/******************* FUNCTION *********************/
const char* TopoHwloc::getLevelName ( int id ) const
{
const char * res = hwloc_obj_type_string((hwloc_obj_type_t)id);
allocAssume(res != NULL,"Failed to conver topological level id ... from hwloc.");//id
return res;
}
void output_synthetic(hwloc_topology_t topology, const char *filename, int logical __hwloc_attribute_unused, int legend __hwloc_attribute_unused, int verbose_mode __hwloc_attribute_unused)
{
FILE *output;
hwloc_obj_t obj = hwloc_get_root_obj(topology);
int arity;
if (!obj->symmetric_subtree) {
fprintf(stderr, "Cannot output assymetric topology in synthetic format.\n");
fprintf(stderr, "Adding --no-io may help making the topology symmetric.\n");
return;
}
if (!filename || !strcmp(filename, "-"))
output = stdout;
else {
output = open_file(filename, "w");
if (!output) {
fprintf(stderr, "Failed to open %s for writing (%s)\n", filename, strerror(errno));
return;
}
}
arity = obj->arity;
while (arity) {
obj = obj->first_child;
fprintf(output, "%s:%u ", hwloc_obj_type_string(obj->type), arity);
arity = obj->arity;
}
fprintf(output, "\n");
if (output != stdout)
fclose(output);
}
static void print_obj_info(hwloc_obj_t obj)
{
int i;
if (obj->type == HWLOC_OBJ_CACHE)
dprint(obj->depth, "[%s] L%u cache size: %lu\n",
hwloc_obj_type_string(obj->type), obj->attr->cache.depth,
obj->attr->cache.size);
else {
if (obj->memory.total_memory || obj->memory.local_memory)
dprint(obj->depth, "[%s:%u] total memory: %lu; local memory: %lu\n",
hwloc_obj_type_string(obj->type), obj->os_index, obj->memory.total_memory,
obj->memory.local_memory);
else
dprint(obj->depth, "[%s:%u]\n", hwloc_obj_type_string(obj->type), obj->os_index);
}
for (i = 0; i < obj->arity; i++)
print_obj_info(obj->children[i]);
}
static int chk_mem_bind(hwloc_topology_t topology, hwloc_nodeset_t nodeset, int print)
{
hwloc_membind_policy_t policy;
hwloc_bitmap_t checkset = hwloc_bitmap_alloc();
if(hwloc_get_membind(topology, checkset, &policy, HWLOC_MEMBIND_THREAD|HWLOC_MEMBIND_BYNODESET) == -1){
perror("get_membind");
hwloc_bitmap_free(checkset);
return -1;
}
if(print){
char * policy_name;
switch(policy){
case HWLOC_MEMBIND_DEFAULT:
policy_name = "DEFAULT";
break;
case HWLOC_MEMBIND_FIRSTTOUCH:
policy_name = "FIRSTTOUCH";
break;
case HWLOC_MEMBIND_BIND:
policy_name = "BIND";
break;
case HWLOC_MEMBIND_INTERLEAVE:
policy_name = "INTERLEAVE";
break;
case HWLOC_MEMBIND_NEXTTOUCH:
policy_name = "NEXTTOUCH";
break;
case HWLOC_MEMBIND_MIXED:
policy_name = "MIXED";
break;
default:
policy_name=NULL;
break;
}
hwloc_obj_t mem_obj = hwloc_get_first_largest_obj_inside_cpuset(topology, checkset);
printf("membind(%s)=%s:%d\n",policy_name,hwloc_obj_type_string(mem_obj->type),mem_obj->logical_index);
}
if(nodeset == NULL)
return -1;
int ret = hwloc_bitmap_isequal(nodeset,checkset);
hwloc_bitmap_free(checkset);
return ret ? 0 : -1;
}
int main(void)
{
hwloc_topology_t topology;
char *string = NULL;
hwloc_obj_t obj;
hwloc_bitmap_t set;
hwloc_topology_init(&topology);
hwloc_topology_set_synthetic(topology, SYNTHETIC_TOPOLOGY_DESCRIPTION);
hwloc_topology_load(topology);
set = hwloc_bitmap_alloc();
hwloc_bitmap_sscanf(set, GIVEN_CPUSET_STRING);
obj = hwloc_get_obj_covering_cpuset(topology, set);
assert(obj);
fprintf(stderr, "found covering object type %s covering cpuset %s\n",
hwloc_obj_type_string(obj->type), GIVEN_CPUSET_STRING);
assert(hwloc_bitmap_isincluded(set, obj->cpuset));
