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;
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;
}
/* ignore procs that have already been bound - should
* never happen, but safer
*/
if (NULL != proc->cpu_bitmap) {
continue;
}
/* 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,
proc->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 */
proc->bind_location = trg_obj;
/* 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)) {
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 this is the default binding policy, then just don't
* bind this proc
*/
data->num_bound--; // maintain count
/* show the proc as not bound */
proc->bind_location = NULL;
hwloc_bitmap_zero(totalcpuset);
break;
}
}
/* 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(&proc->cpu_bitmap, totalcpuset);
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_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;
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;
}
/* ignore procs that have already been bound - should
* never happen, but safer
*/
if (NULL != proc->cpu_bitmap) {
continue;
}
/* get the index of this location */
if (UINT_MAX == (idx = opal_hwloc_base_get_obj_idx(node->topology, proc->locale, OPAL_HWLOC_AVAILABLE))) {
ORTE_ERROR_LOG(ORTE_ERR_BAD_PARAM);
return ORTE_ERR_SILENT;
}
/* track the number bound */
data = (opal_hwloc_obj_data_t*)proc->locale->userdata;
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(proc->locale->type), idx);
/* get the number of cpus under this location */
if (0 == (ncpus = opal_hwloc_base_get_npus(node->topology, proc->locale))) {
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) &&
OPAL_BINDING_POLICY_IS_SET(jdata->map->binding)) {
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;
}
/* bind the proc here */
cpus = opal_hwloc_base_get_available_cpus(node->topology, proc->locale);
hwloc_bitmap_list_asprintf(&proc->cpu_bitmap, cpus);
/* record the location */
proc->bind_location = proc->locale;
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),
proc->cpu_bitmap,
hwloc_obj_type_string(proc->locale->type),
idx, node->name);
}
}
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;
}
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;
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;
}
/* ignore procs that have already been bound - should
* never happen, but safer
*/
if (NULL != proc->cpu_bitmap) {
continue;
}
/* bozo check */
if (NULL == proc->locale) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"BIND UPWARDS: LOCALE FOR PROC %s IS NULL",
ORTE_NAME_PRINT(&proc->name));
return ORTE_ERR_SILENT;
}
/* starting at the locale, move up thru the parents
* to find the target object type
*/
for (obj = proc->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) &&
OPAL_BINDING_POLICY_IS_SET(jdata->map->binding)) {
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;
}
/* bind it here */
cpus = opal_hwloc_base_get_available_cpus(node->topology, obj);
hwloc_bitmap_list_asprintf(&proc->cpu_bitmap, cpus);
/* record the location */
proc->bind_location = obj;
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),
proc->cpu_bitmap,
hwloc_obj_type_string(target),
idx, node->name);
break;
}
}
if (NULL == proc->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;
}
}
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;
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 void set(orte_job_t *jobdat,
orte_proc_t *child,
char ***environ_copy,
int write_fd)
{
hwloc_cpuset_t cpuset;
hwloc_obj_t root;
opal_hwloc_topo_data_t *sum;
orte_app_context_t *context;
int rc=ORTE_ERROR;
char *msg, *param;
char *cpu_bitmap;
opal_output_verbose(2, orte_rtc_base_framework.framework_output,
"%s hwloc:set on child %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(NULL == child) ? "NULL" : ORTE_NAME_PRINT(&child->name));
if (NULL == jobdat || NULL == child) {
/* nothing for us to do */
opal_output_verbose(2, orte_rtc_base_framework.framework_output,
"%s hwloc:set jobdat %s child %s - nothing to do",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
(NULL == jobdat) ? "NULL" : ORTE_JOBID_PRINT(jobdat->jobid),
(NULL == child) ? "NULL" : ORTE_NAME_PRINT(&child->name));
return;
}
context = (orte_app_context_t*)opal_pointer_array_get_item(jobdat->apps, child->app_idx);
/* Set process affinity, if given */
cpu_bitmap = NULL;
if (!orte_get_attribute(&child->attributes, ORTE_PROC_CPU_BITMAP, (void**)&cpu_bitmap, OPAL_STRING) ||
NULL == cpu_bitmap || 0 == strlen(cpu_bitmap)) {
/* if the daemon is bound, then we need to "free" this proc */
if (NULL != orte_daemon_cores) {
root = hwloc_get_root_obj(opal_hwloc_topology);
if (NULL == root->userdata) {
orte_rtc_base_send_warn_show_help(write_fd,
"help-orte-odls-default.txt", "incorrectly bound",
orte_process_info.nodename, context->app,
__FILE__, __LINE__);
}
sum = (opal_hwloc_topo_data_t*)root->userdata;
/* bind this proc to all available processors */
rc = hwloc_set_cpubind(opal_hwloc_topology, sum->available, 0);
/* if we got an error and this wasn't a default binding policy, then report it */
if (rc < 0 && OPAL_BINDING_POLICY_IS_SET(jobdat->map->binding)) {
if (errno == ENOSYS) {
msg = "hwloc indicates cpu binding not supported";
} else if (errno == EXDEV) {
msg = "hwloc indicates cpu binding cannot be enforced";
} else {
char *tmp;
(void)hwloc_bitmap_list_asprintf(&tmp, sum->available);
asprintf(&msg, "hwloc_set_cpubind returned \"%s\" for bitmap \"%s\"",
opal_strerror(rc), tmp);
free(tmp);
}
if (OPAL_BINDING_REQUIRED(jobdat->map->binding)) {
/* If binding is required, send an error up the pipe (which exits
-- it doesn't return). */
orte_rtc_base_send_error_show_help(write_fd, 1, "help-orte-odls-default.txt",
"binding generic error",
orte_process_info.nodename, context->app, msg,
__FILE__, __LINE__);
} else {
orte_rtc_base_send_warn_show_help(write_fd,
"help-orte-odls-default.txt", "not bound",
orte_process_info.nodename, context->app, msg,
__FILE__, __LINE__);
return;
}
}
}
if (0 == rc && opal_hwloc_report_bindings) {
opal_output(0, "MCW rank %d is not bound (or bound to all available processors)", child->name.vpid);
/* avoid reporting it twice */
(void) mca_base_var_env_name ("hwloc_base_report_bindings", ¶m);
opal_unsetenv(param, environ_copy);
free(param);
}
} else {
/* convert the list to a cpuset */
cpuset = hwloc_bitmap_alloc();
if (0 != (rc = hwloc_bitmap_list_sscanf(cpuset, cpu_bitmap))) {
/* See comment above about "This may be a small memory leak" */
asprintf(&msg, "hwloc_bitmap_sscanf returned \"%s\" for the string \"%s\"",
opal_strerror(rc), cpu_bitmap);
if (NULL == msg) {
msg = "failed to convert bitmap list to hwloc bitmap";
}
if (OPAL_BINDING_REQUIRED(jobdat->map->binding) &&
OPAL_BINDING_POLICY_IS_SET(jobdat->map->binding)) {
/* If binding is required and a binding directive was explicitly
* given (i.e., we are not binding due to a default policy),
* send an error up the pipe (which exits -- it doesn't return).
