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;
}
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;
}
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;
}
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;
}
/*
* JOB_MAP
*/
int orte_dt_print_map(char **output, char *prefix, orte_job_map_t *src, opal_data_type_t type)
{
char *tmp=NULL, *tmp2, *tmp3, *pfx, *pfx2;
int32_t i, j;
int rc;
orte_node_t *node;
orte_proc_t *proc;
/* set default result */
*output = NULL;
/* protect against NULL prefix */
if (NULL == prefix) {
asprintf(&pfx2, " ");
} else {
asprintf(&pfx2, "%s", prefix);
}
if (orte_xml_output) {
/* need to create the output in XML format */
asprintf(&tmp, "<map>\n");
/* loop through nodes */
for (i=0; i < src->nodes->size; i++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(src->nodes, i))) {
continue;
}
orte_dt_print_node(&tmp2, "\t", node, ORTE_NODE);
asprintf(&tmp3, "%s%s", tmp, tmp2);
free(tmp2);
free(tmp);
tmp = tmp3;
/* for each node, loop through procs and print their rank */
for (j=0; j < node->procs->size; j++) {
if (NULL == (proc = (orte_proc_t*)opal_pointer_array_get_item(node->procs, j))) {
continue;
}
orte_dt_print_proc(&tmp2, "\t\t", proc, ORTE_PROC);
asprintf(&tmp3, "%s%s", tmp, tmp2);
free(tmp2);
free(tmp);
tmp = tmp3;
}
asprintf(&tmp3, "%s\t</host>\n", tmp);
free(tmp);
tmp = tmp3;
}
asprintf(&tmp2, "%s</map>\n", tmp);
free(tmp);
free(pfx2);
*output = tmp2;
return ORTE_SUCCESS;
}
asprintf(&pfx, "%s\t", pfx2);
if (orte_devel_level_output) {
#if OPAL_HAVE_HWLOC
asprintf(&tmp, "\n%sMapper requested: %s Last mapper: %s Mapping policy: %s Ranking policy: %s\n%sBinding policy: %s Cpu set: %s PPR: %s Cpus-per-rank: %d",
pfx2, (NULL == src->req_mapper) ? "NULL" : src->req_mapper,
(NULL == src->last_mapper) ? "NULL" : src->last_mapper,
orte_rmaps_base_print_mapping(src->mapping),
orte_rmaps_base_print_ranking(src->ranking),
pfx2, opal_hwloc_base_print_binding(src->binding),
(NULL == opal_hwloc_base_cpu_set) ? "NULL" : opal_hwloc_base_cpu_set,
(NULL == src->ppr) ? "NULL" : src->ppr,
(int)src->cpus_per_rank);
#else
asprintf(&tmp, "\n%sMapper requested: %s Last mapper: %s Mapping policy: %s Ranking policy: %s PPR: %s Cpus-per-rank: %d",
pfx2, (NULL == src->req_mapper) ? "NULL" : src->req_mapper,
(NULL == src->last_mapper) ? "NULL" : src->last_mapper,
orte_rmaps_base_print_mapping(src->mapping),
orte_rmaps_base_print_ranking(src->ranking),
(NULL == src->ppr) ? "NULL" : src->ppr,
(int)src->cpus_per_rank);
#endif
if (ORTE_VPID_INVALID == src->daemon_vpid_start) {
asprintf(&tmp2, "%s\n%sNum new daemons: %ld\tNew daemon starting vpid INVALID\n%sNum nodes: %ld",
tmp, pfx, (long)src->num_new_daemons, pfx, (long)src->num_nodes);
} else {
asprintf(&tmp2, "%s\n%sNum new daemons: %ld\tNew daemon starting vpid %ld\n%sNum nodes: %ld",
tmp, pfx, (long)src->num_new_daemons, (long)src->daemon_vpid_start,
pfx, (long)src->num_nodes);
}
free(tmp);
tmp = tmp2;
} else {
/* this is being printed for a user, so let's make it easier to see */
asprintf(&tmp, "\n%s======================== JOB MAP ========================", pfx2);
}
for (i=0; i < src->nodes->size; i++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(src->nodes, i))) {
continue;
}
if (ORTE_SUCCESS != (rc = opal_dss.print(&tmp2, pfx2, node, ORTE_NODE))) {
ORTE_ERROR_LOG(rc);
free(pfx);
free(tmp);
return rc;
}
asprintf(&tmp3, "%s\n%s", tmp, tmp2);
free(tmp);
free(tmp2);
tmp = tmp3;
}
if (!orte_devel_level_output) {
/* this is being printed for a user, so let's make it easier to see */
asprintf(&tmp2, "%s\n\n%s=============================================================\n", tmp, pfx2);
free(tmp);
tmp = tmp2;
}
free(pfx2);
/* set the return */
*output = tmp;
free(pfx);
return ORTE_SUCCESS;
}
static int bind_upwards(orte_job_t *jdata,
orte_node_t *node,
hwloc_obj_type_t target,
unsigned cache_level)
{
/* traverse the hwloc topology tree on each node upwards