hwloc_bitmap_asprintf(&string, obj->cpuset);
fprintf(stderr, "covering object of %s is %s, expected %s\n",
GIVEN_CPUSET_STRING, string, EXPECTED_CPUSET_STRING);
assert(!strcmp(EXPECTED_CPUSET_STRING, string));
free(string);
hwloc_bitmap_sscanf(set, GIVEN_LARGESPLIT_CPUSET_STRING);
obj = hwloc_get_obj_covering_cpuset(topology, set);
assert(obj == hwloc_get_root_obj(topology));
fprintf(stderr, "found system as covering object of first+last cpus cpuset %s\n",
GIVEN_LARGESPLIT_CPUSET_STRING);
hwloc_bitmap_sscanf(set, GIVEN_TOOLARGE_CPUSET_STRING);
obj = hwloc_get_obj_covering_cpuset(topology, set);
assert(!obj);
fprintf(stderr, "found no covering object for too-large cpuset %s\n",
GIVEN_TOOLARGE_CPUSET_STRING);
hwloc_bitmap_free(set);
hwloc_topology_destroy(topology);
return EXIT_SUCCESS;
}
static int chk_cpu_bind(hwloc_topology_t topology, hwloc_cpuset_t cpuset, int print)
{
hwloc_bitmap_t checkset = hwloc_bitmap_alloc();
if(hwloc_get_cpubind(topology, checkset, HWLOC_CPUBIND_THREAD) == -1){
perror("get_cpubind");
hwloc_bitmap_free(checkset);
return -1;
}
if(print){
hwloc_obj_t cpu_obj = hwloc_get_first_largest_obj_inside_cpuset(topology, checkset);
printf("cpubind=%s:%d\n",hwloc_obj_type_string(cpu_obj->type),cpu_obj->logical_index);
}
if(cpuset == NULL)
return -1;
int ret = hwloc_bitmap_isequal(cpuset,checkset);
hwloc_bitmap_free(checkset);
return ret ? 0 : -1;
}
static void _add_hwloc_cpuset(
hwloc_obj_type_t hwtype, hwloc_obj_type_t req_hwtype,
hwloc_obj_t obj, uint32_t taskid, int bind_verbose,
hwloc_bitmap_t cpuset)
{
struct hwloc_obj *pobj;
/* if requested binding overlaps the granularity */
/* use the ancestor cpuset instead of the object one */
if (hwloc_compare_types(hwtype, req_hwtype) > 0) {
/* Get the parent object of req_hwtype or the */
/* one just above if not found (meaning of >0)*/
/* (useful for ldoms binding with !NUMA nodes)*/
pobj = obj->parent;
while (pobj != NULL &&
hwloc_compare_types(pobj->type, req_hwtype) > 0)
pobj = pobj->parent;
if (pobj != NULL) {
if (bind_verbose)
info("task/cgroup: task[%u] higher level %s "
"found", taskid,
hwloc_obj_type_string(pobj->type));
hwloc_bitmap_or(cpuset, cpuset, pobj->allowed_cpuset);
} else {
/* should not be executed */
if (bind_verbose)
info("task/cgroup: task[%u] no higher level "
"found", taskid);
hwloc_bitmap_or(cpuset, cpuset, obj->allowed_cpuset);
}
} else {
hwloc_bitmap_or(cpuset, cpuset, obj->allowed_cpuset);
}
}
/* user to have to play with the cgroup hierarchy to modify it */
extern int task_cgroup_cpuset_set_task_affinity(slurmd_job_t *job)
{
int fstatus = SLURM_ERROR;
#ifndef HAVE_HWLOC
error("task/cgroup: plugin not compiled with hwloc support, "
"skipping affinity.");
return fstatus;
#else
hwloc_obj_type_t socket_or_node;
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;
pid_t pid = job->envtp->task_pid;
cpu_bind_type_t bind_type;
int bind_verbose = 0;
hwloc_topology_t topology;
hwloc_bitmap_t cpuset;
hwloc_obj_type_t hwtype;
hwloc_obj_type_t req_hwtype;
size_t tssize;
cpu_set_t ts;
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);
cpuset = hwloc_bitmap_alloc();
hwloc_topology_load(topology);
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_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) {
hwtype = HWLOC_OBJ_NODE;
nobj = nldoms;
}
/*
* Bind the detected object to the taskid, respecting the
* granularity, using the designated or default distribution
* method (block or cyclic).
