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;
}
int opal_hwloc_base_set_binding_policy(opal_binding_policy_t *policy, char *spec)
{
int i;
opal_binding_policy_t tmp;
char **tmpvals, **quals;
/* set default */
tmp = 0;
/* binding specification */
if (NULL == spec) {
/* default to bind-to core, and that no binding policy was specified */
OPAL_SET_BINDING_POLICY(tmp, OPAL_BIND_TO_CORE);
tmp &= ~OPAL_BIND_GIVEN;
} else if (0 == strncasecmp(spec, "none", strlen("none"))) {
OPAL_SET_BINDING_POLICY(tmp, OPAL_BIND_TO_NONE);
} else {
tmpvals = opal_argv_split(spec, ':');
if (1 < opal_argv_count(tmpvals) || ':' == spec[0]) {
if (':' == spec[0]) {
quals = opal_argv_split(&spec[1], ',');
} else {
quals = opal_argv_split(tmpvals[1], ',');
}
for (i=0; NULL != quals[i]; i++) {
if (0 == strncasecmp(quals[i], "if-supported", strlen(quals[i]))) {
tmp |= OPAL_BIND_IF_SUPPORTED;
} else if (0 == strncasecmp(quals[i], "overload-allowed", strlen(quals[i])) ||
0 == strncasecmp(quals[i], "oversubscribe-allowed", strlen(quals[i]))) {
tmp |= OPAL_BIND_ALLOW_OVERLOAD;
} else {
/* unknown option */
opal_output(0, "Unknown qualifier to binding policy: %s", spec);
return OPAL_ERR_BAD_PARAM;
}
}
opal_argv_free(quals);
}
if (NULL == tmpvals[0] || ':' == spec[0]) {
OPAL_SET_BINDING_POLICY(tmp, OPAL_BIND_TO_CORE);
tmp &= ~OPAL_BIND_GIVEN;
} else {
if (0 == strcasecmp(tmpvals[0], "hwthread")) {
OPAL_SET_BINDING_POLICY(tmp, OPAL_BIND_TO_HWTHREAD);
} else if (0 == strcasecmp(tmpvals[0], "core")) {
OPAL_SET_BINDING_POLICY(tmp, OPAL_BIND_TO_CORE);
} else if (0 == strcasecmp(tmpvals[0], "l1cache")) {
OPAL_SET_BINDING_POLICY(tmp, OPAL_BIND_TO_L1CACHE);
} else if (0 == strcasecmp(tmpvals[0], "l2cache")) {
OPAL_SET_BINDING_POLICY(tmp, OPAL_BIND_TO_L2CACHE);
} else if (0 == strcasecmp(tmpvals[0], "l3cache")) {
OPAL_SET_BINDING_POLICY(tmp, OPAL_BIND_TO_L3CACHE);
} else if (0 == strcasecmp(tmpvals[0], "socket")) {
OPAL_SET_BINDING_POLICY(tmp, OPAL_BIND_TO_SOCKET);
} else if (0 == strcasecmp(tmpvals[0], "numa")) {
OPAL_SET_BINDING_POLICY(tmp, OPAL_BIND_TO_NUMA);
} else if (0 == strcasecmp(tmpvals[0], "board")) {
OPAL_SET_BINDING_POLICY(tmp, OPAL_BIND_TO_BOARD);
} else {
opal_show_help("help-opal-hwloc-base.txt", "invalid binding_policy", true, "binding", spec);
opal_argv_free(tmpvals);
return OPAL_ERR_BAD_PARAM;
}
}
opal_argv_free(tmpvals);
}
*policy = tmp;
return OPAL_SUCCESS;
}
static int opal_hwloc_base_open(mca_base_open_flag_t flags)
{
if (opal_hwloc_base_inited) {
return OPAL_SUCCESS;
}
opal_hwloc_base_inited = true;
#if OPAL_HAVE_HWLOC
{
int rc;
opal_data_type_t tmp;
if (OPAL_SUCCESS != (rc = opal_hwloc_base_set_binding_policy(&opal_hwloc_binding_policy,
opal_hwloc_base_binding_policy))) {
return rc;
}
if (opal_hwloc_base_bind_to_core) {
opal_show_help("help-opal-hwloc-base.txt", "deprecated", true,
"--bind-to-core", "--bind-to core",
"hwloc_base_bind_to_core", "hwloc_base_binding_policy=core");
/* set binding policy to core - error if something else already set */
if (OPAL_BINDING_POLICY_IS_SET(opal_hwloc_binding_policy) &&
OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy) != OPAL_BIND_TO_CORE) {
/* error - cannot redefine the default ranking policy */
opal_show_help("help-opal-hwloc-base.txt", "redefining-policy", true,
"core", opal_hwloc_base_print_binding(opal_hwloc_binding_policy));
return OPAL_ERR_BAD_PARAM;
}
OPAL_SET_BINDING_POLICY(opal_hwloc_binding_policy, OPAL_BIND_TO_CORE);
}
if (opal_hwloc_base_bind_to_socket) {