*/
orte_rtc_base_send_error_show_help(write_fd, 1, "help-orte-odls-default.txt",
"binding generic error",
orte_process_info.nodename,
context->app, msg,
__FILE__, __LINE__);
} else {
orte_rtc_base_send_warn_show_help(write_fd,
"help-orte-odls-default.txt", "not bound",
orte_process_info.nodename, context->app, msg,
__FILE__, __LINE__);
free(cpu_bitmap);
return;
}
}
/* bind as specified */
rc = hwloc_set_cpubind(opal_hwloc_topology, cpuset, 0);
/* if we got an error and this wasn't a default binding policy, then report it */
if (rc < 0 && OPAL_BINDING_POLICY_IS_SET(jobdat->map->binding)) {
char *tmp = NULL;
if (errno == ENOSYS) {
msg = "hwloc indicates cpu binding not supported";
} else if (errno == EXDEV) {
msg = "hwloc indicates cpu binding cannot be enforced";
} else {
asprintf(&msg, "hwloc_set_cpubind returned \"%s\" for bitmap \"%s\"",
opal_strerror(rc), cpu_bitmap);
}
if (OPAL_BINDING_REQUIRED(jobdat->map->binding)) {
/* If binding is required, send an error up the pipe (which exits
-- it doesn't return). */
orte_rtc_base_send_error_show_help(write_fd, 1, "help-orte-odls-default.txt",
"binding generic error",
orte_process_info.nodename, context->app, msg,
__FILE__, __LINE__);
} else {
orte_rtc_base_send_warn_show_help(write_fd,
"help-orte-odls-default.txt", "not bound",
orte_process_info.nodename, context->app, msg,
__FILE__, __LINE__);
if (NULL != tmp) {
free(tmp);
free(msg);
}
return;
}
if (NULL != tmp) {
free(tmp);
free(msg);
}
}
if (0 == rc && opal_hwloc_report_bindings) {
char tmp1[1024], tmp2[1024];
hwloc_cpuset_t mycpus;
/* get the cpus we are bound to */
mycpus = hwloc_bitmap_alloc();
if (hwloc_get_cpubind(opal_hwloc_topology,
mycpus,
HWLOC_CPUBIND_PROCESS) < 0) {
opal_output(0, "MCW rank %d is not bound",
child->name.vpid);
} else {
if (OPAL_ERR_NOT_BOUND == opal_hwloc_base_cset2str(tmp1, sizeof(tmp1), opal_hwloc_topology, mycpus)) {
opal_output(0, "MCW rank %d is not bound (or bound to all available processors)", child->name.vpid);
} else {
opal_hwloc_base_cset2mapstr(tmp2, sizeof(tmp2), opal_hwloc_topology, mycpus);
opal_output(0, "MCW rank %d bound to %s: %s",
child->name.vpid, tmp1, tmp2);
}
}
hwloc_bitmap_free(mycpus);
/* avoid reporting it twice */
(void) mca_base_var_env_name ("hwloc_base_report_bindings", ¶m);
opal_unsetenv(param, environ_copy);
free(param);
}
/* set memory affinity policy - if we get an error, don't report
* anything unless the user actually specified the binding policy
*/
rc = opal_hwloc_base_set_process_membind_policy();
if (ORTE_SUCCESS != rc && OPAL_BINDING_POLICY_IS_SET(jobdat->map->binding)) {
if (errno == ENOSYS) {
msg = "hwloc indicates memory binding not supported";
} else if (errno == EXDEV) {
msg = "hwloc indicates memory binding cannot be enforced";
} else {
msg = "failed to bind memory";
}
if (OPAL_HWLOC_BASE_MBFA_ERROR == opal_hwloc_base_mbfa) {
/* If binding is required, send an error up the pipe (which exits
-- it doesn't return). */
orte_rtc_base_send_error_show_help(write_fd, 1, "help-orte-odls-default.txt",
"memory binding error",
orte_process_info.nodename, context->app, msg,
__FILE__, __LINE__);
} else {
orte_rtc_base_send_warn_show_help(write_fd,
"help-orte-odls-default.txt", "memory not bound",
orte_process_info.nodename, context->app, msg,
__FILE__, __LINE__);
free(cpu_bitmap);
return;
}
}
}
if (NULL != cpu_bitmap) {
free(cpu_bitmap);
}
}
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;
}
int orte_ess_base_proc_binding(void)
{
hwloc_obj_t node, obj;
hwloc_cpuset_t cpus, nodeset;
hwloc_obj_type_t target;
unsigned int cache_level = 0;
struct hwloc_topology_support *support;
char *map;
int ret;
char *error=NULL;
hwloc_cpuset_t mycpus;
/* Determine if we were pre-bound or not */
if (NULL != getenv(OPAL_MCA_PREFIX"orte_bound_at_launch")) {
orte_proc_is_bound = true;
if (NULL != (map = getenv(OPAL_MCA_PREFIX"orte_base_applied_binding"))) {
orte_proc_applied_binding = hwloc_bitmap_alloc();
if (0 != (ret = hwloc_bitmap_list_sscanf(orte_proc_applied_binding, map))) {
error = "applied_binding parse";
goto error;
}
}
}
/* see if we were bound when launched */
if (!orte_proc_is_bound) {
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Not bound at launch",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* we were not bound at launch */
if (NULL == opal_hwloc_topology) {
/* there is nothing we can do, so just return */
return ORTE_SUCCESS;
}
support = (struct hwloc_topology_support*)hwloc_topology_get_support(opal_hwloc_topology);
/* get our node object */
node = hwloc_get_root_obj(opal_hwloc_topology);
nodeset = opal_hwloc_base_get_available_cpus(opal_hwloc_topology, node);
/* get our bindings */
cpus = hwloc_bitmap_alloc();
if (hwloc_get_cpubind(opal_hwloc_topology, cpus, HWLOC_CPUBIND_PROCESS) < 0) {
/* we are NOT bound if get_cpubind fails, nor can we be bound - the
* environment does not support it
*/
hwloc_bitmap_free(cpus);
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Binding not supported",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
goto MOVEON;
}
/* we are bound if the two cpusets are not equal,
* or if there is only ONE cpu available to us
*/
if (0 != hwloc_bitmap_compare(cpus, nodeset) ||
opal_hwloc_base_single_cpu(nodeset) ||
opal_hwloc_base_single_cpu(cpus)) {
/* someone external set it - indicate it is set
* so that we know
*/
orte_proc_is_bound = true;
hwloc_bitmap_list_asprintf(&orte_process_info.cpuset, cpus);
hwloc_bitmap_free(cpus);
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Process was externally bound",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
} else if (support->cpubind->set_thisproc_cpubind &&
OPAL_BINDING_POLICY_IS_SET(opal_hwloc_binding_policy) &&
OPAL_BIND_TO_NONE != OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
/* the system is capable of doing processor affinity, but it
* has not yet been set - see if a slot_list was given
*/
hwloc_bitmap_zero(cpus);
if (OPAL_BIND_TO_CPUSET == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
if (OPAL_SUCCESS != (ret = opal_hwloc_base_slot_list_parse(opal_hwloc_base_slot_list,
opal_hwloc_topology,
OPAL_HWLOC_LOGICAL, cpus))) {
error = "Setting processor affinity failed";
hwloc_bitmap_free(cpus);
goto error;
}
if (0 > hwloc_set_cpubind(opal_hwloc_topology, cpus, 0)) {
error = "Setting processor affinity failed";
hwloc_bitmap_free(cpus);
goto error;
}
hwloc_bitmap_list_asprintf(&orte_process_info.cpuset, cpus);
hwloc_bitmap_free(cpus);
orte_proc_is_bound = true;
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Process bound according to slot_list",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
} else {
/* cleanup */
hwloc_bitmap_free(cpus);
/* get the node rank */
if (ORTE_NODE_RANK_INVALID == orte_process_info.my_node_rank) {
/* this is not an error - could be due to being
* direct launched - so just ignore and leave
* us unbound
*/