* 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 {
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 (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);
}
} 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
}
/* 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, match;
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;
/* Allocate and initialize hwloc objects */
hwloc_topology_init(&topology);
hwloc_topology_load(topology);
cpuset = hwloc_bitmap_alloc();
int spec_threads = 0;
if (job->batch) {
jnpus = job->cpus;
job->cpus_per_task = job->cpus;
} else
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 ;
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);
//info("PU:%d CORE:%d SOCK:%d LDOM:%d", npus, ncores, nsockets, nldoms);
hwtype = HWLOC_OBJ_MACHINE;
nobj = 1;
if ((job->job_core_spec != (uint16_t) NO_VAL) &&
(job->job_core_spec & CORE_SPEC_THREAD) &&
(job->job_core_spec != CORE_SPEC_THREAD)) {
spec_threads = job->job_core_spec & (~CORE_SPEC_THREAD);
}
if (npus >= (jnpus + spec_threads) || 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) &&
(nobj < jnpus)) {
info("task/cgroup: task[%u] not enough %s objects (%d < %d), "
"disabling affinity",
taskid, hwloc_obj_type_string(hwtype), nobj, jnpus);
} 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);
match = hwloc_bitmap_isequal(obj->complete_cpuset,
obj->allowed_cpuset);
if ((job->job_core_spec == (uint16_t) NO_VAL) && !match) {
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 (job->job_core_spec != (uint16_t) NO_VAL)
_validate_mask(taskid, obj, &ts);
if ((rc = sched_setaffinity(pid, tssize, &ts))) {
error("task/cgroup: task[%u] unable to set "
"mask 0x%s", taskid,
cpuset_to_str(&ts, mstr));
error("sched_setaffinity rc = %d", rc);
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 dist %u",
taskid, hwloc_obj_type_string(hwtype),
job->task_dist);
}
/* See srun man page for detailed information on --distribution
* option.
*
* You can see the equivalent code for the
* task/affinity plugin in
* src/plugins/task/affinity/dist_tasks.c, around line 368
*/
switch (job->task_dist & SLURM_DIST_NODESOCKMASK) {
case SLURM_DIST_BLOCK_BLOCK:
case SLURM_DIST_CYCLIC_BLOCK:
case SLURM_DIST_PLANE:
/* tasks are distributed in blocks within a plane */
_task_cgroup_cpuset_dist_block(topology,
hwtype, req_hwtype,
nobj, job, bind_verbose, cpuset);
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) {
_task_cgroup_cpuset_dist_block(topology,
hwtype, req_hwtype,
nobj, job, bind_verbose, cpuset);
break;
}
/* We want to fall through here if we aren't doing a
default dist block.
*/
default:
_task_cgroup_cpuset_dist_cyclic(topology,
hwtype, req_hwtype,
job, bind_verbose, cpuset);
break;
}
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] set taskset '%s'",
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
}
/* mapping by hwloc object looks a lot like mapping by node,
* but has the added complication of possibly having different
* numbers of objects on each node
*/
int orte_rmaps_rr_byobj(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,
hwloc_obj_type_t target, unsigned cache_level)
{
int i, j, nprocs_mapped;
orte_node_t *node;
orte_proc_t *proc;
opal_list_item_t *item;
int num_procs_to_assign, nperobj, nprocs, nxtra_objs=0;
int extra_procs_to_assign=0, nxtra_nodes=0, idx;
hwloc_obj_t obj=NULL;
unsigned int nobjs;
float balance;
bool add_one=false;
/* there are two modes for mapping by object: span and not-span. The
* span mode essentially operates as if there was just a single
* "super-node" in the system - i.e., it balances the load across
* all objects of the indicated type regardless of their location.
* In essence, it acts as if we placed one proc on each object, cycling
* across all objects on all nodes, and then wrapped around to place
* another proc on each object, doing so until all procs were placed.
*
* In contrast, the non-span mode operates similar to byslot mapping.
* All slots on each node are filled, assigning each proc to an object
* on that node in a balanced fashion, and then the mapper moves on
* to the next node. Thus, procs tend to be "front loaded" onto the
* list of nodes, as opposed to being "load balanced" in the span mode
*/
if (ORTE_MAPPING_SPAN & jdata->map->mapping) {
return byobj_span(jdata, app, node_list, num_slots,
num_procs, target, cache_level);
}
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr: mapping no-span by %s for job %s slots %d num_procs %lu",
hwloc_obj_type_string(target),
ORTE_JOBID_PRINT(jdata->jobid),
(int)num_slots, (unsigned long)num_procs);
/* quick check to see if we can map all the procs - can't
* do more because we don't know how many total objects exist
* across all the nodes
*/
if (num_slots < (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;
}
/* compute how many extra procs to put on each node */
if (1 == opal_list_get_size(node_list)) {
/* if there is only one node, then they all have to go on it */
extra_procs_to_assign = app->num_procs;
} else {
balance = (float)(((jdata->num_procs + app->num_procs)*orte_rmaps_base.cpus_per_rank) - num_slots) / (float)opal_list_get_size(node_list);
extra_procs_to_assign = (int)balance;
if (0 < (balance - (float)extra_procs_to_assign)) {
/* compute how many nodes need an extra proc */
nxtra_nodes = ((jdata->num_procs + app->num_procs)*orte_rmaps_base.cpus_per_rank) - num_slots - (extra_procs_to_assign * opal_list_get_size(node_list));
/* add one so that we add an extra proc to the first nodes