opal_show_help("help-opal-hwloc-base.txt", "deprecated", true,
"--bind-to-socket", "--bind-to socket",
"hwloc_base_bind_to_socket", "hwloc_base_binding_policy=socket");
/* set binding policy to socket - error if something else already set */
if (OPAL_BINDING_POLICY_IS_SET(opal_hwloc_binding_policy) &&
OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy) != OPAL_BIND_TO_SOCKET) {
/* error - cannot redefine the default ranking policy */
opal_show_help("help-opal-hwloc-base.txt", "redefining-policy", true,
"socket", opal_hwloc_base_print_binding(opal_hwloc_binding_policy));
return OPAL_ERR_SILENT;
}
OPAL_SET_BINDING_POLICY(opal_hwloc_binding_policy, OPAL_BIND_TO_SOCKET);
}
/* did the user provide a slot list? */
if (NULL != opal_hwloc_base_slot_list) {
/* if we already were given a policy, then this is an error */
if (OPAL_BINDING_POLICY_IS_SET(opal_hwloc_binding_policy)) {
opal_show_help("help-opal-hwloc-base.txt", "redefining-policy", true,
"socket", opal_hwloc_base_print_binding(opal_hwloc_binding_policy));
return OPAL_ERR_SILENT;
}
OPAL_SET_BINDING_POLICY(opal_hwloc_binding_policy, OPAL_BIND_TO_CPUSET);
}
/* cpu allocation specification */
if (NULL != opal_hwloc_base_cpu_set) {
if (!OPAL_BINDING_POLICY_IS_SET(opal_hwloc_binding_policy)) {
/* it is okay if a binding policy was already given - just ensure that
* we do bind to the given cpus if provided, otherwise this would be
* ignored if someone didn't also specify a binding policy
*/
OPAL_SET_BINDING_POLICY(opal_hwloc_binding_policy, OPAL_BIND_TO_CPUSET);
}
}
/* if we are binding to hwthreads, then we must use hwthreads as cpus */
if (OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy) == OPAL_BIND_TO_HWTHREAD) {
opal_hwloc_use_hwthreads_as_cpus = true;
}
/* to support tools such as ompi_info, add the components
* to a list
*/
if (OPAL_SUCCESS !=
mca_base_framework_components_open(&opal_hwloc_base_framework, flags)) {
return OPAL_ERROR;
}
/* declare the hwloc data types */
tmp = OPAL_HWLOC_TOPO;
if (OPAL_SUCCESS != (rc = opal_dss.register_type(opal_hwloc_pack,
opal_hwloc_unpack,
(opal_dss_copy_fn_t)opal_hwloc_copy,
(opal_dss_compare_fn_t)opal_hwloc_compare,
(opal_dss_print_fn_t)opal_hwloc_print,
OPAL_DSS_STRUCTURED,
"OPAL_HWLOC_TOPO", &tmp))) {
return rc;
}
}
#endif
return OPAL_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;
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;
}
/**
* 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;
}
#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
}
}
static int allocate(orte_job_t *jdata, opal_list_t *nodes)
{
char **nodelist;
orte_node_t *node;
int i, num_nodes;
char *affinity_file;
struct stat buf;
char *ptr;
/* get the list of allocated nodes */
if ((num_nodes = lsb_getalloc(&nodelist)) < 0) {
orte_show_help("help-ras-lsf.txt", "nodelist-failed", true);
return ORTE_ERR_NOT_AVAILABLE;
}
node = NULL;
/* step through the list */
for (i = 0; i < num_nodes; i++) {
if( !orte_keep_fqdn_hostnames && !opal_net_isaddr(nodelist[i]) ) {
if (NULL != (ptr = strchr(nodelist[i], '.'))) {
*ptr = '\0';
}
}
/* is this a repeat of the current node? */
if (NULL != node && 0 == strcmp(nodelist[i], node->name)) {
/* it is a repeat - just bump the slot count */
++node->slots;
opal_output_verbose(10, orte_ras_base_framework.framework_output,
"ras/lsf: +++ Node (%s) [slots=%d]", node->name, node->slots);
continue;
}
/* not a repeat - create a node entry for it */
node = OBJ_NEW(orte_node_t);
node->name = strdup(nodelist[i]);
node->slots_inuse = 0;
node->slots_max = 0;