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Process not bound - no node rank available",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
goto MOVEON;
}
/* if the binding policy is hwthread, then we bind to the nrank-th
* hwthread on this node
*/
if (OPAL_BIND_TO_HWTHREAD == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
if (NULL == (obj = opal_hwloc_base_get_obj_by_type(opal_hwloc_topology, HWLOC_OBJ_PU,
0, orte_process_info.my_node_rank, OPAL_HWLOC_LOGICAL))) {
ret = ORTE_ERR_NOT_FOUND;
error = "Getting hwthread object";
goto error;
}
cpus = opal_hwloc_base_get_available_cpus(opal_hwloc_topology, obj);
if (0 > hwloc_set_cpubind(opal_hwloc_topology, cpus, 0)) {
ret = ORTE_ERROR;
error = "Setting processor affinity failed";
goto error;
}
hwloc_bitmap_list_asprintf(&orte_process_info.cpuset, cpus);
hwloc_bitmap_free(cpus);
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Process bound to hwthread",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
} else if (OPAL_BIND_TO_CORE == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
/* if the binding policy is core, then we bind to the nrank-th
* core on this node
*/
if (NULL == (obj = opal_hwloc_base_get_obj_by_type(opal_hwloc_topology, HWLOC_OBJ_CORE,
0, orte_process_info.my_node_rank, OPAL_HWLOC_LOGICAL))) {
ret = ORTE_ERR_NOT_FOUND;
error = "Getting core object";
goto error;
}
cpus = opal_hwloc_base_get_available_cpus(opal_hwloc_topology, obj);
if (0 > hwloc_set_cpubind(opal_hwloc_topology, cpus, 0)) {
error = "Setting processor affinity failed";
ret = ORTE_ERROR;
goto error;
}
hwloc_bitmap_list_asprintf(&orte_process_info.cpuset, cpus);
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Process bound to core",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
} else {
/* for all higher binding policies, we bind to the specified
* object that the nrank-th core belongs to
*/
if (NULL == (obj = opal_hwloc_base_get_obj_by_type(opal_hwloc_topology, HWLOC_OBJ_CORE,
0, orte_process_info.my_node_rank, OPAL_HWLOC_LOGICAL))) {
ret = ORTE_ERR_NOT_FOUND;
error = "Getting core object";
goto error;
}
if (OPAL_BIND_TO_L1CACHE == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
target = HWLOC_OBJ_CACHE;
cache_level = 1;
} else if (OPAL_BIND_TO_L2CACHE == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
target = HWLOC_OBJ_CACHE;
cache_level = 2;
} else if (OPAL_BIND_TO_L3CACHE == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
target = HWLOC_OBJ_CACHE;
cache_level = 3;
} else if (OPAL_BIND_TO_SOCKET == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
target = HWLOC_OBJ_SOCKET;
} else if (OPAL_BIND_TO_NUMA == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
target = HWLOC_OBJ_NODE;
} else {
ret = ORTE_ERR_NOT_FOUND;
error = "Binding policy not known";
goto error;
}
for (obj = obj->parent; NULL != obj; obj = obj->parent) {
if (target == obj->type) {
if (HWLOC_OBJ_CACHE == target && cache_level != obj->attr->cache.depth) {
continue;
}
/* this is the place! */
cpus = opal_hwloc_base_get_available_cpus(opal_hwloc_topology, obj);
if (0 > hwloc_set_cpubind(opal_hwloc_topology, cpus, 0)) {
ret = ORTE_ERROR;
error = "Setting processor affinity failed";
goto error;
}
hwloc_bitmap_list_asprintf(&orte_process_info.cpuset, cpus);
orte_proc_is_bound = true;
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Process bound to %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
hwloc_obj_type_string(target)));
break;
}
}
if (!orte_proc_is_bound) {
ret = ORTE_ERROR;
error = "Setting processor affinity failed";
goto error;
}
}
}
}
} else {
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Process bound at launch",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
}
MOVEON:
/* get or update our local cpuset - it will get used multiple
* times, so it's more efficient to keep a global copy
*/
opal_hwloc_base_get_local_cpuset();
/* get the cpus we are bound to */
mycpus = hwloc_bitmap_alloc();
if (hwloc_get_cpubind(opal_hwloc_topology,
mycpus,
HWLOC_CPUBIND_PROCESS) < 0) {
if (NULL != orte_process_info.cpuset) {
free(orte_process_info.cpuset);
orte_process_info.cpuset = NULL;
}
if (opal_hwloc_report_bindings || 4 < opal_output_get_verbosity(orte_ess_base_framework.framework_output)) {
opal_output(0, "MCW rank %d is not bound",
ORTE_PROC_MY_NAME->vpid);
}
} else {
/* store/update the string representation of our local binding */
if (NULL != orte_process_info.cpuset) {
free(orte_process_info.cpuset);
orte_process_info.cpuset = NULL;
}
hwloc_bitmap_list_asprintf(&orte_process_info.cpuset, mycpus);
/* report the binding, if requested */
if (opal_hwloc_report_bindings || 4 < opal_output_get_verbosity(orte_ess_base_framework.framework_output)) {
char tmp1[1024], tmp2[1024];
if (OPAL_ERR_NOT_BOUND == opal_hwloc_base_cset2str(tmp1, sizeof(tmp1), opal_hwloc_topology, mycpus)) {
opal_output(0, "MCW rank %d is not bound (or bound to all available processors)", ORTE_PROC_MY_NAME->vpid);
} else {
opal_hwloc_base_cset2mapstr(tmp2, sizeof(tmp2), opal_hwloc_topology, mycpus);
opal_output(0, "MCW rank %d bound to %s: %s",
ORTE_PROC_MY_NAME->vpid, tmp1, tmp2);
}
}
}
hwloc_bitmap_free(mycpus);
/* push our cpuset so others can calculate our locality */
if (NULL != orte_process_info.cpuset) {
OPAL_MODEX_SEND_VALUE(ret, OPAL_PMIX_GLOBAL, OPAL_PMIX_CPUSET,
orte_process_info.cpuset, OPAL_STRING);
}
return ORTE_SUCCESS;
error:
if (ORTE_ERR_SILENT != ret) {
orte_show_help("help-orte-runtime",
"orte_init:startup:internal-failure",
true, error, ORTE_ERROR_NAME(ret), ret);
}
return ORTE_ERR_SILENT;
}
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;
}
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;
}
int orte_rmaps_base_compute_bindings(orte_job_t *jdata)
{
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: compute bindings for job %s with policy %s",
ORTE_JOBID_PRINT(jdata->jobid),
opal_hwloc_base_print_binding(jdata->map->binding));
if (ORTE_MAPPING_BYUSER == ORTE_GET_MAPPING_POLICY(orte_rmaps_base.mapping)) {
/* user specified binding by rankfile - nothing for us to do */
return ORTE_SUCCESS;
}
if (OPAL_BIND_TO_CPUSET == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
int rc;
/* cpuset was given - setup the bindings */
if (ORTE_SUCCESS != (rc = bind_to_cpuset(jdata))) {
ORTE_ERROR_LOG(rc);
}
return rc;
}
if (!OPAL_BINDING_POLICY_IS_SET(jdata->map->binding) ||
OPAL_BIND_TO_NONE == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
/* no binding requested */
return ORTE_SUCCESS;
}
if (OPAL_BIND_TO_BOARD == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
/* doesn't do anything at this time */
return ORTE_SUCCESS;
}
/* binding requested */
/* if the job was mapped by the corresponding target, then
* we bind in place
*
* otherwise, we have to bind either up or down the hwloc
* tree. If we are binding upwards (e.g., mapped to hwthread
* but binding to core), then we just climb the tree to find
* the first matching object.