* until all procs are mapped
*/
extra_procs_to_assign++;
/* flag that we added one */
add_one = true;
}
}
}
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr: mapping no-span by %s extra_procs %d extra_nodes %d",
hwloc_obj_type_string(target),
extra_procs_to_assign, nxtra_nodes);
nprocs_mapped = 0;
for (item = opal_list_get_first(node_list);
item != opal_list_get_end(node_list);
item = opal_list_get_next(item)) {
node = (orte_node_t*)item;
/* bozo check */
if (NULL == node->topology) {
orte_show_help("help-orte-rmaps-ppr.txt", "ppr-topo-missing",
true, node->name);
return ORTE_ERR_SILENT;
}
/* add this node to the map, if reqd */
if (!node->mapped) {
if (ORTE_SUCCESS > (idx = opal_pointer_array_add(jdata->map->nodes, (void*)node))) {
ORTE_ERROR_LOG(idx);
return idx;
}
node->mapped = true;
OBJ_RETAIN(node); /* maintain accounting on object */
++(jdata->map->num_nodes);
}
/* compute the number of procs to go on this node */
if (add_one) {
if (0 == nxtra_nodes) {
--extra_procs_to_assign;
add_one = false;
} else {
--nxtra_nodes;
}
}
if (node->slots <= node->slots_inuse) {
/* everybody takes at least the extras */
num_procs_to_assign = extra_procs_to_assign;
} else {
num_procs_to_assign = (node->slots - node->slots_inuse)/orte_rmaps_base.cpus_per_rank + extra_procs_to_assign;
if (app->num_procs < num_procs_to_assign) {
/* might have more slots than procs */
num_procs_to_assign = app->num_procs;
}
}
/* get the number of objects of this type on this node */
nobjs = opal_hwloc_base_get_nbobjs_by_type(node->topology, target, cache_level, OPAL_HWLOC_AVAILABLE);
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:byobj: nprocs-to-assign %d for %d objs on node %s", num_procs_to_assign, nobjs, node->name);
/* if there are no objects of this type, then report the error
* and abort - this can happen, for example, on systems that
* don't report "sockets" as an independent object. However, IF
* this object is the default one - i.e., not specified by the
* user - then we can fall back to mapping by slot
*/
if (0 == nobjs) {
if (ORTE_MAPPING_GIVEN & ORTE_GET_MAPPING_DIRECTIVE(orte_rmaps_base.mapping)) {
orte_show_help("help-orte-rmaps-base.txt", "orte-rmaps-base:no-objects",
true, hwloc_obj_type_string(target), node->name);
return ORTE_ERR_SILENT;
} else {
/* this was the default mapping policy, so clear the map
* of any prior work and indicate that map-by slot is reqd
*/
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;
}
for (idx=0; idx < node->procs->size; idx++) {
if (NULL == (proc = (orte_proc_t*)opal_pointer_array_get_item(node->procs, idx))) {
continue;
}
if (proc->name.jobid != jdata->jobid) {
continue;
}
--node->num_procs;
OBJ_RELEASE(proc);
opal_pointer_array_set_item(node->procs, idx, NULL);
}
if (0 == node->num_procs) {
node->mapped = false;
OBJ_RELEASE(node);
opal_pointer_array_set_item(jdata->map->nodes, i, NULL);
}
}
return ORTE_ERR_NOT_SUPPORTED;
}
}
/* compute the number of procs to go on each object */
nperobj = num_procs_to_assign / nobjs;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:byobj: placing %d procs on each object", nperobj);
if ((int)(nperobj * nobjs) < num_procs_to_assign) {
/* compute how many objs need an extra proc */
nxtra_objs = num_procs_to_assign - nperobj * nobjs;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:byobj: adding 1 extra proc to the first %d objects, if needed", nxtra_objs);
}
/* loop through the number of objects */
for (i=0; i < (int)nobjs && nprocs_mapped < (int)app->num_procs; i++) {
/* get the hwloc object */
if (NULL == (obj = opal_hwloc_base_get_obj_by_type(node->topology, target, cache_level, i, OPAL_HWLOC_AVAILABLE))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
return ORTE_ERR_NOT_FOUND;
}
/* map the reqd number of procs */
if (0 < nxtra_objs) {
nprocs = nperobj + 1;
--nxtra_objs;
} else {
nprocs = nperobj;
}
for (j=0; j < nprocs && nprocs_mapped < app->num_procs; j++) {
if (NULL == (proc = orte_rmaps_base_setup_proc(jdata, node, app->idx))) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
nprocs_mapped++;
proc->locale = obj;
}
}
/* not all nodes are equal, so only set oversubscribed for
* this node if it is in that state
*/
if (node->slots < (int)node->num_procs) {
/* flag the node as oversubscribed so that sched-yield gets
* properly set
*/
node->oversubscribed = true;
}
if (nprocs_mapped == app->num_procs) {
/* we are done */
break;
}
}
return ORTE_SUCCESS;
}
static int byobj_span(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,
hwloc_obj_type_t target, unsigned cache_level)
{
int i, j, nprocs_mapped, lag, delta, navg;
orte_node_t *node;
orte_proc_t *proc;
opal_list_item_t *item;
int num_procs_to_assign, nperobj, nprocs, nxtra_objs=0;
int extra_procs_to_assign=0, nxtra_nodes=0, idx;
hwloc_obj_t obj=NULL;
unsigned int nobjs;
float balance;
bool add_one=false;
bool oversubscribed=false;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr: mapping span by %s for job %s slots %d num_procs %lu",
hwloc_obj_type_string(target),
ORTE_JOBID_PRINT(jdata->jobid),
(int)num_slots, (unsigned long)num_procs);
/* quick check to see if we can map all the procs - can't
* do more because we don't know how many total objects exist
* across all the nodes
*/
if (num_slots < (int)app->num_procs) {
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;
}
oversubscribed = true;
}
/* divide the procs evenly across all nodes - this is the
* average we have to maintain as we go, but we adjust
* the number on each node to reflect its available slots.
