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;
}
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_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;
}
예제 #4
0
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
        }
    }
예제 #5
0
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;
}
예제 #6
0
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_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;
}