node->slots = 1;
node->state = ORTE_NODE_STATE_UP;
opal_list_append(nodes, &node->super);
opal_output_verbose(10, orte_ras_base_framework.framework_output,
"ras/lsf: New Node (%s) [slots=%d]", node->name, node->slots);
}
/* release the nodelist from lsf */
opal_argv_free(nodelist);
/* check for an affinity file */
if (NULL != (affinity_file = getenv("LSB_AFFINITY_HOSTFILE"))) {
/* check to see if the file is empty - if it is,
* then affinity wasn't actually set for this job */
if (0 != stat(affinity_file, &buf)) {
orte_show_help("help-ras-lsf.txt", "affinity-file-not-found", true, affinity_file);
return ORTE_ERR_SILENT;
}
if (0 == buf.st_size) {
/* no affinity, so just return */
return ORTE_SUCCESS;
}
/* the affinity file sequentially lists rank locations, with
* cpusets given as physical cpu-ids. Setup the job object
* so it knows to process this accordingly */
if (NULL == jdata->map) {
jdata->map = OBJ_NEW(orte_job_map_t);
}
ORTE_SET_MAPPING_POLICY(jdata->map->mapping, ORTE_MAPPING_SEQ);
jdata->map->req_mapper = strdup("seq"); // need sequential mapper
/* tell the sequential mapper that all cpusets are to be treated as "physical" */
orte_set_attribute(&jdata->attributes, ORTE_JOB_PHYSICAL_CPUIDS, true, NULL, OPAL_BOOL);
/* LSF provides its info as hwthreads, so set the hwthread-as-cpus flag */
opal_hwloc_use_hwthreads_as_cpus = true;
/* don't override something provided by the user, but default to bind-to hwthread */
if (!OPAL_BINDING_POLICY_IS_SET(opal_hwloc_binding_policy)) {
OPAL_SET_BINDING_POLICY(opal_hwloc_binding_policy, OPAL_BIND_TO_HWTHREAD);
}
/*
* Do not set the hostfile attribute on each app_context since that
* would confuse the sequential mapper when it tries to assign bindings
* when running an MPMD job.
* Instead just overwrite the orte_default_hostfile so it will be
* general for all of the app_contexts.
*/
if( NULL != orte_default_hostfile ) {
free(orte_default_hostfile);
orte_default_hostfile = NULL;
}
orte_default_hostfile = strdup(affinity_file);
opal_output_verbose(10, orte_ras_base_framework.framework_output,
"ras/lsf: Set default_hostfile to %s",orte_default_hostfile);
return ORTE_SUCCESS;
}
return ORTE_SUCCESS;
}
static int bind_in_place(orte_job_t *jdata,
hwloc_obj_type_t target,
unsigned cache_level)
{
/* traverse the hwloc topology tree on each node downwards
* until we find an unused object of type target - and then bind
* the process to that target
*/
int i, j;
orte_job_map_t *map;
orte_node_t *node;
orte_proc_t *proc;
hwloc_cpuset_t cpus;
unsigned int idx, ncpus;
struct hwloc_topology_support *support;
opal_hwloc_obj_data_t *data;
hwloc_obj_t locale, sib;
char *cpu_bitmap;
bool found;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bind in place for job %s with bindings %s",
ORTE_JOBID_PRINT(jdata->jobid),
opal_hwloc_base_print_binding(jdata->map->binding));
/* initialize */
map = jdata->map;
for (i=0; i < map->nodes->size; i++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(map->nodes, i))) {
continue;
}
if (!orte_do_not_launch) {
/* if we don't want to launch, then we are just testing the system,
* so ignore questions about support capabilities
*/
support = (struct hwloc_topology_support*)hwloc_topology_get_support(node->topology);
/* check if topology supports cpubind - have to be careful here
* as Linux doesn't currently support thread-level binding. This
* may change in the future, though, and it isn't clear how hwloc
* interprets the current behavior. So check both flags to be sure.