* Obviously, if all nodes have the same number of slots,
* then the avg is what we get on each node - this is
* the most common situation.
*/
navg = app->num_procs / opal_list_get_size(node_list);
if (0 == navg) {
/* if there are less procs than nodes, we have to
* place at least one/node
*/
navg = 1;
}
/* compute how many extra procs to put on each node */
balance = (float)((jdata->num_procs + app->num_procs) - (navg * opal_list_get_size(node_list))) / (float)opal_list_get_size(node_list);
extra_procs_to_assign = (int)balance;
if (0 < (balance - (float)extra_procs_to_assign)) {
/* compute how many nodes need an extra proc */
nxtra_nodes = (jdata->num_procs + app->num_procs) - ((navg + extra_procs_to_assign) * opal_list_get_size(node_list));
/* add one so that we add an extra proc to the first nodes
* until all procs are mapped
*/
extra_procs_to_assign++;
/* flag that we added one */
add_one = true;
}
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr: mapping by %s navg %d extra_procs %d extra_nodes %d",
hwloc_obj_type_string(target),
navg, extra_procs_to_assign, nxtra_nodes);
nprocs_mapped = 0;
lag = 0;
for (item = opal_list_get_first(node_list);
item != opal_list_get_end(node_list);
item = opal_list_get_next(item)) {
node = (orte_node_t*)item;
/* bozo check */
if (NULL == node->topology) {
orte_show_help("help-orte-rmaps-ppr.txt", "ppr-topo-missing",
true, node->name);
return ORTE_ERR_SILENT;
}
/* add this node to the map, if reqd */
if (!node->mapped) {
if (ORTE_SUCCESS > (idx = opal_pointer_array_add(jdata->map->nodes, (void*)node))) {
ORTE_ERROR_LOG(idx);
return idx;
}
node->mapped = true;
OBJ_RETAIN(node); /* maintain accounting on object */
++(jdata->map->num_nodes);
}
/* compute the number of procs to go on this node */
if (add_one) {
if (0 == nxtra_nodes) {
--extra_procs_to_assign;
add_one = false;
} else {
--nxtra_nodes;
}
}
if (oversubscribed) {
/* everybody just takes their share */
num_procs_to_assign = navg + extra_procs_to_assign;
} else {
/* if we are not oversubscribed, then there are enough
* slots to handle all the procs. However, not every
* node will have the same number of slots, so we
* have to track how many procs to "shift" elsewhere
* to make up the difference
*/
if (node->slots <= node->slots_inuse) {
/* if there are no extras to take, then we can
* safely remove this node as we don't need it
*/
if (0 == extra_procs_to_assign) {
opal_pointer_array_set_item(jdata->map->nodes, node->index, NULL);
OBJ_RELEASE(node);
--(jdata->map->num_nodes);
/* update how many we are lagging behind */
lag += navg;
continue;
}
/* everybody has to take at least the extras */
num_procs_to_assign = extra_procs_to_assign;
/* update how many we are lagging behind */
lag += navg;
} else {
/* if slots < avg, then take all */
if ((node->slots - node->slots_inuse) < navg) {
num_procs_to_assign = (node->slots - node->slots_inuse) + extra_procs_to_assign;
/* update how many we are lagging behind */
lag += navg - (node->slots - node->slots_inuse);
} else {
/* take the avg plus as much of the "lag" as we can */
delta = 0;
if (0 < lag) {
delta = (node->slots - node->slots_inuse) - navg;
if (lag < delta) {
delta = lag;
}
lag -= delta;
}
num_procs_to_assign = navg + delta + extra_procs_to_assign;
}
}
}
/* get the number of objects of this type on this node */
nobjs = opal_hwloc_base_get_nbobjs_by_type(node->topology, target, cache_level, OPAL_HWLOC_AVAILABLE);
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:byobj: found %d objs on node %s", nobjs, node->name);
/* compute the number of procs to go on each object */
nperobj = num_procs_to_assign / nobjs;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:byobj: placing %d procs on each object", nperobj);
if ((int)(nperobj * nobjs) < num_procs_to_assign) {
/* compute how many objs need an extra proc */
nxtra_objs = num_procs_to_assign - nperobj * nobjs;