*
* if we are binding downwards (e.g., mapped to node and bind
* to core), then we have to do a round-robin assigment of
* procs to the resources below.
*/
if (ORTE_MAPPING_BYDIST == ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
int rc = ORTE_SUCCESS;
if (OPAL_BIND_TO_NUMA == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bindings for job %s - dist to numa",
ORTE_JOBID_PRINT(jdata->jobid));
if (ORTE_SUCCESS != (rc = bind_in_place(jdata, HWLOC_OBJ_NODE, 0))) {
ORTE_ERROR_LOG(rc);
}
} else if (OPAL_BIND_TO_NUMA < OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
if (OPAL_BIND_TO_HWTHREAD == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
if (ORTE_SUCCESS != (rc = bind_downwards(jdata, HWLOC_OBJ_PU, 0))) {
ORTE_ERROR_LOG(rc);
}
} else if (OPAL_BIND_TO_CORE == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
if (ORTE_SUCCESS != (rc = bind_downwards(jdata, HWLOC_OBJ_CORE, 0))) {
ORTE_ERROR_LOG(rc);
}
} else if (OPAL_BIND_TO_L1CACHE == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
if (ORTE_SUCCESS != (rc = bind_downwards(jdata, HWLOC_OBJ_CACHE, 1))) {
ORTE_ERROR_LOG(rc);
}
} else if (OPAL_BIND_TO_L2CACHE == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
if (ORTE_SUCCESS != (rc = bind_downwards(jdata, HWLOC_OBJ_CACHE, 2))) {
ORTE_ERROR_LOG(rc);
}
} else if (OPAL_BIND_TO_L3CACHE == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
if (ORTE_SUCCESS != (rc = bind_downwards(jdata, HWLOC_OBJ_CACHE, 3))) {
ORTE_ERROR_LOG(rc);
}
} else if (OPAL_BIND_TO_SOCKET == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
if (ORTE_SUCCESS != (rc = bind_downwards(jdata, HWLOC_OBJ_SOCKET, 0))) {
ORTE_ERROR_LOG(rc);
}
}
}
/* if the binding policy is less than numa, then we are unbound - so
* just ignore this and return (should have been caught in prior
* tests anyway as only options meeting that criteria are "none"
* and "board")
*/
return rc;
} else if (OPAL_BIND_TO_HWTHREAD == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
int rc;
if (ORTE_MAPPING_BYHWTHREAD == ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bindings for job %s - hwthread to hwthread",
ORTE_JOBID_PRINT(jdata->jobid));
if (ORTE_SUCCESS != (rc = bind_in_place(jdata, HWLOC_OBJ_PU, 0))) {
ORTE_ERROR_LOG(rc);
}
return rc;
}
/* HW threads are at the bottom, so we must have mapped somewhere
* above this level - and we need to bind downwards
*/
if (ORTE_SUCCESS != (rc = bind_downwards(jdata, HWLOC_OBJ_PU, 0))) {
ORTE_ERROR_LOG(rc);
}
return rc;
} else if (OPAL_BIND_TO_CORE == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
int rc;
if (ORTE_MAPPING_BYCORE == ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bindings for job %s - core to core",
ORTE_JOBID_PRINT(jdata->jobid));
if (ORTE_SUCCESS != (rc = bind_in_place(jdata, HWLOC_OBJ_CORE, 0))) {
ORTE_ERROR_LOG(rc);
}
return rc;
}
/* if the mapping policy used is less than bycore, then it is a
* downward binding - i.e., the locale is above the binding location.
* for example, if we map-to-socket and bind-to-core, then we compare
* the mapping value of ORTE_MAPPING_BYCORE to ORTE_MAPPING_BYSOCKET.
* In this case, BYCORE > BYSOCKET, so we know that the locale is
* above the desired binding level (sockets are at a higher level than
* the desired core binding level), and we will have to bind downwards
*/
if (ORTE_MAPPING_BYCORE > ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
if (ORTE_SUCCESS != (rc = bind_downwards(jdata, HWLOC_OBJ_CORE, 0))) {
ORTE_ERROR_LOG(rc);
}
} else {
if (ORTE_SUCCESS != (rc = bind_upwards(jdata, HWLOC_OBJ_CORE, 0))) {
ORTE_ERROR_LOG(rc);
}
}
return rc;
} else if (OPAL_BIND_TO_L1CACHE == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
int rc;
if (ORTE_MAPPING_BYL1CACHE == ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bindings for job %s - L1cache to L1cache",
ORTE_JOBID_PRINT(jdata->jobid));
if (ORTE_SUCCESS != (rc = bind_in_place(jdata, HWLOC_OBJ_CACHE, 1))) {
ORTE_ERROR_LOG(rc);
}
return rc;
}
/* if the mapping policy is less than l1cache, then it is a
* downward binding
*/
if (ORTE_MAPPING_BYL1CACHE > ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
if (ORTE_SUCCESS != (rc = bind_downwards(jdata, HWLOC_OBJ_CACHE, 1))) {
ORTE_ERROR_LOG(rc);
}
} else {
if (ORTE_SUCCESS != (rc = bind_upwards(jdata, HWLOC_OBJ_CACHE, 1))) {
ORTE_ERROR_LOG(rc);
}
}
return rc;
} else if (OPAL_BIND_TO_L2CACHE == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
int rc;
if (ORTE_MAPPING_BYL2CACHE == ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bindings for job %s - L2cache to L2cache",
ORTE_JOBID_PRINT(jdata->jobid));
if (ORTE_SUCCESS != (rc = bind_in_place(jdata, HWLOC_OBJ_CACHE, 2))) {
ORTE_ERROR_LOG(rc);
}
return rc;
}
/* if the mapping policy is less than l2cache, then it is a
* downward binding
*/
if (ORTE_MAPPING_BYL2CACHE > ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
if (ORTE_SUCCESS != (rc = bind_downwards(jdata, HWLOC_OBJ_CACHE, 2))) {
ORTE_ERROR_LOG(rc);
}
} else {
if (ORTE_SUCCESS != (rc = bind_upwards(jdata, HWLOC_OBJ_CACHE, 2))) {
ORTE_ERROR_LOG(rc);
}
}
return rc;
} else if (OPAL_BIND_TO_L3CACHE == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
int rc;
if (ORTE_MAPPING_BYL3CACHE == ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bindings for job %s - L3cache to L3cache",
ORTE_JOBID_PRINT(jdata->jobid));
if (ORTE_SUCCESS != (rc = bind_in_place(jdata, HWLOC_OBJ_CACHE, 3))) {
ORTE_ERROR_LOG(rc);
}
return rc;
}
/* if the mapping policy is less than l3cache, then it is a
* downward binding
*/
if (ORTE_MAPPING_BYL3CACHE > ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
if (ORTE_SUCCESS != (rc = bind_downwards(jdata, HWLOC_OBJ_CACHE, 3))) {
ORTE_ERROR_LOG(rc);
}
} else {
if (ORTE_SUCCESS != (rc = bind_upwards(jdata, HWLOC_OBJ_CACHE, 3))) {
ORTE_ERROR_LOG(rc);
}
}
return rc;
} else if (OPAL_BIND_TO_SOCKET == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
int rc;
if (ORTE_MAPPING_BYSOCKET == ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bindings for job %s - socket to socket",
ORTE_JOBID_PRINT(jdata->jobid));
if (ORTE_SUCCESS != (rc = bind_in_place(jdata, HWLOC_OBJ_SOCKET, 0))) {
ORTE_ERROR_LOG(rc);
}
return rc;
}
/* if the mapping policy is less than bysocket, then it is a
* downward binding
*/
if (ORTE_MAPPING_BYSOCKET > ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
if (ORTE_SUCCESS != (rc = bind_downwards(jdata, HWLOC_OBJ_SOCKET, 0))) {
ORTE_ERROR_LOG(rc);
}
} else {
if (ORTE_SUCCESS != (rc = bind_upwards(jdata, HWLOC_OBJ_SOCKET, 0))) {
ORTE_ERROR_LOG(rc);
}
}
return rc;
} else if (OPAL_BIND_TO_NUMA == OPAL_GET_BINDING_POLICY(jdata->map->binding)) {
int rc;
