/*
* task_pre_launch() is called prior to exec of application task.
* It is followed by TaskProlog program (from slurm.conf) and
* --task-prolog (from srun command line).
*/
extern int task_pre_launch (slurmd_job_t *job)
{
char base[PATH_MAX], path[PATH_MAX];
int rc = SLURM_SUCCESS;
debug("affinity task_pre_launch:%u.%u, task:%u bind:%u",
job->jobid, job->stepid, job->envtp->procid,
job->cpu_bind_type);
if (conf->task_plugin_param & CPU_BIND_CPUSETS) {
info("Using cpuset affinity for tasks");
#ifdef MULTIPLE_SLURMD
if (snprintf(base, PATH_MAX, "%s/slurm_%s_%u",
CPUSET_DIR,
(conf->node_name != NULL)?conf->node_name:"",
job->jobid) > PATH_MAX) {
error("cpuset path too long");
return SLURM_ERROR;
}
#else
if (snprintf(base, PATH_MAX, "%s/slurm%u",
CPUSET_DIR, job->jobid) > PATH_MAX) {
error("cpuset path too long");
return SLURM_ERROR;
}
#endif
if (snprintf(path, PATH_MAX, "%s/slurm%u.%u_%d",
base, job->jobid, job->stepid,
job->envtp->localid) > PATH_MAX) {
error("cpuset path too long");
return SLURM_ERROR;
}
} else
info("Using sched_affinity for tasks");
/*** CPU binding support ***/
if (job->cpu_bind_type) {
cpu_set_t new_mask, cur_mask;
pid_t mypid = job->envtp->task_pid;
slurm_getaffinity(mypid, sizeof(cur_mask), &cur_mask);
if (get_cpuset(&new_mask, job) &&
(!(job->cpu_bind_type & CPU_BIND_NONE))) {
if (conf->task_plugin_param & CPU_BIND_CPUSETS) {
rc = slurm_set_cpuset(base, path, mypid,
sizeof(new_mask),
&new_mask);
slurm_get_cpuset(path, mypid,
sizeof(cur_mask),
&cur_mask);
} else {
rc = slurm_setaffinity(mypid,
sizeof(new_mask),
&new_mask);
slurm_getaffinity(mypid,
sizeof(cur_mask),
&cur_mask);
}
}
slurm_chkaffinity(rc ? &cur_mask : &new_mask,
job, rc);
} else if (job->mem_bind_type &&
(conf->task_plugin_param & CPU_BIND_CPUSETS)) {
cpu_set_t cur_mask;
pid_t mypid = job->envtp->task_pid;
/* Establish cpuset just for the memory binding */
slurm_getaffinity(mypid, sizeof(cur_mask), &cur_mask);
rc = slurm_set_cpuset(base, path,
(pid_t) job->envtp->task_pid,
sizeof(cur_mask), &cur_mask);
}
#ifdef HAVE_NUMA
if ((conf->task_plugin_param & CPU_BIND_CPUSETS) &&
(slurm_memset_available() >= 0)) {
nodemask_t new_mask, cur_mask;
cur_mask = numa_get_membind();
if (get_memset(&new_mask, job) &&
(!(job->mem_bind_type & MEM_BIND_NONE))) {
slurm_set_memset(path, &new_mask);
if (numa_available() >= 0)
numa_set_membind(&new_mask);
cur_mask = new_mask;
}
slurm_chk_memset(&cur_mask, job);
} else if (job->mem_bind_type && (numa_available() >= 0)) {
nodemask_t new_mask, cur_mask;
cur_mask = numa_get_membind();
if (get_memset(&new_mask, job)
&& (!(job->mem_bind_type & MEM_BIND_NONE))) {
numa_set_membind(&new_mask);
cur_mask = new_mask;
}
slurm_chk_memset(&cur_mask, job);
}
#endif
return rc;
}
/* user to have to play with the cgroup hierarchy to modify it */
extern int task_cgroup_cpuset_set_task_affinity(stepd_step_rec_t *job)
{
int fstatus = SLURM_ERROR;
#ifndef HAVE_HWLOC
error("task/cgroup: plugin not compiled with hwloc support, "
"skipping affinity.");
return fstatus;
#else
char mstr[1 + CPU_SETSIZE / 4];
cpu_bind_type_t bind_type;
cpu_set_t ts;
hwloc_obj_t obj;
hwloc_obj_type_t socket_or_node;
hwloc_topology_t topology;
hwloc_bitmap_t cpuset;
hwloc_obj_type_t hwtype;
hwloc_obj_type_t req_hwtype;
int bind_verbose = 0;
int rc = SLURM_SUCCESS;
pid_t pid = job->envtp->task_pid;
size_t tssize;
uint32_t nldoms;
uint32_t nsockets;
uint32_t ncores;
uint32_t npus;
uint32_t nobj;
uint32_t taskid = job->envtp->localid;
uint32_t jntasks = job->node_tasks;
uint32_t jnpus = jntasks * job->cpus_per_task;
bind_type = job->cpu_bind_type;
if (conf->task_plugin_param & CPU_BIND_VERBOSE ||
bind_type & CPU_BIND_VERBOSE)
bind_verbose = 1 ;
/* Allocate and initialize hwloc objects */
hwloc_topology_init(&topology);
hwloc_topology_load(topology);
cpuset = hwloc_bitmap_alloc();
if ( hwloc_get_type_depth(topology, HWLOC_OBJ_NODE) >
hwloc_get_type_depth(topology, HWLOC_OBJ_SOCKET) ) {
/* One socket contains multiple NUMA-nodes
* like AMD Opteron 6000 series etc.