opal_output_verbose(2, orte_rmaps_base_framework.framework_output,
"mca:rmaps:rr:byobj: adding 1 extra proc to the first %d objects, if needed", nxtra_objs);
}
/* loop through the number of objects */
for (i=0; i < (int)nobjs && nprocs_mapped < (int)app->num_procs; i++) {
/* get the hwloc object */
if (NULL == (obj = opal_hwloc_base_get_obj_by_type(node->topology, target, cache_level, i, OPAL_HWLOC_AVAILABLE))) {
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
return ORTE_ERR_NOT_FOUND;
}
/* map the reqd number of procs */
if (0 < nxtra_objs) {
nprocs = nperobj + 1;
--nxtra_objs;
} else {
nprocs = nperobj;
}
for (j=0; j < nprocs && nprocs_mapped < app->num_procs; j++) {
if (NULL == (proc = orte_rmaps_base_setup_proc(jdata, node, app->idx))) {
return ORTE_ERR_OUT_OF_RESOURCE;
}
nprocs_mapped++;
proc->locale = obj;
}
/* keep track of the node we last used */
jdata->bookmark = node;
}
/* not all nodes are equal, so only set oversubscribed for
* this node if it is in that state
*/
if (node->slots < (int)node->num_procs) {
/* flag the node as oversubscribed so that sched-yield gets
* properly set
*/
node->oversubscribed = true;
}
if (nprocs_mapped == app->num_procs) {
/* we are done */
break;
}
}
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;
}
static void
hwloc_info_show_obj(hwloc_obj_t obj, const char *type, const char *prefix, int verbose)
{
char s[128];
unsigned i;
if (verbose < 0)
return;
printf("%s type = %s\n", prefix, hwloc_obj_type_string(obj->type));
printf("%s full type = %s\n", prefix, type);
printf("%s logical index = %u\n", prefix, obj->logical_index);
if (obj->os_index != (unsigned) -1)
printf("%s os index = %u\n", prefix, obj->os_index);
if (obj->name)
printf("%s name = %s\n", prefix, obj->name);
if (obj->depth != (unsigned) -1)
printf("%s depth = %u\n", prefix, obj->depth);
printf("%s sibling rank = %u\n", prefix, obj->sibling_rank);
printf("%s children = %u\n", prefix, obj->arity);
if (obj->memory.local_memory)
printf("%s local memory = %llu\n", prefix, (unsigned long long) obj->memory.local_memory);
if (obj->memory.total_memory)
printf("%s total memory = %llu\n", prefix, (unsigned long long) obj->memory.total_memory);
if (obj->cpuset) {
hwloc_bitmap_snprintf(s, sizeof(s), obj->cpuset);
printf("%s cpuset = %s\n", prefix, s);
}
if (obj->complete_cpuset) {
hwloc_bitmap_snprintf(s, sizeof(s), obj->complete_cpuset);
printf("%s complete cpuset = %s\n", prefix, s);
}
if (obj->online_cpuset) {
hwloc_bitmap_snprintf(s, sizeof(s), obj->online_cpuset);
printf("%s online cpuset = %s\n", prefix, s);
}
if (obj->allowed_cpuset) {
hwloc_bitmap_snprintf(s, sizeof(s), obj->allowed_cpuset);
printf("%s allowed cpuset = %s\n", prefix, s);
}
if (obj->nodeset) {
hwloc_bitmap_snprintf(s, sizeof(s), obj->nodeset);
printf("%s nodeset = %s\n", prefix, s);
}
if (obj->complete_nodeset) {
hwloc_bitmap_snprintf(s, sizeof(s), obj->complete_nodeset);
printf("%s complete nodeset = %s\n", prefix, s);
}
if (obj->allowed_nodeset) {
hwloc_bitmap_snprintf(s, sizeof(s), obj->allowed_nodeset);
printf("%s allowed nodeset = %s\n", prefix, s);
}
switch (obj->type) {
case HWLOC_OBJ_CACHE:
printf("%s attr cache depth = %u\n", prefix, obj->attr->cache.depth);
switch (obj->attr->cache.type) {
case HWLOC_OBJ_CACHE_UNIFIED: printf("%s attr cache type = Unified\n", prefix); break;
case HWLOC_OBJ_CACHE_DATA: printf("%s attr cache type = Data\n", prefix); break;
case HWLOC_OBJ_CACHE_INSTRUCTION: printf("%s attr cache type = Instruction\n", prefix); break;
}
printf("%s attr cache size = %llu\n", prefix, (unsigned long long) obj->attr->cache.size);
printf("%s attr cache line size = %u\n", prefix, obj->attr->cache.linesize);
if (obj->attr->cache.associativity == -1)
printf("%s attr cache ways = Fully-associative\n", prefix);
else if (obj->attr->cache.associativity != 0)
printf("%s attr cache ways = %d\n", prefix, obj->attr->cache.associativity);
