( NULL != jdata->map->ppr) {
rmaps_lama_cmd_mppr = rmaps_lama_covert_ppr(jdata->map->ppr);
}
}
/*
* Oversubscription
*/
if( ORTE_MAPPING_NO_OVERSUBSCRIBE & ORTE_GET_MAPPING_DIRECTIVE(jdata->map->mapping) ) {
rmaps_lama_can_oversubscribe = false;
}
else {
rmaps_lama_can_oversubscribe = true;
}
/*
* Display revised values
*/
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:lama: Revised Parameters -----");
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:lama: Map : %s",
rmaps_lama_cmd_map);
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:lama: Bind : %s",
rmaps_lama_cmd_bind);
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:lama: MPPR : %s",
rmaps_lama_cmd_mppr);
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps:lama: Order : %s",
rmaps_lama_cmd_ordering);
cleanup:
return exit_status;
}