* until we find an object of type target - and then bind
* the process to that target
*/
int j;
orte_job_map_t *map;
orte_proc_t *proc;
hwloc_obj_t obj;
hwloc_cpuset_t cpus;
unsigned int idx, ncpus;
opal_hwloc_obj_data_t *data;
hwloc_obj_t locale;
char *cpu_bitmap;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bind upwards for job %s with bindings %s",
ORTE_JOBID_PRINT(jdata->jobid),
opal_hwloc_base_print_binding(jdata->map->binding));
/* initialize */
map = jdata->map;
/* cycle thru the procs */
for (j=0; j < node->procs->size; j++) {
if (NULL == (proc = (orte_proc_t*)opal_pointer_array_get_item(node->procs, j))) {
continue;
}
/* ignore procs from other jobs */
if (proc->name.jobid != jdata->jobid) {
continue;
}
/* bozo check */
if (!orte_get_attribute(&proc->attributes, ORTE_PROC_HWLOC_LOCALE, (void**)&locale, OPAL_PTR)) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-locale", true, ORTE_NAME_PRINT(&proc->name));
return ORTE_ERR_SILENT;
}
/* starting at the locale, move up thru the parents
* to find the target object type
*/
cpu_bitmap = NULL;
for (obj = locale->parent; NULL != obj; obj = obj->parent) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s bind:upward target %s type %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
hwloc_obj_type_string(target),
hwloc_obj_type_string(obj->type));
if (target == obj->type) {
if (HWLOC_OBJ_CACHE == target && cache_level != obj->attr->cache.depth) {
continue;
}
/* get its index */
if (UINT_MAX == (idx = opal_hwloc_base_get_obj_idx(node->topology, obj, OPAL_HWLOC_AVAILABLE))) {
ORTE_ERROR_LOG(ORTE_ERR_BAD_PARAM);
return ORTE_ERR_SILENT;
}
/* track the number bound */
data = (opal_hwloc_obj_data_t*)obj->userdata;
data->num_bound++;
/* get the number of cpus under this location */
if (0 == (ncpus = opal_hwloc_base_get_npus(node->topology, obj))) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-available-cpus", true, node->name);
return ORTE_ERR_SILENT;
}
/* error out if adding a proc would cause overload and that wasn't allowed,
* and it wasn't a default binding policy (i.e., the user requested it)
*/
if (ncpus < data->num_bound &&
!OPAL_BIND_OVERLOAD_ALLOWED(jdata->map->binding)) {
if (OPAL_BINDING_POLICY_IS_SET(jdata->map->binding)) {
/* if the user specified a binding policy, then we cannot meet
* it since overload isn't allowed, so error out - have the
* message indicate that setting overload allowed will remove
* this restriction */
orte_show_help("help-orte-rmaps-base.txt", "rmaps:binding-overload", true,
opal_hwloc_base_print_binding(map->binding), node->name,
data->num_bound, ncpus);
return ORTE_ERR_SILENT;
} else {
/* if we have the default binding policy, then just don't bind */
OPAL_SET_BINDING_POLICY(map->binding, OPAL_BIND_TO_NONE);
unbind_procs(jdata);
return ORTE_SUCCESS;
}
}
/* bind it here */
cpus = opal_hwloc_base_get_available_cpus(node->topology, obj);
hwloc_bitmap_list_asprintf(&cpu_bitmap, cpus);
orte_set_attribute(&proc->attributes, ORTE_PROC_CPU_BITMAP, ORTE_ATTR_GLOBAL, cpu_bitmap, OPAL_STRING);
/* record the location */
orte_set_attribute(&proc->attributes, ORTE_PROC_HWLOC_BOUND, ORTE_ATTR_LOCAL, obj, OPAL_PTR);
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s BOUND PROC %s TO %s[%s:%u] on node %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&proc->name),
cpu_bitmap,
hwloc_obj_type_string(target),
idx, node->name);
break;
}
}
if (NULL == cpu_bitmap && OPAL_BINDING_REQUIRED(jdata->map->binding)) {
/* didn't find anyone to bind to - this is an error
* unless the user specified if-supported
*/
orte_show_help("help-orte-rmaps-base.txt", "rmaps:binding-target-not-found", true,
opal_hwloc_base_print_binding(map->binding), node->name);
return ORTE_ERR_SILENT;
}
if (NULL != cpu_bitmap) {
free(cpu_bitmap);
}
}
return ORTE_SUCCESS;
}
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;
}
static int bind_upwards(orte_job_t *jdata,