if (ORTE_MAPPING_BYNUMA == ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bindings for job %s - numa to numa",
ORTE_JOBID_PRINT(jdata->jobid));
if (ORTE_SUCCESS != (rc = bind_in_place(jdata, HWLOC_OBJ_NODE, 0))) {
ORTE_ERROR_LOG(rc);
}
return rc;
}
/* if the mapping policy is less than numa, then it is a
* downward binding
*/
if (ORTE_MAPPING_BYNUMA > ORTE_GET_MAPPING_POLICY(jdata->map->mapping)) {
if (ORTE_SUCCESS != (rc = bind_downwards(jdata, HWLOC_OBJ_NODE, 0))) {
ORTE_ERROR_LOG(rc);
}
} else {
if (ORTE_SUCCESS != (rc = bind_upwards(jdata, HWLOC_OBJ_NODE, 0))) {
ORTE_ERROR_LOG(rc);
}
}
return rc;
} else {
ORTE_ERROR_LOG(ORTE_ERR_NOT_SUPPORTED);
return ORTE_ERR_NOT_SUPPORTED;
}
return ORTE_SUCCESS;
}
int orte_ess_base_proc_binding(void)
{
#if OPAL_HAVE_HWLOC
hwloc_obj_t node, obj;
hwloc_cpuset_t cpus, nodeset;
hwloc_obj_type_t target;
unsigned int cache_level = 0;
struct hwloc_topology_support *support;
char *map;
int ret;
char *error;
/* Determine if we were pre-bound or not */
if (NULL != getenv("OMPI_MCA_orte_bound_at_launch")) {
orte_proc_is_bound = true;
if (NULL != (map = getenv("OMPI_MCA_orte_base_applied_binding"))) {
orte_proc_applied_binding = hwloc_bitmap_alloc();
if (0 != (ret = hwloc_bitmap_list_sscanf(orte_proc_applied_binding, map))) {
error = "applied_binding parse";
goto error;
}
}
}
/* see if we were bound when launched */
if (!orte_proc_is_bound) {
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_output,
"%s Not bound at launch",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* we were not bound at launch */
if (NULL != opal_hwloc_topology) {
support = (struct hwloc_topology_support*)hwloc_topology_get_support(opal_hwloc_topology);
/* get our node object */
node = hwloc_get_root_obj(opal_hwloc_topology);
nodeset = opal_hwloc_base_get_available_cpus(opal_hwloc_topology, node);
/* get our bindings */
cpus = hwloc_bitmap_alloc();
if (hwloc_get_cpubind(opal_hwloc_topology, cpus, HWLOC_CPUBIND_PROCESS) < 0) {
/* we are NOT bound if get_cpubind fails, nor can we be bound - the
* environment does not support it
*/
hwloc_bitmap_free(cpus);
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_output,
"%s Binding not supported",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
goto MOVEON;
}
/* we are bound if the two cpusets are not equal,
* or if there is only ONE cpu available to us
*/
if (0 != hwloc_bitmap_compare(cpus, nodeset) ||
opal_hwloc_base_single_cpu(nodeset) ||
opal_hwloc_base_single_cpu(cpus)) {
/* someone external set it - indicate it is set
* so that we know
*/
orte_proc_is_bound = true;
hwloc_bitmap_free(cpus);
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_output,
"%s Process was externally bound",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
} else if (support->cpubind->set_thisproc_cpubind &&
OPAL_BINDING_POLICY_IS_SET(opal_hwloc_binding_policy) &&
OPAL_BIND_TO_NONE != OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
/* the system is capable of doing processor affinity, but it
* has not yet been set - see if a slot_list was given
*/
hwloc_bitmap_zero(cpus);
if (OPAL_BIND_TO_CPUSET == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
if (OPAL_SUCCESS != (ret = opal_hwloc_base_slot_list_parse(opal_hwloc_base_slot_list,
opal_hwloc_topology, cpus))) {
error = "Setting processor affinity failed";
hwloc_bitmap_free(cpus);
goto error;
}
if (0 > hwloc_set_cpubind(opal_hwloc_topology, cpus, 0)) {
error = "Setting processor affinity failed";
hwloc_bitmap_free(cpus);
goto error;
}
/* try to find a level and index for this location */
opal_hwloc_base_get_level_and_index(cpus, &orte_process_info.bind_level, &orte_process_info.bind_idx);
/* cleanup */
hwloc_bitmap_free(cpus);
orte_proc_is_bound = true;
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_output,
"%s Process bound according to slot_list",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
} else {
/* cleanup */
hwloc_bitmap_free(cpus);
/* get the node rank */
if (ORTE_NODE_RANK_INVALID == orte_process_info.my_node_rank) {
/* this is not an error - could be due to being
* direct launched - so just ignore and leave
* us unbound
*/
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_output,
"%s Process not bound - no node rank available",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
goto MOVEON;
}
/* if the binding policy is hwthread, then we bind to the nrank-th
* hwthread on this node
*/
if (OPAL_BIND_TO_HWTHREAD == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
if (NULL == (obj = opal_hwloc_base_get_obj_by_type(opal_hwloc_topology, HWLOC_OBJ_PU,
0, orte_process_info.my_node_rank, OPAL_HWLOC_LOGICAL))) {
ret = ORTE_ERR_NOT_FOUND;
error = "Getting hwthread object";
goto error;
}
cpus = opal_hwloc_base_get_available_cpus(opal_hwloc_topology, obj);
if (0 > hwloc_set_cpubind(opal_hwloc_topology, cpus, 0)) {
ret = ORTE_ERROR;
error = "Setting processor affinity failed";
goto error;
}
orte_process_info.bind_level = OPAL_HWLOC_HWTHREAD_LEVEL;
orte_process_info.bind_idx = orte_process_info.my_node_rank;
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_output,
"%s Process bound to hwthread",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
} else if (OPAL_BIND_TO_CORE == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
/* if the binding policy is core, then we bind to the nrank-th
* core on this node
*/
if (NULL == (obj = opal_hwloc_base_get_obj_by_type(opal_hwloc_topology, HWLOC_OBJ_CORE,
0, orte_process_info.my_node_rank, OPAL_HWLOC_LOGICAL))) {
ret = ORTE_ERR_NOT_FOUND;
error = "Getting core object";
goto error;
}
cpus = opal_hwloc_base_get_available_cpus(opal_hwloc_topology, obj);
if (0 > hwloc_set_cpubind(opal_hwloc_topology, cpus, 0)) {
error = "Setting processor affinity failed";
ret = ORTE_ERROR;
goto error;
}
orte_process_info.bind_level = OPAL_HWLOC_CORE_LEVEL;
orte_process_info.bind_idx = orte_process_info.my_node_rank;
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_output,
"%s Process bound to core",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
} else {
/* for all higher binding policies, we bind to the specified
* object that the nrank-th core belongs to
*/
if (NULL == (obj = opal_hwloc_base_get_obj_by_type(opal_hwloc_topology, HWLOC_OBJ_CORE,
0, orte_process_info.my_node_rank, OPAL_HWLOC_LOGICAL))) {
ret = ORTE_ERR_NOT_FOUND;
error = "Getting core object";
goto error;
}
if (OPAL_BIND_TO_L1CACHE == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
target = HWLOC_OBJ_CACHE;
cache_level = 1;
orte_process_info.bind_level = OPAL_HWLOC_L1CACHE_LEVEL;
} else if (OPAL_BIND_TO_L2CACHE == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
target = HWLOC_OBJ_CACHE;
cache_level = 2;