break;
case HWLOC_OBJ_GROUP:
printf("%s attr group depth = %u\n", prefix, obj->attr->group.depth);
break;
case HWLOC_OBJ_BRIDGE:
switch (obj->attr->bridge.upstream_type) {
case HWLOC_OBJ_BRIDGE_HOST:
printf("%s attr bridge upstream type = Host\n", prefix);
break;
case HWLOC_OBJ_BRIDGE_PCI:
printf("%s attr bridge upstream type = PCI\n", prefix);
printf("%s attr PCI bus id = %04x:%02x:%02x.%01x\n",
prefix, obj->attr->pcidev.domain, obj->attr->pcidev.bus, obj->attr->pcidev.dev, obj->attr->pcidev.func);
printf("%s attr PCI class = %04x\n",
prefix, obj->attr->pcidev.class_id);
printf("%s attr PCI id = %04x:%04x\n",
prefix, obj->attr->pcidev.vendor_id, obj->attr->pcidev.device_id);
if (obj->attr->pcidev.linkspeed)
printf("%s attr PCI linkspeed = %f GB/s\n", prefix, obj->attr->pcidev.linkspeed);
break;
}
switch (obj->attr->bridge.downstream_type) {
case HWLOC_OBJ_BRIDGE_HOST:
assert(0);
case HWLOC_OBJ_BRIDGE_PCI:
printf("%s attr bridge downstream type = PCI\n", prefix);
printf("%s attr PCI secondary bus = %02x\n",
prefix, obj->attr->bridge.downstream.pci.secondary_bus);
printf("%s attr PCI subordinate bus = %02x\n",
prefix, obj->attr->bridge.downstream.pci.subordinate_bus);
break;
}
break;
case HWLOC_OBJ_PCI_DEVICE:
printf("%s attr PCI bus id = %04x:%02x:%02x.%01x\n",
prefix, obj->attr->pcidev.domain, obj->attr->pcidev.bus, obj->attr->pcidev.dev, obj->attr->pcidev.func);
printf("%s attr PCI class = %04x\n",
prefix, obj->attr->pcidev.class_id);
printf("%s attr PCI id = %04x:%04x\n",
prefix, obj->attr->pcidev.vendor_id, obj->attr->pcidev.device_id);
if (obj->attr->pcidev.linkspeed)
printf("%s attr PCI linkspeed = %f GB/s\n", prefix, obj->attr->pcidev.linkspeed);
break;
case HWLOC_OBJ_OS_DEVICE:
printf("%s attr osdev type = %s\n", prefix, type);
break;
default:
/* nothing to show */
break;
}
for(i=0; i<obj->infos_count; i++) {
printf("%s info %s = %s\n", prefix, obj->infos[i].name, obj->infos[i].value);
}
}
/* user to have to play with the cgroup hierarchy to modify it */
extern int task_cgroup_cpuset_set_task_affinity(slurmd_job_t *job)
{
int fstatus = SLURM_ERROR;
#ifndef HAVE_HWLOC
error("task/cgroup: plugin not compiled with hwloc support, "
"skipping affinity.");
return fstatus;
#else
uint32_t i;
uint32_t nldoms;
uint32_t nsockets;
uint32_t ncores;
uint32_t npus;
uint32_t nobj;
uint32_t pfirst,plast;
uint32_t taskid = job->envtp->localid;
uint32_t jntasks = job->node_tasks;
uint32_t jnpus = jntasks * job->cpus_per_task;
pid_t pid = job->envtp->task_pid;
cpu_bind_type_t bind_type;
int verbose;
hwloc_topology_t topology;
#if HWLOC_API_VERSION <= 0x00010000
hwloc_cpuset_t cpuset,ct;
#else
hwloc_bitmap_t cpuset,ct;
#endif
hwloc_obj_t obj;
struct hwloc_obj *pobj;
hwloc_obj_type_t hwtype;
hwloc_obj_type_t req_hwtype;
int hwdepth;
size_t tssize;
cpu_set_t ts;
bind_type = job->cpu_bind_type ;
if (conf->task_plugin_param & CPU_BIND_VERBOSE ||
bind_type & CPU_BIND_VERBOSE)
verbose = 1 ;
if (bind_type & CPU_BIND_NONE) {
if (verbose)
info("task/cgroup: task[%u] is requesting no affinity",
taskid);
return 0;
} else if (bind_type & CPU_BIND_TO_THREADS) {
if (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 (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 (verbose)
info("task/cgroup: task[%u] is requesting "
"socket level binding",taskid);
req_hwtype = HWLOC_OBJ_SOCKET;
} else if (bind_type & CPU_BIND_TO_LDOMS) {
if (verbose)
info("task/cgroup: task[%u] is requesting "
"ldom level binding",taskid);
req_hwtype = HWLOC_OBJ_NODE;
} else {
if (verbose)
info("task/cgroup: task[%u] using core level binding"
" by default",taskid);
req_hwtype = HWLOC_OBJ_CORE;
}
/* Allocate and initialize hwloc objects */
hwloc_topology_init(&topology);
#if HWLOC_API_VERSION <= 0x00010000
cpuset = hwloc_cpuset_alloc() ;
#else
cpuset = hwloc_bitmap_alloc() ;
#endif
/*
* 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.)