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 i, j;
orte_job_map_t *map;
orte_node_t *node;
orte_proc_t *proc;
hwloc_obj_t obj;
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 upwards 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(opal_hwloc_binding_policy)) {
/* 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
*/
if (!support->membind->set_thisproc_membind &&
!support->membind->set_thisthread_membind) {
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;
}
}
}
/* clear the topology of any prior usage numbers */
opal_hwloc_base_clear_usage(node->topology);
/* 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(0, "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(0, "%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))) {
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 */
if (ncpus < data->num_bound &&
!OPAL_BIND_OVERLOAD_ALLOWED(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);
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;
}
static int bind_downwards(orte_job_t *jdata,
hwloc_obj_type_t target,
unsigned cache_level)
{
int i, j;
orte_job_map_t *map;
orte_node_t *node;
orte_proc_t *proc;
hwloc_obj_t trg_obj, nxt_obj;
hwloc_cpuset_t cpus;
unsigned int ncpus;
struct hwloc_topology_support *support;
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();
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(opal_hwloc_binding_policy)) {
/* 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
*/
if (!support->membind->set_thisproc_membind &&
!support->membind->set_thisthread_membind) {
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;
}
}
}
/* clear the topology of any prior usage numbers */
opal_hwloc_base_clear_usage(node->topology);
/* 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;
}
/* 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 */
if (ncpus < data->num_bound &&
!OPAL_BIND_OVERLOAD_ALLOWED(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;
}
/* bind the proc here */
cpus = opal_hwloc_base_get_available_cpus(node->topology, trg_obj);
hwloc_bitmap_or(totalcpuset, totalcpuset, 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];
opal_hwloc_base_cset2str(tmp1, sizeof(tmp1), totalcpuset);
opal_hwloc_base_cset2mapstr(tmp2, sizeof(tmp2), 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;
}
/* stuff proc attributes for sending back to a proc */
int orte_pmix_server_register_nspace(orte_job_t *jdata, bool force)
{
int rc;
orte_proc_t *pptr;
int i, k, n;
opal_list_t *info, *pmap;
opal_value_t *kv;
orte_node_t *node, *mynode;
opal_vpid_t vpid;
char **list, **procs, **micro, *tmp, *regex;
orte_job_t *dmns;
orte_job_map_t *map;
orte_app_context_t *app;
uid_t uid;
gid_t gid;
opal_list_t *cache;
hwloc_obj_t machine;
opal_output_verbose(2, orte_pmix_server_globals.output,
"%s register nspace for %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_JOBID_PRINT(jdata->jobid));
/* setup the info list */
info = OBJ_NEW(opal_list_t);
uid = geteuid();
gid = getegid();
/* pass our nspace/rank */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_SERVER_NSPACE);
kv->data.string = strdup(ORTE_JOBID_PRINT(ORTE_PROC_MY_NAME->jobid));
kv->type = OPAL_STRING;
opal_list_append(info, &kv->super);
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_SERVER_RANK);
kv->data.uint32 = ORTE_PROC_MY_NAME->vpid;
kv->type = OPAL_UINT32;
opal_list_append(info, &kv->super);
/* jobid */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_JOBID);
kv->data.string = strdup(ORTE_JOBID_PRINT(jdata->jobid));
kv->type = OPAL_STRING;
opal_list_append(info, &kv->super);
/* offset */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_NPROC_OFFSET);
kv->data.uint32 = jdata->offset;
kv->type = OPAL_UINT32;
opal_list_append(info, &kv->super);
/* check for cached values to add to the job info */
cache = NULL;
if (orte_get_attribute(&jdata->attributes, ORTE_JOB_INFO_CACHE, (void**)&cache, OPAL_PTR) &&