orte_process_info.bind_level = OPAL_HWLOC_L2CACHE_LEVEL;
} else if (OPAL_BIND_TO_L3CACHE == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
target = HWLOC_OBJ_CACHE;
cache_level = 3;
orte_process_info.bind_level = OPAL_HWLOC_L3CACHE_LEVEL;
} else if (OPAL_BIND_TO_SOCKET == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
target = HWLOC_OBJ_SOCKET;
orte_process_info.bind_level = OPAL_HWLOC_SOCKET_LEVEL;
} else if (OPAL_BIND_TO_NUMA == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
target = HWLOC_OBJ_NODE;
orte_process_info.bind_level = OPAL_HWLOC_NUMA_LEVEL;
} else {
ret = ORTE_ERR_NOT_FOUND;
error = "Binding policy not known";
goto error;
}
for (obj = obj->parent; NULL != obj; obj = obj->parent) {
if (target == obj->type) {
if (HWLOC_OBJ_CACHE == target && cache_level != obj->attr->cache.depth) {
continue;
}
/* this is the place! */
cpus = opal_hwloc_base_get_available_cpus(opal_hwloc_topology, obj);
if (0 > hwloc_set_cpubind(opal_hwloc_topology, cpus, 0)) {
ret = ORTE_ERROR;
error = "Setting processor affinity failed";
goto error;
}
orte_process_info.bind_idx = opal_hwloc_base_get_obj_idx(opal_hwloc_topology,
obj, OPAL_HWLOC_LOGICAL);
orte_proc_is_bound = true;
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_output,
"%s Process bound to %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
opal_hwloc_base_print_level(orte_process_info.bind_level)));
break;
}
}
if (!orte_proc_is_bound) {
ret = ORTE_ERROR;
error = "Setting processor affinity failed";
goto error;
}
}
}
}
}
} else {
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_output,
"%s Process bound at launch",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
}
MOVEON:
/* get or update our local cpuset - it will get used multiple
* times, so it's more efficient to keep a global copy
*/
opal_hwloc_base_get_local_cpuset();
/* report bindings, if requested */
if (opal_hwloc_report_bindings) {
char bindings[64];
hwloc_obj_t root;
hwloc_cpuset_t cpus;
/* get the root object for this node */
root = hwloc_get_root_obj(opal_hwloc_topology);
cpus = opal_hwloc_base_get_available_cpus(opal_hwloc_topology, root);
/* we are not bound if this equals our cpuset */
if (0 == hwloc_bitmap_compare(cpus, opal_hwloc_my_cpuset)) {
opal_output(0, "%s is not bound",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
} else {
hwloc_bitmap_list_snprintf(bindings, 64, opal_hwloc_my_cpuset);
opal_output(0, "%s is bound to cpus %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
bindings);
}
}
return ORTE_SUCCESS;
error:
if (ORTE_ERR_SILENT != ret) {
orte_show_help("help-orte-runtime",
"orte_init:startup:internal-failure",
true, error, ORTE_ERROR_NAME(ret), ret);
}
return ORTE_ERR_SILENT;
#else
return ORTE_SUCCESS;
#endif
}
static int do_child(orte_app_context_t* context,
orte_proc_t *child,
char **environ_copy,
orte_job_t *jobdat, int write_fd,
orte_iof_base_io_conf_t opts)
{
int i, rc;
sigset_t sigs;
long fd, fdmax = sysconf(_SC_OPEN_MAX);
char *param, *msg;
if (orte_forward_job_control) {
/* Set a new process group for this child, so that a
SIGSTOP can be sent to it without being sent to the
orted. */
setpgid(0, 0);
}
/* Setup the pipe to be close-on-exec */
fcntl(write_fd, F_SETFD, FD_CLOEXEC);
if (NULL != child) {
/* setup stdout/stderr so that any error messages that we
may print out will get displayed back at orterun.
NOTE: Definitely do this AFTER we check contexts so
that any error message from those two functions doesn't
come out to the user. IF we didn't do it in this order,
THEN a user who gives us a bad executable name or
working directory would get N error messages, where
N=num_procs. This would be very annoying for large
jobs, so instead we set things up so that orterun
always outputs a nice, single message indicating what
happened
*/
if (ORTE_SUCCESS != (i = orte_iof_base_setup_child(&opts,
&environ_copy))) {
ORTE_ERROR_LOG(i);
send_error_show_help(write_fd, 1,
"help-orte-odls-default.txt",
"iof setup failed",
orte_process_info.nodename, context->app);
/* Does not return */
}
#if OPAL_HAVE_HWLOC
{
hwloc_cpuset_t cpuset;
hwloc_obj_t root;
opal_hwloc_topo_data_t *sum;
/* Set process affinity, if given */
if (NULL == child->cpu_bitmap) {
/* if the daemon is bound, then we need to "free" this proc */
if (NULL != orte_daemon_cores) {
root = hwloc_get_root_obj(opal_hwloc_topology);
if (NULL == root->userdata) {
send_warn_show_help(write_fd,
"help-orte-odls-default.txt", "incorrectly bound",
orte_process_info.nodename, context->app,
__FILE__, __LINE__);
}
sum = (opal_hwloc_topo_data_t*)root->userdata;
/* bind this proc to all available processors */
hwloc_set_cpubind(opal_hwloc_topology, sum->available, 0);
}
if (opal_hwloc_report_bindings) {
opal_output(0, "MCW rank %d is not bound (or bound to all available processors)", child->name.vpid);
/* avoid reporting it twice */
(void) mca_base_var_env_name ("hwloc_base_report_bindings", ¶m);
opal_unsetenv(param, &environ_copy);
free(param);
}
} else {
if (0 == strlen(child->cpu_bitmap)) {
/* this proc is not bound */
if (opal_hwloc_report_bindings) {
opal_output(0, "MCW rank %d is not bound (or bound to all available processors)", child->name.vpid);
/* avoid reporting it twice */
(void) mca_base_var_env_name ("hwloc_base_report_bindings", ¶m);
opal_unsetenv(param, &environ_copy);
free(param);
}
/* if the daemon is bound, then we need to "free" this proc */
if (NULL != orte_daemon_cores) {
root = hwloc_get_root_obj(opal_hwloc_topology);
if (NULL == root->userdata) {
send_warn_show_help(write_fd,
"help-orte-odls-default.txt", "incorrectly bound",
orte_process_info.nodename, context->app,
__FILE__, __LINE__);
}
sum = (opal_hwloc_topo_data_t*)root->userdata;
/* bind this proc to all available processors */
hwloc_set_cpubind(opal_hwloc_topology, sum->available, 0);
}
/* provide a nice string representation of what we bound to */
(void) mca_base_var_env_name ("orte_base_applied_binding", ¶m);
opal_setenv(param, child->cpu_bitmap, true, &environ_copy);
free(param);
goto PROCEED;
}
/* convert the list to a cpu bitmap */
cpuset = hwloc_bitmap_alloc();
if (0 != (rc = hwloc_bitmap_list_sscanf(cpuset, child->cpu_bitmap))) {
/* See comment above about "This may be a small memory leak" */
asprintf(&msg, "hwloc_bitmap_sscanf returned \"%s\" for the string \"%s\"",
opal_strerror(rc), child->cpu_bitmap);
if (NULL == msg) {
msg = "failed to convert bitmap list to hwloc bitmap";
}
if (OPAL_BINDING_REQUIRED(jobdat->map->binding) &&
OPAL_BINDING_POLICY_IS_SET(jobdat->map->binding)) {
/* If binding is required and a binding directive was explicitly
* given (i.e., we are not binding due to a default policy),
* send an error up the pipe (which exits -- it doesn't return).
