/* * lllp_distribution * * Note: lllp stands for Lowest Level of Logical Processors. * * When automatic binding is enabled: * - no binding flags set >= CPU_BIND_NONE, and * - a auto binding level selected CPU_BIND_TO_{SOCKETS,CORES,THREADS} * Otherwise limit job step to the allocated CPUs * * generate the appropriate cpu_bind type and string which results in * the specified lllp distribution. * * IN/OUT- job launch request (cpu_bind_type and cpu_bind updated) * IN- global task id array */ void lllp_distribution(launch_tasks_request_msg_t *req, uint32_t node_id) { int rc = SLURM_SUCCESS; bitstr_t **masks = NULL; char buf_type[100]; int maxtasks = req->tasks_to_launch[(int)node_id]; int whole_nodes, whole_sockets, whole_cores, whole_threads; int part_sockets, part_cores; const uint32_t *gtid = req->global_task_ids[(int)node_id]; static uint16_t bind_entity = CPU_BIND_TO_THREADS | CPU_BIND_TO_CORES | CPU_BIND_TO_SOCKETS | CPU_BIND_TO_LDOMS; static uint16_t bind_mode = CPU_BIND_NONE | CPU_BIND_MASK | CPU_BIND_RANK | CPU_BIND_MAP | CPU_BIND_LDMASK | CPU_BIND_LDRANK | CPU_BIND_LDMAP; static int only_one_thread_per_core = -1; if (only_one_thread_per_core == -1) { if (conf->cpus == (conf->sockets * conf->cores)) only_one_thread_per_core = 1; else only_one_thread_per_core = 0; } /* If we are telling the system we only want to use 1 thread * per core with the CPUs node option this is the easiest way * to portray that to the affinity plugin. */ if (only_one_thread_per_core) req->cpu_bind_type |= CPU_BIND_ONE_THREAD_PER_CORE; if (req->cpu_bind_type & bind_mode) { /* Explicit step binding specified by user */ char *avail_mask = _alloc_mask(req, &whole_nodes, &whole_sockets, &whole_cores, &whole_threads, &part_sockets, &part_cores); if ((whole_nodes == 0) && avail_mask && (req->job_core_spec == (uint16_t) NO_VAL)) { info("task/affinity: entire node must be allocated, " "disabling affinity"); xfree(req->cpu_bind); req->cpu_bind = avail_mask; req->cpu_bind_type &= (~bind_mode); req->cpu_bind_type |= CPU_BIND_MASK; } else { if (req->job_core_spec == (uint16_t) NO_VAL) { if (req->cpu_bind_type & CPU_BIND_MASK) _validate_mask(req, avail_mask); else if (req->cpu_bind_type & CPU_BIND_MAP) _validate_map(req, avail_mask); } xfree(avail_mask); } slurm_sprint_cpu_bind_type(buf_type, req->cpu_bind_type); info("lllp_distribution jobid [%u] manual binding: %s", req->job_id, buf_type); return; } if (!(req->cpu_bind_type & bind_entity)) { /* No bind unit (sockets, cores) specified by user, * pick something reasonable */ uint32_t task_plugin_param = slurm_get_task_plugin_param(); bool auto_def_set = false; int spec_thread_cnt = 0; int max_tasks = req->tasks_to_launch[(int)node_id] * req->cpus_per_task; char *avail_mask = _alloc_mask(req, &whole_nodes, &whole_sockets, &whole_cores, &whole_threads, &part_sockets, &part_cores); debug("binding tasks:%d to " "nodes:%d sockets:%d:%d cores:%d:%d threads:%d", max_tasks, whole_nodes, whole_sockets ,part_sockets, whole_cores, part_cores, whole_threads); if ((req->job_core_spec != (uint16_t) NO_VAL) && (req->job_core_spec & CORE_SPEC_THREAD) && (req->job_core_spec != CORE_SPEC_THREAD)) { spec_thread_cnt = req->job_core_spec & (~CORE_SPEC_THREAD); } if (((max_tasks == whole_sockets) && (part_sockets == 0)) || (spec_thread_cnt && (max_tasks == (whole_sockets + part_sockets)))) { req->cpu_bind_type |= CPU_BIND_TO_SOCKETS; goto make_auto; } if (((max_tasks == whole_cores) && (part_cores == 0)) || (spec_thread_cnt && (max_tasks == (whole_cores + part_cores)))) { req->cpu_bind_type |= CPU_BIND_TO_CORES; goto make_auto; } if (max_tasks == whole_threads) { req->cpu_bind_type |= CPU_BIND_TO_THREADS; goto make_auto; } if (task_plugin_param & CPU_AUTO_BIND_TO_THREADS) { auto_def_set = true; req->cpu_bind_type |= CPU_BIND_TO_THREADS; goto make_auto; } else if (task_plugin_param & CPU_AUTO_BIND_TO_CORES) { auto_def_set = true; req->cpu_bind_type |= CPU_BIND_TO_CORES; goto make_auto; } else if (task_plugin_param & CPU_AUTO_BIND_TO_SOCKETS) { auto_def_set = true; req->cpu_bind_type |= CPU_BIND_TO_SOCKETS; goto make_auto; } if (avail_mask) { xfree(req->cpu_bind); req->cpu_bind = avail_mask; req->cpu_bind_type |= CPU_BIND_MASK; } slurm_sprint_cpu_bind_type(buf_type, req->cpu_bind_type); info("lllp_distribution jobid [%u] auto binding off: %s", req->job_id, buf_type); return; make_auto: xfree(avail_mask); slurm_sprint_cpu_bind_type(buf_type, req->cpu_bind_type); info("lllp_distribution jobid [%u] %s auto binding: " "%s, dist %d", req->job_id, (auto_def_set) ? "default" : "implicit", buf_type, req->task_dist); } else { /* Explicit bind unit (sockets, cores) specified by user */ slurm_sprint_cpu_bind_type(buf_type, req->cpu_bind_type); info("lllp_distribution jobid [%u] binding: %s, dist %d", req->job_id, buf_type, req->task_dist); } switch (req->task_dist & SLURM_DIST_STATE_BASE) { case SLURM_DIST_BLOCK_BLOCK: case SLURM_DIST_CYCLIC_BLOCK: case SLURM_DIST_PLANE: /* tasks are distributed in blocks within a plane */ rc = _task_layout_lllp_block(req, node_id, &masks); break; case SLURM_DIST_ARBITRARY: case SLURM_DIST_BLOCK: case SLURM_DIST_CYCLIC: case SLURM_DIST_UNKNOWN: if (slurm_get_select_type_param() & CR_CORE_DEFAULT_DIST_BLOCK) { rc = _task_layout_lllp_block(req, node_id, &masks); break; } /* We want to fall through here if we aren't doing a default dist block. */ default: rc = _task_layout_lllp_cyclic(req, node_id, &masks); break; } /* FIXME: I'm worried about core_bitmap with CPU_BIND_TO_SOCKETS & * max_cores - does select/cons_res plugin allocate whole * socket??? Maybe not. Check srun man page. */ if (rc == SLURM_SUCCESS) { _task_layout_display_masks(req, gtid, maxtasks, masks); /* translate abstract masks to actual hardware layout */ _lllp_map_abstract_masks(maxtasks, masks); _task_layout_display_masks(req, gtid, maxtasks, masks); #ifdef HAVE_NUMA if (req->cpu_bind_type & CPU_BIND_TO_LDOMS) { _match_masks_to_ldom(maxtasks, masks); _task_layout_display_masks(req, gtid, maxtasks, masks); } #endif /* convert masks into cpu_bind mask string */ _lllp_generate_cpu_bind(req, maxtasks, masks); } else { char *avail_mask = _alloc_mask(req, &whole_nodes, &whole_sockets, &whole_cores, &whole_threads, &part_sockets, &part_cores); if (avail_mask) { xfree(req->cpu_bind); req->cpu_bind = avail_mask; req->cpu_bind_type &= (~bind_mode); req->cpu_bind_type |= CPU_BIND_MASK; } slurm_sprint_cpu_bind_type(buf_type, req->cpu_bind_type); error("lllp_distribution jobid [%u] overriding binding: %s", req->job_id, buf_type); error("Verify socket/core/thread counts in configuration"); } if (masks) _lllp_free_masks(maxtasks, masks); }
/* * lllp_distribution * * Note: lllp stands for Lowest Level of Logical Processors. * * When automatic binding is enabled: * - no binding flags set >= CPU_BIND_NONE, and * - a auto binding level selected CPU_BIND_TO_{SOCKETS,CORES,THREADS} * Otherwise limit job step to the allocated CPUs * * generate the appropriate cpu_bind type and string which results in * the specified lllp distribution. * * IN/OUT- job launch request (cpu_bind_type and cpu_bind updated) * IN- global task id array */ void lllp_distribution(launch_tasks_request_msg_t *req, uint32_t node_id) { int rc = SLURM_SUCCESS; bitstr_t **masks = NULL; char buf_type[100]; int maxtasks = req->tasks_to_launch[(int)node_id]; int whole_nodes, whole_sockets, whole_cores, whole_threads; int part_sockets, part_cores; const uint32_t *gtid = req->global_task_ids[(int)node_id]; static uint16_t bind_entity = CPU_BIND_TO_THREADS | CPU_BIND_TO_CORES | CPU_BIND_TO_SOCKETS | CPU_BIND_TO_LDOMS; static uint16_t bind_mode = CPU_BIND_NONE | CPU_BIND_MASK | CPU_BIND_RANK | CPU_BIND_MAP | CPU_BIND_LDMASK | CPU_BIND_LDRANK | CPU_BIND_LDMAP; if (req->cpu_bind_type & bind_mode) { /* Explicit step binding specified by user */ char *avail_mask = _alloc_mask(req, &whole_nodes, &whole_sockets, &whole_cores, &whole_threads, &part_sockets, &part_cores); if ((whole_nodes == 0) && avail_mask) { /* Step does NOT have access to whole node, * bind to full mask of available processors */ xfree(req->cpu_bind); req->cpu_bind = avail_mask; req->cpu_bind_type &= (~bind_mode); req->cpu_bind_type |= CPU_BIND_MASK; } else { /* Step does have access to whole node, * bind to whatever step wants */ xfree(avail_mask); } slurm_sprint_cpu_bind_type(buf_type, req->cpu_bind_type); info("lllp_distribution jobid [%u] manual binding: %s", req->job_id, buf_type); return; } if (!(req->cpu_bind_type & bind_entity)) { /* No bind unit (sockets, cores) specified by user, * pick something reasonable */ int max_tasks = req->tasks_to_launch[(int)node_id]; char *avail_mask = _alloc_mask(req, &whole_nodes, &whole_sockets, &whole_cores, &whole_threads, &part_sockets, &part_cores); debug("binding tasks:%d to " "nodes:%d sockets:%d:%d cores:%d:%d threads:%d", max_tasks, whole_nodes, whole_sockets ,part_sockets, whole_cores, part_cores, whole_threads); if ((max_tasks == whole_sockets) && (part_sockets == 0)) { req->cpu_bind_type |= CPU_BIND_TO_SOCKETS; goto make_auto; } if ((max_tasks == whole_cores) && (part_cores == 0)) { req->cpu_bind_type |= CPU_BIND_TO_CORES; goto make_auto; } if (max_tasks == whole_threads) { req->cpu_bind_type |= CPU_BIND_TO_THREADS; goto make_auto; } if (avail_mask) { xfree(req->cpu_bind); req->cpu_bind = avail_mask; req->cpu_bind_type |= CPU_BIND_MASK; } slurm_sprint_cpu_bind_type(buf_type, req->cpu_bind_type); info("lllp_distribution jobid [%u] auto binding off: %s", req->job_id, buf_type); return; make_auto: xfree(avail_mask); slurm_sprint_cpu_bind_type(buf_type, req->cpu_bind_type); info("lllp_distribution jobid [%u] implicit auto binding: " "%s, dist %d", req->job_id, buf_type, req->task_dist); } else { /* Explicit bind unit (sockets, cores) specified by user */ slurm_sprint_cpu_bind_type(buf_type, req->cpu_bind_type); info("lllp_distribution jobid [%u] binding: %s, dist %d", req->job_id, buf_type, req->task_dist); } switch (req->task_dist) { case SLURM_DIST_BLOCK_BLOCK: case SLURM_DIST_CYCLIC_BLOCK: case SLURM_DIST_PLANE: /* tasks are distributed in blocks within a plane */ rc = _task_layout_lllp_block(req, node_id, &masks); break; case SLURM_DIST_CYCLIC: case SLURM_DIST_BLOCK: case SLURM_DIST_CYCLIC_CYCLIC: case SLURM_DIST_BLOCK_CYCLIC: rc = _task_layout_lllp_cyclic(req, node_id, &masks); break; default: if (req->cpus_per_task > 1) rc = _task_layout_lllp_multi(req, node_id, &masks); else rc = _task_layout_lllp_cyclic(req, node_id, &masks); req->task_dist = SLURM_DIST_BLOCK_CYCLIC; break; } /* FIXME: I'm worried about core_bitmap with CPU_BIND_TO_SOCKETS & * max_cores - does select/cons_res plugin allocate whole * socket??? Maybe not. Check srun man page. */ if (rc == SLURM_SUCCESS) { _task_layout_display_masks(req, gtid, maxtasks, masks); /* translate abstract masks to actual hardware layout */ _lllp_map_abstract_masks(maxtasks, masks); _task_layout_display_masks(req, gtid, maxtasks, masks); #ifdef HAVE_NUMA if (req->cpu_bind_type & CPU_BIND_TO_LDOMS) { _match_masks_to_ldom(maxtasks, masks); _task_layout_display_masks(req, gtid, maxtasks, masks); } #endif /* convert masks into cpu_bind mask string */ _lllp_generate_cpu_bind(req, maxtasks, masks); } else { char *avail_mask = _alloc_mask(req, &whole_nodes, &whole_sockets, &whole_cores, &whole_threads, &part_sockets, &part_cores); if (avail_mask) { xfree(req->cpu_bind); req->cpu_bind = avail_mask; req->cpu_bind_type &= (~bind_mode); req->cpu_bind_type |= CPU_BIND_MASK; } slurm_sprint_cpu_bind_type(buf_type, req->cpu_bind_type); error("lllp_distribution jobid [%u] overriding binding: %s", req->job_id, buf_type); error("Verify socket/core/thread counts in configuration"); } if (masks) _lllp_free_masks(maxtasks, masks); }