NULL != cache) {
while (NULL != (kv = (opal_value_t*)opal_list_remove_first(cache))) {
opal_list_append(info, &kv->super);
}
orte_remove_attribute(&jdata->attributes, ORTE_JOB_INFO_CACHE);
OBJ_RELEASE(cache);
}
/* assemble the node and proc map info */
list = NULL;
procs = NULL;
map = jdata->map;
for (i=0; i < map->nodes->size; i++) {
micro = NULL;
if (NULL != (node = (orte_node_t*)opal_pointer_array_get_item(map->nodes, i))) {
opal_argv_append_nosize(&list, node->name);
/* assemble all the ranks for this job that are on this node */
for (k=0; k < node->procs->size; k++) {
if (NULL != (pptr = (orte_proc_t*)opal_pointer_array_get_item(node->procs, k))) {
if (jdata->jobid == pptr->name.jobid) {
opal_argv_append_nosize(µ, ORTE_VPID_PRINT(pptr->name.vpid));
}
}
}
/* assemble the rank/node map */
if (NULL != micro) {
tmp = opal_argv_join(micro, ',');
opal_argv_free(micro);
opal_argv_append_nosize(&procs, tmp);
free(tmp);
}
}
}
/* let the PMIx server generate the nodemap regex */
if (NULL != list) {
tmp = opal_argv_join(list, ',');
opal_argv_free(list);
list = NULL;
if (OPAL_SUCCESS != (rc = opal_pmix.generate_regex(tmp, ®ex))) {
ORTE_ERROR_LOG(rc);
free(tmp);
OPAL_LIST_RELEASE(info);
return rc;
}
free(tmp);
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_NODE_MAP);
kv->type = OPAL_STRING;
kv->data.string = regex;
opal_list_append(info, &kv->super);
}
/* let the PMIx server generate the procmap regex */
if (NULL != procs) {
tmp = opal_argv_join(procs, ';');
opal_argv_free(procs);
procs = NULL;
if (OPAL_SUCCESS != (rc = opal_pmix.generate_ppn(tmp, ®ex))) {
ORTE_ERROR_LOG(rc);
free(tmp);
OPAL_LIST_RELEASE(info);
return rc;
}
free(tmp);
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_PROC_MAP);
kv->type = OPAL_STRING;
kv->data.string = regex;
opal_list_append(info, &kv->super);
}
/* get our local node */
if (NULL == (dmns = orte_get_job_data_object(ORTE_PROC_MY_NAME->jobid))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
OPAL_LIST_RELEASE(info);
return ORTE_ERR_NOT_FOUND;
}
if (NULL == (pptr = (orte_proc_t*)opal_pointer_array_get_item(dmns->procs, ORTE_PROC_MY_NAME->vpid))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
OPAL_LIST_RELEASE(info);
return ORTE_ERR_NOT_FOUND;
}
mynode = pptr->node;
if (NULL == mynode) {
/* cannot happen */
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
OPAL_LIST_RELEASE(info);
return ORTE_ERR_NOT_FOUND;
}
/* pass our node ID */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_NODEID);
kv->type = OPAL_UINT32;
kv->data.uint32 = mynode->index;
opal_list_append(info, &kv->super);
/* pass our node size */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_NODE_SIZE);
kv->type = OPAL_UINT32;
kv->data.uint32 = mynode->num_procs;
opal_list_append(info, &kv->super);
/* pass the number of nodes in the job */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_NUM_NODES);
kv->type = OPAL_UINT32;
kv->data.uint32 = map->num_nodes;
opal_list_append(info, &kv->super);
/* univ size */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_UNIV_SIZE);
kv->type = OPAL_UINT32;
kv->data.uint32 = jdata->total_slots_alloc;
opal_list_append(info, &kv->super);
/* job size */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_JOB_SIZE);
kv->type = OPAL_UINT32;
kv->data.uint32 = jdata->num_procs;
opal_list_append(info, &kv->super);
/* number of apps in this job */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_JOB_NUM_APPS);
kv->type = OPAL_UINT32;
kv->data.uint32 = jdata->num_apps;
opal_list_append(info, &kv->super);
/* local size */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_LOCAL_SIZE);
kv->type = OPAL_UINT32;
kv->data.uint32 = jdata->num_local_procs;
opal_list_append(info, &kv->super);
/* max procs */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_MAX_PROCS);
kv->type = OPAL_UINT32;
kv->data.uint32 = jdata->total_slots_alloc;
opal_list_append(info, &kv->super);