*/
send_error_show_help(write_fd, 1, "help-orte-odls-default.txt",
"binding generic error",
orte_process_info.nodename,
context->app, msg,
__FILE__, __LINE__);
} else {
send_warn_show_help(write_fd,
"help-orte-odls-default.txt", "not bound",
orte_process_info.nodename, context->app, msg,
__FILE__, __LINE__);
goto PROCEED;
}
}
/* bind as specified */
rc = hwloc_set_cpubind(opal_hwloc_topology, cpuset, 0);
/* if we got an error and this wasn't a default binding policy, then report it */
if (rc < 0 && OPAL_BINDING_POLICY_IS_SET(jobdat->map->binding)) {
char *tmp = NULL;
if (errno == ENOSYS) {
msg = "hwloc indicates cpu binding not supported";
} else if (errno == EXDEV) {
msg = "hwloc indicates cpu binding cannot be enforced";
} else {
asprintf(&msg, "hwloc_set_cpubind returned \"%s\" for bitmap \"%s\"",
opal_strerror(rc), child->cpu_bitmap);
}
if (OPAL_BINDING_REQUIRED(jobdat->map->binding)) {
/* If binding is required, send an error up the pipe (which exits
-- it doesn't return). */
send_error_show_help(write_fd, 1, "help-orte-odls-default.txt",
"binding generic error",
orte_process_info.nodename, context->app, msg,
__FILE__, __LINE__);
} else {
send_warn_show_help(write_fd,
"help-orte-odls-default.txt", "not bound",
orte_process_info.nodename, context->app, msg,
__FILE__, __LINE__);
if (NULL != tmp) {
free(tmp);
free(msg);
}
goto PROCEED;
}
if (NULL != tmp) {
free(tmp);
free(msg);
}
}
if (0 == rc && opal_hwloc_report_bindings) {
char tmp1[1024], tmp2[1024];
hwloc_cpuset_t mycpus;
/* get the cpus we are bound to */
mycpus = hwloc_bitmap_alloc();
if (hwloc_get_cpubind(opal_hwloc_topology,
mycpus,
HWLOC_CPUBIND_PROCESS) < 0) {
opal_output(0, "MCW rank %d is not bound",
child->name.vpid);
} else {
if (OPAL_ERR_NOT_BOUND == opal_hwloc_base_cset2str(tmp1, sizeof(tmp1), opal_hwloc_topology, mycpus)) {
opal_output(0, "MCW rank %d is not bound (or bound to all available processors)", child->name.vpid);
} else {
opal_hwloc_base_cset2mapstr(tmp2, sizeof(tmp2), opal_hwloc_topology, mycpus);
opal_output(0, "MCW rank %d bound to %s: %s",
child->name.vpid, tmp1, tmp2);
}
}
hwloc_bitmap_free(mycpus);
/* avoid reporting it twice */
(void) mca_base_var_env_name ("hwloc_base_report_bindings", ¶m);
opal_unsetenv(param, &environ_copy);
free(param);
}
/* set memory affinity policy - if we get an error, don't report
* anything unless the user actually specified the binding policy
*/
rc = opal_hwloc_base_set_process_membind_policy();
if (ORTE_SUCCESS != rc && OPAL_BINDING_POLICY_IS_SET(jobdat->map->binding)) {
if (errno == ENOSYS) {
msg = "hwloc indicates memory binding not supported";
} else if (errno == EXDEV) {
msg = "hwloc indicates memory binding cannot be enforced";
} else {
msg = "failed to bind memory";
}
if (OPAL_HWLOC_BASE_MBFA_ERROR == opal_hwloc_base_mbfa) {
/* If binding is required, send an error up the pipe (which exits
-- it doesn't return). */
send_error_show_help(write_fd, 1, "help-orte-odls-default.txt",
"memory binding error",
orte_process_info.nodename, context->app, msg,
__FILE__, __LINE__);
} else {
send_warn_show_help(write_fd,
"help-orte-odls-default.txt", "memory not bound",
orte_process_info.nodename, context->app, msg,
__FILE__, __LINE__);
goto PROCEED;
}
}
}
}
#endif
} else if (!(ORTE_JOB_CONTROL_FORWARD_OUTPUT & jobdat->controls)) {
/* tie stdin/out/err/internal to /dev/null */
int fdnull;
for (i=0; i < 3; i++) {
fdnull = open("/dev/null", O_RDONLY, 0);
if (fdnull > i && i != write_fd) {
dup2(fdnull, i);
}
close(fdnull);
}
fdnull = open("/dev/null", O_RDONLY, 0);
if (fdnull > opts.p_internal[1]) {
dup2(fdnull, opts.p_internal[1]);
}
close(fdnull);
}
#if OPAL_HAVE_HWLOC
PROCEED:
#endif
/* if the user requested it, set the system resource limits */
if (OPAL_SUCCESS != (rc = opal_util_init_sys_limits(&msg))) {
send_error_show_help(write_fd, 1, "help-orte-odls-default.txt",
"set limit",
orte_process_info.nodename, context->app,
__FILE__, __LINE__, msg);
}
/* ensure we only do this once */
(void) mca_base_var_env_name("opal_set_max_sys_limits", ¶m);
opal_unsetenv(param, &environ_copy);
free(param);
/* close all file descriptors w/ exception of stdin/stdout/stderr,
the pipe used for the IOF INTERNAL messages, and the pipe up to
the parent. */
for(fd=3; fd<fdmax; fd++) {
if (fd != opts.p_internal[1] && fd != write_fd) {
close(fd);
}
}
if (context->argv == NULL) {
context->argv = malloc(sizeof(char*)*2);
context->argv[0] = strdup(context->app);
context->argv[1] = NULL;
}
/* Set signal handlers back to the default. Do this close to
the exev() because the event library may (and likely will)
reset them. If we don't do this, the event library may
have left some set that, at least on some OS's, don't get
reset via fork() or exec(). Hence, the launched process
could be unkillable (for example). */
set_handler_default(SIGTERM);
set_handler_default(SIGINT);
set_handler_default(SIGHUP);
set_handler_default(SIGPIPE);
set_handler_default(SIGCHLD);
/* Unblock all signals, for many of the same reasons that we
set the default handlers, above. This is noticable on
Linux where the event library blocks SIGTERM, but we don't
want that blocked by the launched process. */
sigprocmask(0, 0, &sigs);
sigprocmask(SIG_UNBLOCK, &sigs, 0);
/* Exec the new executable */
if (10 < opal_output_get_verbosity(orte_odls_base_framework.framework_output)) {
int jout;
opal_output(0, "%s STARTING %s", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), context->app);
for (jout=0; NULL != context->argv[jout]; jout++) {
opal_output(0, "%s\tARGV[%d]: %s", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), jout, context->argv[jout]);
}
for (jout=0; NULL != environ_copy[jout]; jout++) {
opal_output(0, "%s\tENVIRON[%d]: %s", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), jout, environ_copy[jout]);
}
}
execve(context->app, context->argv, environ_copy);
send_error_show_help(write_fd, 1,
"help-orte-odls-default.txt", "execve error",
orte_process_info.nodename, context->app, strerror(errno));
/* Does not return */
}
int orte_rmaps_base_compute_bindings(orte_job_t *jdata)
{
hwloc_obj_type_t hwb, hwm;
unsigned clvl=0, clvm=0;
opal_binding_policy_t bind;