*/
if (!support->cpubind->set_thisproc_cpubind &&
!support->cpubind->set_thisthread_cpubind) {
if (!OPAL_BINDING_REQUIRED(map->binding) ||
!OPAL_BINDING_POLICY_IS_SET(map->binding)) {
/* we are not required to bind, so ignore this */
continue;
}
orte_show_help("help-orte-rmaps-base.txt", "rmaps:cpubind-not-supported", true, node->name);
return ORTE_ERR_SILENT;
}
/* check if topology supports membind - have to be careful here
* as hwloc treats this differently than I (at least) would have
* expected. Per hwloc, Linux memory binding is at the thread,
* and not process, level. Thus, hwloc sets the "thisproc" flag
* to "false" on all Linux systems, and uses the "thisthread" flag
* to indicate binding capability - don't warn if the user didn't
* specifically request binding
*/
if (!support->membind->set_thisproc_membind &&
!support->membind->set_thisthread_membind &&
OPAL_BINDING_POLICY_IS_SET(map->binding)) {
if (OPAL_HWLOC_BASE_MBFA_WARN == opal_hwloc_base_mbfa && !membind_warned) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:membind-not-supported", true, node->name);
membind_warned = true;
} else if (OPAL_HWLOC_BASE_MBFA_ERROR == opal_hwloc_base_mbfa) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:membind-not-supported-fatal", true, node->name);
return ORTE_ERR_SILENT;
}
}
}
/* some systems do not report cores, and so we can get a situation where our
* default binding policy will fail for no necessary reason. So if we are
* computing a binding due to our default policy, and no cores are found
* on this node, just silently skip it - we will not bind
*/
if (!OPAL_BINDING_POLICY_IS_SET(map->binding) &&
HWLOC_TYPE_DEPTH_UNKNOWN == hwloc_get_type_depth(node->topology, HWLOC_OBJ_CORE)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"Unable to bind-to core by default on node %s as no cores detected",
node->name);
continue;
}
/* we share topologies in order
* to save space, so we need to reset the usage info to reflect
* our own current state
*/
reset_usage(node, jdata->jobid);
/* cycle thru the procs */
for (j=0; j < node->procs->size; j++) {
if (NULL == (proc = (orte_proc_t*)opal_pointer_array_get_item(node->procs, j))) {
continue;
}
/* ignore procs from other jobs */
if (proc->name.jobid != jdata->jobid) {
continue;
}
/* bozo check */
if (!orte_get_attribute(&proc->attributes, ORTE_PROC_HWLOC_LOCALE, (void**)&locale, OPAL_PTR)) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-locale", true, ORTE_NAME_PRINT(&proc->name));
return ORTE_ERR_SILENT;
}
/* get the index of this location */
if (UINT_MAX == (idx = opal_hwloc_base_get_obj_idx(node->topology, locale, OPAL_HWLOC_AVAILABLE))) {
ORTE_ERROR_LOG(ORTE_ERR_BAD_PARAM);
return ORTE_ERR_SILENT;
}
data = (opal_hwloc_obj_data_t*)locale->userdata;
/* get the number of cpus under this location */
if (0 == (ncpus = opal_hwloc_base_get_npus(node->topology, locale))) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-available-cpus", true, node->name);
return ORTE_ERR_SILENT;
}
/* if we don't have enough cpus to support this additional proc, try
* shifting the location to a cousin that can support it - the important
* thing is that we maintain the same level in the topology */
if (ncpus < (data->num_bound+1)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s bind_in_place: searching right",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
sib = locale;
found = false;
while (NULL != (sib = sib->next_cousin)) {
data = (opal_hwloc_obj_data_t*)sib->userdata;
ncpus = opal_hwloc_base_get_npus(node->topology, sib);