/* topology signature */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_TOPOLOGY_SIGNATURE);
kv->type = OPAL_STRING;
kv->data.string = strdup(orte_topo_signature);
opal_list_append(info, &kv->super);
/* total available physical memory */
machine = hwloc_get_next_obj_by_type (opal_hwloc_topology, HWLOC_OBJ_MACHINE, NULL);
if (NULL != machine) {
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_AVAIL_PHYS_MEMORY);
kv->type = OPAL_UINT64;
#if HWLOC_API_VERSION < 0x20000
kv->data.uint64 = machine->memory.total_memory;
#else
kv->data.uint64 = machine->total_memory;
#endif
opal_list_append(info, &kv->super);
}
/* pass the mapping policy used for this job */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_MAPBY);
kv->type = OPAL_STRING;
kv->data.string = strdup(orte_rmaps_base_print_mapping(jdata->map->mapping));
opal_list_append(info, &kv->super);
/* pass the ranking policy used for this job */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_RANKBY);
kv->type = OPAL_STRING;
kv->data.string = strdup(orte_rmaps_base_print_ranking(jdata->map->ranking));
opal_list_append(info, &kv->super);
/* pass the binding policy used for this job */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_BINDTO);
kv->type = OPAL_STRING;
kv->data.string = strdup(opal_hwloc_base_print_binding(jdata->map->binding));
opal_list_append(info, &kv->super);
/* register any local clients */
vpid = ORTE_VPID_MAX;
micro = NULL;
for (i=0; i < mynode->procs->size; i++) {
if (NULL == (pptr = (orte_proc_t*)opal_pointer_array_get_item(mynode->procs, i))) {
continue;
}
if (pptr->name.jobid == jdata->jobid) {
opal_argv_append_nosize(µ, ORTE_VPID_PRINT(pptr->name.vpid));
if (pptr->name.vpid < vpid) {
vpid = pptr->name.vpid;
}
/* go ahead and register this client */
if (OPAL_SUCCESS != (rc = opal_pmix.server_register_client(&pptr->name, uid, gid,
(void*)pptr, NULL, NULL))) {
ORTE_ERROR_LOG(rc);
}
}
}
if (NULL != micro) {
/* pass the local peers */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_LOCAL_PEERS);
kv->type = OPAL_STRING;
kv->data.string = opal_argv_join(micro, ',');
opal_argv_free(micro);
opal_list_append(info, &kv->super);
}
/* pass the local ldr */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_LOCALLDR);
kv->type = OPAL_VPID;
kv->data.name.vpid = vpid;
opal_list_append(info, &kv->super);
/* for each proc in this job, create an object that
* includes the info describing the proc so the recipient has a complete
* picture. This allows procs to connect to each other without
* any further info exchange, assuming the underlying transports
* support it. We also pass all the proc-specific data here so
* that each proc can lookup info about every other proc in the job */
for (n=0; n < map->nodes->size; n++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(map->nodes, n))) {
continue;
}
/* cycle across each proc on this node, passing all data that
* varies by proc */
for (i=0; i < node->procs->size; i++) {
if (NULL == (pptr = (orte_proc_t*)opal_pointer_array_get_item(node->procs, i))) {
continue;
}
/* only consider procs from this job */
if (pptr->name.jobid != jdata->jobid) {
continue;
}
/* setup the proc map object */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_PROC_DATA);
kv->type = OPAL_PTR;
kv->data.ptr = OBJ_NEW(opal_list_t);
opal_list_append(info, &kv->super);
pmap = kv->data.ptr;
/* must start with rank */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_RANK);
kv->type = OPAL_VPID;
kv->data.name.vpid = pptr->name.vpid;
opal_list_append(pmap, &kv->super);
/* location, for local procs */
if (node == mynode) {
tmp = NULL;
if (orte_get_attribute(&pptr->attributes, ORTE_PROC_CPU_BITMAP, (void**)&tmp, OPAL_STRING) &&
NULL != tmp) {
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_LOCALITY_STRING);
kv->type = OPAL_STRING;
kv->data.string = opal_hwloc_base_get_locality_string(opal_hwloc_topology, tmp);
opal_list_append(pmap, &kv->super);