* In such case, use NUMA-node instead of socket. */
socket_or_node = HWLOC_OBJ_NODE;
} else {
socket_or_node = HWLOC_OBJ_SOCKET;
}
if (bind_type & CPU_BIND_NONE) {
if (bind_verbose)
info("task/cgroup: task[%u] is requesting no affinity",
taskid);
return 0;
} else if (bind_type & CPU_BIND_TO_THREADS) {
if (bind_verbose)
info("task/cgroup: task[%u] is requesting "
"thread level binding",taskid);
req_hwtype = HWLOC_OBJ_PU;
} else if (bind_type & CPU_BIND_TO_CORES) {
if (bind_verbose)
info("task/cgroup: task[%u] is requesting "
"core level binding",taskid);
req_hwtype = HWLOC_OBJ_CORE;
} else if (bind_type & CPU_BIND_TO_SOCKETS) {
if (bind_verbose)
info("task/cgroup: task[%u] is requesting "
"socket level binding",taskid);
req_hwtype = socket_or_node;
} else if (bind_type & CPU_BIND_TO_LDOMS) {
if (bind_verbose)
info("task/cgroup: task[%u] is requesting "
"ldom level binding",taskid);
req_hwtype = HWLOC_OBJ_NODE;
} else if (bind_type & CPU_BIND_TO_BOARDS) {
if (bind_verbose)
info("task/cgroup: task[%u] is requesting "
"board level binding",taskid);
req_hwtype = HWLOC_OBJ_GROUP;
} else if (bind_type & bind_mode_ldom) {
req_hwtype = HWLOC_OBJ_NODE;
} else {
if (bind_verbose)
info("task/cgroup: task[%u] using core level binding"
" by default",taskid);
req_hwtype = HWLOC_OBJ_CORE;
}
/*
* Perform the topology detection. It will only get allowed PUs.
* Detect in the same time the granularity to use for binding.
* The granularity can be relaxed from threads to cores if enough
* cores are available as with hyperthread support, ntasks-per-core
* param can let us have access to more threads per core for each
* task
* Revert back to machine granularity if no finer-grained granularity
* matching the request is found. This will result in no affinity
* applied.
* The detected granularity will be used to find where to best place
* the task, then the cpu_bind option will be used to relax the
* affinity constraint and use more PUs. (i.e. use a core granularity
* to dispatch the tasks across the sockets and then provide access
* to each task to the cores of its socket.)
*/
npus = (uint32_t) hwloc_get_nbobjs_by_type(topology,
HWLOC_OBJ_PU);
ncores = (uint32_t) hwloc_get_nbobjs_by_type(topology,
HWLOC_OBJ_CORE);
nsockets = (uint32_t) hwloc_get_nbobjs_by_type(topology,
socket_or_node);
nldoms = (uint32_t) hwloc_get_nbobjs_by_type(topology,
HWLOC_OBJ_NODE);
hwtype = HWLOC_OBJ_MACHINE;
nobj = 1;
if (npus >= jnpus || bind_type & CPU_BIND_TO_THREADS) {
hwtype = HWLOC_OBJ_PU;
nobj = npus;
}
if (ncores >= jnpus || bind_type & CPU_BIND_TO_CORES) {
hwtype = HWLOC_OBJ_CORE;
nobj = ncores;
}
if (nsockets >= jntasks &&
bind_type & CPU_BIND_TO_SOCKETS) {
hwtype = socket_or_node;
nobj = nsockets;
}
/*
* HWLOC returns all the NUMA nodes available regardless of the
* number of underlying sockets available (regardless of the allowed
* resources). So there is no guarantee that each ldom will be populated
* with usable sockets. So add a simple check that at least ensure that
* we have as many sockets as ldoms before moving to ldoms granularity
*/
if (nldoms >= jntasks &&
nsockets >= nldoms &&
bind_type & (CPU_BIND_TO_LDOMS | bind_mode_ldom)) {
hwtype = HWLOC_OBJ_NODE;
nobj = nldoms;
}
/*
* If not enough objects to do the job, revert to no affinity mode
*/
if (hwloc_compare_types(hwtype,HWLOC_OBJ_MACHINE) == 0) {
info("task/cgroup: task[%u] disabling affinity because of %s "
"granularity",taskid,hwloc_obj_type_string(hwtype));
} else if (hwloc_compare_types(hwtype,HWLOC_OBJ_CORE) >= 0 &&
jnpus > nobj) {
info("task/cgroup: task[%u] not enough %s objects, disabling "
"affinity",taskid,hwloc_obj_type_string(hwtype));