if (data->num_bound < ncpus) {
found = true;
locale = sib;
break;
}
}
if (!found) {
/* try the other direction */
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s bind_in_place: searching left",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
sib = locale;
while (NULL != (sib = sib->prev_cousin)) {
data = (opal_hwloc_obj_data_t*)sib->userdata;
ncpus = opal_hwloc_base_get_npus(node->topology, sib);
if (data->num_bound < ncpus) {
found = true;
locale = sib;
break;
}
}
}
if (!found) {
/* no place to put this - see if overload is allowed */
if (!OPAL_BIND_OVERLOAD_ALLOWED(jdata->map->binding)) {
if (OPAL_BINDING_POLICY_IS_SET(jdata->map->binding)) {
/* if the user specified a binding policy, then we cannot meet
* it since overload isn't allowed, so error out - have the
* message indicate that setting overload allowed will remove
* this restriction */
orte_show_help("help-orte-rmaps-base.txt", "rmaps:binding-overload", true,
opal_hwloc_base_print_binding(map->binding), node->name,
data->num_bound, ncpus);
return ORTE_ERR_SILENT;
} else {
/* if we have the default binding policy, then just don't bind */
OPAL_SET_BINDING_POLICY(map->binding, OPAL_BIND_TO_NONE);
unbind_procs(jdata);
return ORTE_SUCCESS;
}
}
}
}
/* track the number bound */
data = (opal_hwloc_obj_data_t*)locale->userdata; // just in case it changed
data->num_bound++;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"BINDING PROC %s TO %s NUMBER %u",
ORTE_NAME_PRINT(&proc->name),
hwloc_obj_type_string(locale->type), idx);
/* bind the proc here */
cpus = opal_hwloc_base_get_available_cpus(node->topology, locale);
hwloc_bitmap_list_asprintf(&cpu_bitmap, cpus);
orte_set_attribute(&proc->attributes, ORTE_PROC_CPU_BITMAP, ORTE_ATTR_GLOBAL, cpu_bitmap, OPAL_STRING);
/* update the location, in case it changed */
orte_set_attribute(&proc->attributes, ORTE_PROC_HWLOC_BOUND, ORTE_ATTR_LOCAL, locale, OPAL_PTR);
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s BOUND PROC %s TO %s[%s:%u] on node %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&proc->name),
cpu_bitmap, hwloc_obj_type_string(locale->type),
idx, node->name);
if (NULL != cpu_bitmap) {
free(cpu_bitmap);
}
}
}
return ORTE_SUCCESS;
}
static int bind_downwards(orte_job_t *jdata,
orte_node_t *node,
hwloc_obj_type_t target,
unsigned cache_level)
{
int j;
orte_job_map_t *map;
orte_proc_t *proc;
hwloc_obj_t trg_obj, nxt_obj;
hwloc_cpuset_t cpus;
unsigned int ncpus;
opal_hwloc_obj_data_t *data;
int total_cpus;
hwloc_cpuset_t totalcpuset;
hwloc_obj_t locale;
char *cpu_bitmap;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bind downward for job %s with bindings %s",
ORTE_JOBID_PRINT(jdata->jobid),
opal_hwloc_base_print_binding(jdata->map->binding));
/* initialize */
map = jdata->map;
totalcpuset = hwloc_bitmap_alloc();
/* cycle thru the procs */
for (j=0; j < node->procs->size; j++) {
if (NULL == (proc = (orte_proc_t*)opal_pointer_array_get_item(node->procs, j))) {
continue;
}
/* ignore procs from other jobs */
if (proc->name.jobid != jdata->jobid) {
continue;
}
/* bozo check */
locale = NULL;
if (!orte_get_attribute(&proc->attributes, ORTE_PROC_HWLOC_LOCALE, (void**)&locale, OPAL_PTR)) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-locale", true, ORTE_NAME_PRINT(&proc->name));
hwloc_bitmap_free(totalcpuset);
return ORTE_ERR_SILENT;
}
/* we don't know if the target is a direct child of this locale,
* or if it is some depth below it, so we have to conduct a bit
* of a search. Let hwloc find the min usage one for us.