orte_mapping_policy_t map;
orte_node_t *node;
int i, rc;
struct hwloc_topology_support *support;
bool force_down = false;
hwloc_cpuset_t totalcpuset;
int bind_depth, map_depth;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: compute bindings for job %s with policy %s",
ORTE_JOBID_PRINT(jdata->jobid),
opal_hwloc_base_print_binding(jdata->map->binding));
map = ORTE_GET_MAPPING_POLICY(jdata->map->mapping);
bind = OPAL_GET_BINDING_POLICY(jdata->map->binding);
if (ORTE_MAPPING_BYUSER == map) {
/* user specified binding by rankfile - nothing for us to do */
return ORTE_SUCCESS;
}
if (OPAL_BIND_TO_CPUSET == bind) {
int rc;
/* cpuset was given - setup the bindings */
if (ORTE_SUCCESS != (rc = bind_to_cpuset(jdata))) {
ORTE_ERROR_LOG(rc);
}
return rc;
}
if (OPAL_BIND_TO_NONE == bind) {
/* no binding requested */
return ORTE_SUCCESS;
}
if (OPAL_BIND_TO_BOARD == bind) {
/* doesn't do anything at this time */
return ORTE_SUCCESS;
}
/* binding requested - convert the binding level to the hwloc obj type */
switch (bind) {
case OPAL_BIND_TO_NUMA:
hwb = HWLOC_OBJ_NODE;
break;
case OPAL_BIND_TO_SOCKET:
hwb = HWLOC_OBJ_SOCKET;
break;
case OPAL_BIND_TO_L3CACHE:
hwb = HWLOC_OBJ_CACHE;
clvl = 3;
break;
case OPAL_BIND_TO_L2CACHE:
hwb = HWLOC_OBJ_CACHE;
clvl = 2;
break;
case OPAL_BIND_TO_L1CACHE:
hwb = HWLOC_OBJ_CACHE;
clvl = 1;
break;
case OPAL_BIND_TO_CORE:
hwb = HWLOC_OBJ_CORE;
break;
case OPAL_BIND_TO_HWTHREAD:
hwb = HWLOC_OBJ_PU;
break;
default:
ORTE_ERROR_LOG(ORTE_ERR_BAD_PARAM);
return ORTE_ERR_BAD_PARAM;
}
/* do the same for the mapping policy */
switch (map) {
case ORTE_MAPPING_BYNODE:
case ORTE_MAPPING_BYSLOT:
case ORTE_MAPPING_SEQ:
hwm = HWLOC_OBJ_MACHINE;
break;
case ORTE_MAPPING_BYDIST:
case ORTE_MAPPING_BYNUMA:
hwm = HWLOC_OBJ_NODE;
break;
case ORTE_MAPPING_BYSOCKET:
hwm = HWLOC_OBJ_SOCKET;
break;
case ORTE_MAPPING_BYL3CACHE:
hwm = HWLOC_OBJ_CACHE;
clvm = 3;
break;
case ORTE_MAPPING_BYL2CACHE:
hwm = HWLOC_OBJ_CACHE;
clvm = 2;
break;
case ORTE_MAPPING_BYL1CACHE:
hwm = HWLOC_OBJ_CACHE;
clvm = 1;
break;
case ORTE_MAPPING_BYCORE:
hwm = HWLOC_OBJ_CORE;
break;
case ORTE_MAPPING_BYHWTHREAD:
hwm = HWLOC_OBJ_PU;
break;
default:
ORTE_ERROR_LOG(ORTE_ERR_BAD_PARAM);
return ORTE_ERR_BAD_PARAM;
}
/* if the job was mapped by the corresponding target, then
* we bind in place
*
* otherwise, we have to bind either up or down the hwloc
* tree. If we are binding upwards (e.g., mapped to hwthread
* but binding to core), then we just climb the tree to find
* the first matching object.
*
* if we are binding downwards (e.g., mapped to node and bind
* to core), then we have to do a round-robin assigment of
* procs to the resources below.
*/
if (ORTE_MAPPING_BYDIST == map) {
int rc = ORTE_SUCCESS;
if (OPAL_BIND_TO_NUMA == bind) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bindings for job %s - dist to numa",
ORTE_JOBID_PRINT(jdata->jobid));
if (ORTE_SUCCESS != (rc = bind_in_place(jdata, HWLOC_OBJ_NODE, 0))) {
ORTE_ERROR_LOG(rc);
}
} else if (OPAL_BIND_TO_NUMA < bind) {
/* bind every proc downwards */
force_down = true;
goto execute;
}
/* if the binding policy is less than numa, then we are unbound - so
* just ignore this and return (should have been caught in prior
* tests anyway as only options meeting that criteria are "none"
* and "board")
*/
return rc;
}
/* now deal with the remaining binding policies based on hardware */
if (bind == map) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bindings for job %s - bind in place",
ORTE_JOBID_PRINT(jdata->jobid));
if (ORTE_SUCCESS != (rc = bind_in_place(jdata, hwb, clvl))) {
ORTE_ERROR_LOG(rc);
}
return rc;
}
/* we need to handle the remaining binding options on a per-node
* basis because different nodes could potentially have different
* topologies, with different relative depths for the two levels
*/
execute:
/* initialize */
totalcpuset = hwloc_bitmap_alloc();
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;
}
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(jdata->map->binding) ||
!OPAL_BINDING_POLICY_IS_SET(jdata->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);
hwloc_bitmap_free(totalcpuset);
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(jdata->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);
hwloc_bitmap_free(totalcpuset);
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(jdata->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);
if (force_down) {
if (ORTE_SUCCESS != (rc = bind_downwards(jdata, node, hwb, clvl))) {
ORTE_ERROR_LOG(rc);
return rc;
}
} else {
/* determine the relative depth on this node */
if (HWLOC_OBJ_CACHE == hwb) {
/* must use a unique function because blasted hwloc
* just doesn't deal with caches very well...sigh
*/
bind_depth = hwloc_get_cache_type_depth(node->topology, clvl, -1);
} else {
bind_depth = hwloc_get_type_depth(node->topology, hwb);
}
if (0 > bind_depth) {
/* didn't find such an object */
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:no-objects",
true, hwloc_obj_type_string(hwb), node->name);
return ORTE_ERR_SILENT;
}
if (HWLOC_OBJ_CACHE == hwm) {
/* must use a unique function because blasted hwloc
* just doesn't deal with caches very well...sigh
*/
map_depth = hwloc_get_cache_type_depth(node->topology, clvm, -1);
} else {
map_depth = hwloc_get_type_depth(node->topology, hwm);
}
if (0 > map_depth) {
/* didn't find such an object */
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:no-objects",
true, hwloc_obj_type_string(hwm), node->name);
return ORTE_ERR_SILENT;
}
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s bind_depth: %d map_depth %d",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
bind_depth, map_depth);
if (bind_depth > map_depth) {
if (ORTE_SUCCESS != (rc = bind_downwards(jdata, node, hwb, clvl))) {
ORTE_ERROR_LOG(rc);
return rc;
}
} else {
if (ORTE_SUCCESS != (rc = bind_upwards(jdata, node, hwb, clvl))) {
ORTE_ERROR_LOG(rc);
return rc;
}
}
}
}
return ORTE_SUCCESS;
}
/*
* 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);
}
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;
}
/*********************************
* 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;
}