} else if (bind_type & bind_mode) {
/* Explicit binding mode specified by the user
* Bind the taskid in accordance with the specified mode
*/
obj = hwloc_get_obj_by_type(topology, HWLOC_OBJ_MACHINE, 0);
if (!hwloc_bitmap_isequal(obj->complete_cpuset,
obj->allowed_cpuset)) {
info("task/cgroup: entire node must be allocated, "
"disabling affinity, task[%u]", taskid);
fprintf(stderr, "Requested cpu_bind option requires "
"entire node to be allocated; disabling "
"affinity\n");
} else {
if (bind_verbose)
info("task/cgroup: task[%u] is requesting "
"explicit binding mode",taskid);
_get_sched_cpuset(topology, hwtype, req_hwtype, &ts,
job);
tssize = sizeof(cpu_set_t);
fstatus = SLURM_SUCCESS;
if ((rc = sched_setaffinity(pid, tssize, &ts))) {
error("task/cgroup: task[%u] unable to set "
"mask 0x%s", taskid,
cpuset_to_str(&ts, mstr));
fstatus = SLURM_ERROR;
} else if (bind_verbose)
info("task/cgroup: task[%u] mask 0x%s",
taskid, cpuset_to_str(&ts, mstr));
slurm_chkaffinity(&ts, job, rc);
}
} else {
/* Bind the detected object to the taskid, respecting the
* granularity, using the designated or default distribution
* method (block or cyclic). */
char *str;
if (bind_verbose) {
info("task/cgroup: task[%u] using %s granularity",
taskid,hwloc_obj_type_string(hwtype));
}
/* There are two "distributions," controlled by the
* -m option of srun and friends. The first is the
* distribution of tasks to nodes. The second is the
* distribution of allocated cpus to tasks for
* binding. This code is handling the second
* distribution. Here's how the values get set, based
* on the value of -m
*
* SLURM_DIST_CYCLIC = srun -m cyclic
* SLURM_DIST_BLOCK = srun -m block
* SLURM_DIST_CYCLIC_CYCLIC = srun -m cyclic:cyclic
* SLURM_DIST_BLOCK_CYCLIC = srun -m block:cyclic
*
* In the first two cases, the user only specified the
* first distribution. The second distribution
* defaults to cyclic. In the second two cases, the
* user explicitly requested a second distribution of
* cyclic. So all these four cases correspond to a
* second distribution of cyclic. So we want to call
* _task_cgroup_cpuset_dist_cyclic.
*
* If the user explicitly specifies a second
* distribution of block, or if
* CR_CORE_DEFAULT_DIST_BLOCK is configured and the
* user does not explicitly specify a second
* distribution of cyclic, the second distribution is
* block, and we need to call
* _task_cgroup_cpuset_dist_block. In these cases,
* task_dist would be set to SLURM_DIST_CYCLIC_BLOCK
* or SLURM_DIST_BLOCK_BLOCK.
*
* You can see the equivalent code for the
* task/affinity plugin in
* src/plugins/task/affinity/dist_tasks.c, around line 384.
*/
switch (job->task_dist) {
case SLURM_DIST_CYCLIC:
case SLURM_DIST_BLOCK:
case SLURM_DIST_CYCLIC_CYCLIC:
case SLURM_DIST_BLOCK_CYCLIC:
_task_cgroup_cpuset_dist_cyclic(
topology, hwtype, req_hwtype,
job, bind_verbose, cpuset);
break;
default:
_task_cgroup_cpuset_dist_block(
topology, hwtype, req_hwtype,
nobj, job, bind_verbose, cpuset);
}
hwloc_bitmap_asprintf(&str, cpuset);
tssize = sizeof(cpu_set_t);
if (hwloc_cpuset_to_glibc_sched_affinity(topology,cpuset,
&ts,tssize) == 0) {
fstatus = SLURM_SUCCESS;
if ((rc = sched_setaffinity(pid,tssize,&ts))) {
error("task/cgroup: task[%u] unable to set "
"taskset '%s'",taskid,str);
fstatus = SLURM_ERROR;
} else if (bind_verbose) {
info("task/cgroup: task[%u] taskset '%s' is set"
,taskid,str);
}
slurm_chkaffinity(&ts, job, rc);
} else {
error("task/cgroup: task[%u] unable to build "
"taskset '%s'",taskid,str);
fstatus = SLURM_ERROR;
}
free(str);
}
/* Destroy hwloc objects */
hwloc_bitmap_free(cpuset);
hwloc_topology_destroy(topology);
return fstatus;
#endif
}