*/
trg_obj = opal_hwloc_base_find_min_bound_target_under_obj(node->topology, locale,
target, cache_level);
if (NULL == trg_obj) {
/* there aren't any such targets under this object */
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-available-cpus", true, node->name);
hwloc_bitmap_free(totalcpuset);
return ORTE_ERR_SILENT;
}
/* record the location */
orte_set_attribute(&proc->attributes, ORTE_PROC_HWLOC_BOUND, ORTE_ATTR_LOCAL, trg_obj, OPAL_PTR);
/* start with a clean slate */
hwloc_bitmap_zero(totalcpuset);
total_cpus = 0;
nxt_obj = trg_obj;
do {
if (NULL == nxt_obj) {
/* could not find enough cpus to meet request */
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-available-cpus", true, node->name);
hwloc_bitmap_free(totalcpuset);
return ORTE_ERR_SILENT;
}
trg_obj = nxt_obj;
/* get the number of cpus under this location */
ncpus = opal_hwloc_base_get_npus(node->topology, trg_obj);
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s GOT %d CPUS",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), ncpus);
/* track the number bound */
if (NULL == (data = (opal_hwloc_obj_data_t*)trg_obj->userdata)) {
data = OBJ_NEW(opal_hwloc_obj_data_t);
trg_obj->userdata = data;
}
data->num_bound++;
/* error out if adding a proc would cause overload and that wasn't allowed,
* and it wasn't a default binding policy (i.e., the user requested it)
*/
if (ncpus < data->num_bound &&
!OPAL_BIND_OVERLOAD_ALLOWED(jdata->map->binding)) {
if (OPAL_BINDING_POLICY_IS_SET(jdata->map->binding)) {
/* if the user specified a binding policy, then we cannot meet
* it since overload isn't allowed, so error out - have the
* message indicate that setting overload allowed will remove
* this restriction */
orte_show_help("help-orte-rmaps-base.txt", "rmaps:binding-overload", true,
opal_hwloc_base_print_binding(map->binding), node->name,
data->num_bound, ncpus);
hwloc_bitmap_free(totalcpuset);
return ORTE_ERR_SILENT;
} else {
/* if we have the default binding policy, then just don't bind */
OPAL_SET_BINDING_POLICY(map->binding, OPAL_BIND_TO_NONE);
unbind_procs(jdata);
hwloc_bitmap_zero(totalcpuset);
return ORTE_SUCCESS;
}
}
/* bind the proc here */
cpus = opal_hwloc_base_get_available_cpus(node->topology, trg_obj);
hwloc_bitmap_or(totalcpuset, totalcpuset, cpus);
/* track total #cpus */
total_cpus += ncpus;
/* move to the next location, in case we need it */
nxt_obj = trg_obj->next_cousin;
} while (total_cpus < orte_rmaps_base.cpus_per_rank);
hwloc_bitmap_list_asprintf(&cpu_bitmap, totalcpuset);
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s PROC %s BITMAP %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&proc->name), cpu_bitmap);
orte_set_attribute(&proc->attributes, ORTE_PROC_CPU_BITMAP, ORTE_ATTR_GLOBAL, cpu_bitmap, OPAL_STRING);
if (NULL != cpu_bitmap) {
free(cpu_bitmap);
}
if (4 < opal_output_get_verbosity(orte_rmaps_base_framework.framework_output)) {
char tmp1[1024], tmp2[1024];
if (OPAL_ERR_NOT_BOUND == opal_hwloc_base_cset2str(tmp1, sizeof(tmp1),
node->topology, totalcpuset)) {
opal_output(orte_rmaps_base_framework.framework_output,
"%s PROC %s ON %s IS NOT BOUND",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&proc->name), node->name);
} else {
opal_hwloc_base_cset2mapstr(tmp2, sizeof(tmp2), node->topology, totalcpuset);
opal_output(orte_rmaps_base_framework.framework_output,
"%s BOUND PROC %s[%s] TO %s: %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&proc->name), node->name,
tmp1, tmp2);
}
}
}
hwloc_bitmap_free(totalcpuset);
return ORTE_SUCCESS;
}
static int bind_upwards(orte_job_t *jdata,
orte_node_t *node,
hwloc_obj_type_t target,
unsigned cache_level)
{
/* traverse the hwloc topology tree on each node upwards
* until we find an object of type target - and then bind