free(tmp);
} else {
/* the proc is not bound */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_LOCALITY_STRING);
kv->type = OPAL_STRING;
kv->data.string = NULL;
opal_list_append(pmap, &kv->super);
}
}
/* global/univ rank */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_GLOBAL_RANK);
kv->type = OPAL_VPID;
kv->data.name.vpid = pptr->name.vpid + jdata->offset;
opal_list_append(pmap, &kv->super);
if (1 < jdata->num_apps) {
/* appnum */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_APPNUM);
kv->type = OPAL_UINT32;
kv->data.uint32 = pptr->app_idx;
opal_list_append(pmap, &kv->super);
/* app ldr */
app = (orte_app_context_t*)opal_pointer_array_get_item(jdata->apps, pptr->app_idx);
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_APPLDR);
kv->type = OPAL_VPID;
kv->data.name.vpid = app->first_rank;
opal_list_append(pmap, &kv->super);
/* app rank */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_APP_RANK);
kv->type = OPAL_VPID;
kv->data.name.vpid = pptr->app_rank;
opal_list_append(pmap, &kv->super);
/* app size */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_APP_SIZE);
kv->type = OPAL_UINT32;
kv->data.uint32 = app->num_procs;
opal_list_append(info, &kv->super);
}
/* local rank */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_LOCAL_RANK);
kv->type = OPAL_UINT16;
kv->data.uint16 = pptr->local_rank;
opal_list_append(pmap, &kv->super);
/* node rank */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_NODE_RANK);
kv->type = OPAL_UINT16;
kv->data.uint32 = pptr->node_rank;
opal_list_append(pmap, &kv->super);
/* node ID */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_NODEID);
kv->type = OPAL_UINT32;
kv->data.uint32 = pptr->node->index;
opal_list_append(pmap, &kv->super);
if (map->num_nodes < orte_hostname_cutoff) {
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_HOSTNAME);
kv->type = OPAL_STRING;
kv->data.string = strdup(pptr->node->name);
opal_list_append(pmap, &kv->super);
}
}
}
/* mark the job as registered */
orte_set_attribute(&jdata->attributes, ORTE_JOB_NSPACE_REGISTERED, ORTE_ATTR_LOCAL, NULL, OPAL_BOOL);
/* pass it down */
/* we are in an event, so no need to callback */
rc = opal_pmix.server_register_nspace(jdata->jobid,
jdata->num_local_procs,
info, NULL, NULL);
OPAL_LIST_RELEASE(info);
/* if the user has connected us to an external server, then we must
* assume there is going to be some cross-mpirun exchange, and so
* we protect against that situation by publishing the job info
* for this job - this allows any subsequent "connect" to retrieve
* the job info */
if (NULL != orte_data_server_uri) {
opal_buffer_t buf;
OBJ_CONSTRUCT(&buf, opal_buffer_t);
if (OPAL_SUCCESS != (rc = opal_dss.pack(&buf, &jdata, 1, ORTE_JOB))) {
ORTE_ERROR_LOG(rc);
OBJ_DESTRUCT(&buf);
return rc;
}
info = OBJ_NEW(opal_list_t);
/* create a key-value with the key being the string jobid
* and the value being the byte object */
kv = OBJ_NEW(opal_value_t);
orte_util_convert_jobid_to_string(&kv->key, jdata->jobid);
kv->type = OPAL_BYTE_OBJECT;
opal_dss.unload(&buf, (void**)&kv->data.bo.bytes, &kv->data.bo.size);
OBJ_DESTRUCT(&buf);
opal_list_append(info, &kv->super);
/* set the range to be session */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_RANGE);
kv->type = OPAL_UINT;
kv->data.uint = OPAL_PMIX_RANGE_SESSION;
opal_list_append(info, &kv->super);
/* set the persistence to be app */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_PERSISTENCE);
kv->type = OPAL_INT;
kv->data.integer = OPAL_PMIX_PERSIST_APP;
opal_list_append(info, &kv->super);
/* add our effective userid to the directives */
kv = OBJ_NEW(opal_value_t);
kv->key = strdup(OPAL_PMIX_USERID);
kv->type = OPAL_UINT32;
kv->data.uint32 = geteuid();
opal_list_append(info, &kv->super);
/* now publish it */
if (ORTE_SUCCESS != (rc = pmix_server_publish_fn(ORTE_PROC_MY_NAME,
info, mycbfunc, info))) {
ORTE_ERROR_LOG(rc);
}
}
return rc;
}