* the process to that target
*/
int j;
orte_job_map_t *map;
orte_proc_t *proc;
hwloc_obj_t obj;
hwloc_cpuset_t cpus;
unsigned int idx, ncpus;
opal_hwloc_obj_data_t *data;
hwloc_obj_t locale;
char *cpu_bitmap;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bind upwards for job %s with bindings %s",
ORTE_JOBID_PRINT(jdata->jobid),
opal_hwloc_base_print_binding(jdata->map->binding));
/* initialize */
map = jdata->map;
/* cycle thru the procs */
for (j=0; j < node->procs->size; j++) {
if (NULL == (proc = (orte_proc_t*)opal_pointer_array_get_item(node->procs, j))) {
continue;
}
/* ignore procs from other jobs */
if (proc->name.jobid != jdata->jobid) {
continue;
}
/* bozo check */
if (!orte_get_attribute(&proc->attributes, ORTE_PROC_HWLOC_LOCALE, (void**)&locale, OPAL_PTR)) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-locale", true, ORTE_NAME_PRINT(&proc->name));
return ORTE_ERR_SILENT;
}
/* starting at the locale, move up thru the parents
* to find the target object type
*/
cpu_bitmap = NULL;
for (obj = locale->parent; NULL != obj; obj = obj->parent) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s bind:upward target %s type %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
hwloc_obj_type_string(target),
hwloc_obj_type_string(obj->type));
if (target == obj->type) {
if (HWLOC_OBJ_CACHE == target && cache_level != obj->attr->cache.depth) {
continue;
}
/* get its index */
if (UINT_MAX == (idx = opal_hwloc_base_get_obj_idx(node->topology, obj, OPAL_HWLOC_AVAILABLE))) {
ORTE_ERROR_LOG(ORTE_ERR_BAD_PARAM);
return ORTE_ERR_SILENT;
}
/* track the number bound */
data = (opal_hwloc_obj_data_t*)obj->userdata;
data->num_bound++;
/* get the number of cpus under this location */
if (0 == (ncpus = opal_hwloc_base_get_npus(node->topology, obj))) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-available-cpus", true, node->name);
return ORTE_ERR_SILENT;
}
/* error out if adding a proc would cause overload and that wasn't allowed,
* and it wasn't a default binding policy (i.e., the user requested it)
*/
if (ncpus < data->num_bound &&
!OPAL_BIND_OVERLOAD_ALLOWED(jdata->map->binding)) {
if (OPAL_BINDING_POLICY_IS_SET(jdata->map->binding)) {
/* if the user specified a binding policy, then we cannot meet
* it since overload isn't allowed, so error out - have the
* message indicate that setting overload allowed will remove
* this restriction */
orte_show_help("help-orte-rmaps-base.txt", "rmaps:binding-overload", true,
opal_hwloc_base_print_binding(map->binding), node->name,
data->num_bound, ncpus);
return ORTE_ERR_SILENT;
} else {
/* if we have the default binding policy, then just don't bind */
OPAL_SET_BINDING_POLICY(map->binding, OPAL_BIND_TO_NONE);
unbind_procs(jdata);
return ORTE_SUCCESS;
}
}
/* bind it here */
cpus = opal_hwloc_base_get_available_cpus(node->topology, obj);
hwloc_bitmap_list_asprintf(&cpu_bitmap, cpus);
orte_set_attribute(&proc->attributes, ORTE_PROC_CPU_BITMAP, ORTE_ATTR_GLOBAL, cpu_bitmap, OPAL_STRING);
/* record the location */
orte_set_attribute(&proc->attributes, ORTE_PROC_HWLOC_BOUND, ORTE_ATTR_LOCAL, obj, OPAL_PTR);
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s BOUND PROC %s TO %s[%s:%u] on node %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&proc->name),
cpu_bitmap,
hwloc_obj_type_string(target),
idx, node->name);
break;
}
}
if (NULL == cpu_bitmap && OPAL_BINDING_REQUIRED(jdata->map->binding)) {
/* didn't find anyone to bind to - this is an error
* unless the user specified if-supported
*/
orte_show_help("help-orte-rmaps-base.txt", "rmaps:binding-target-not-found", true,
opal_hwloc_base_print_binding(map->binding), node->name);
return ORTE_ERR_SILENT;
}
if (NULL != cpu_bitmap) {
free(cpu_bitmap);
}
}
return ORTE_SUCCESS;
}