*/
hwloc_topology_load(topology);
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,
HWLOC_OBJ_SOCKET);
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 = HWLOC_OBJ_SOCKET;
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) {
hwtype = HWLOC_OBJ_NODE;
nobj = nldoms;
}
/*
* Perform a block binding on the detected object respecting the
* granularity.
* 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 (verbose) {
info("task/cgroup: task[%u] using %s granularity",
taskid,hwloc_obj_type_string(hwtype));
}
if (hwloc_compare_types(hwtype,HWLOC_OBJ_CORE) >= 0) {
/* cores or threads granularity */
pfirst = taskid * job->cpus_per_task ;
plast = pfirst + job->cpus_per_task - 1;
} else {
/* sockets or ldoms granularity */
pfirst = taskid;
plast = pfirst;
}
hwdepth = hwloc_get_type_depth(topology,hwtype);
for (i = pfirst; i <= plast && i < nobj ; i++) {
obj = hwloc_get_obj_by_depth(topology,hwdepth,(int)i);
/* if requested binding overlap the granularity */
/* use the ancestor cpuset instead of the object one */
if (hwloc_compare_types(hwtype,req_hwtype) > 0) {
/* Get the parent object of req_hwtype or the */
/* one just above if not found (meaning of >0)*/
/* (useful for ldoms binding with !NUMA nodes)*/
pobj = obj->parent;
while (pobj != NULL &&
hwloc_compare_types(pobj->type,
req_hwtype) > 0)
pobj = pobj->parent;
if (pobj != NULL) {
if (verbose)
info("task/cgroup: task[%u] "
"higher level %s found",
taskid,
hwloc_obj_type_string(
pobj->type));
#if HWLOC_API_VERSION <= 0x00010000
ct = hwloc_cpuset_dup(pobj->
allowed_cpuset);
hwloc_cpuset_or(cpuset,cpuset,ct);
hwloc_cpuset_free(ct);
#else
ct = hwloc_bitmap_dup(pobj->
allowed_cpuset);
hwloc_bitmap_or(cpuset,cpuset,ct);
hwloc_bitmap_free(ct);
#endif
} else {
/* should not be executed */
if (verbose)
info("task/cgroup: task[%u] "
"no higher level found",
taskid);
#if HWLOC_API_VERSION <= 0x00010000
ct = hwloc_cpuset_dup(obj->
allowed_cpuset);
hwloc_cpuset_or(cpuset,cpuset,ct);
hwloc_cpuset_free(ct);
#else
ct = hwloc_bitmap_dup(obj->
allowed_cpuset);
hwloc_bitmap_or(cpuset,cpuset,ct);
hwloc_bitmap_free(ct);
#endif
}
} else {
#if HWLOC_API_VERSION <= 0x00010000
ct = hwloc_cpuset_dup(obj->allowed_cpuset);
hwloc_cpuset_or(cpuset,cpuset,ct);
hwloc_cpuset_free(ct);
#else
ct = hwloc_bitmap_dup(obj->allowed_cpuset);
hwloc_bitmap_or(cpuset,cpuset,ct);
hwloc_bitmap_free(ct);
#endif
}
}
char *str;
#if HWLOC_API_VERSION <= 0x00010000
hwloc_cpuset_asprintf(&str,cpuset);
#else
hwloc_bitmap_asprintf(&str,cpuset);
#endif
tssize = sizeof(cpu_set_t);
if (hwloc_cpuset_to_glibc_sched_affinity(topology,cpuset,
&ts,tssize) == 0) {
fstatus = SLURM_SUCCESS;
if (sched_setaffinity(pid,tssize,&ts)) {
error("task/cgroup: task[%u] unable to set "
"taskset '%s'",taskid,str);
fstatus = SLURM_ERROR;
} else if (verbose) {
info("task/cgroup: task[%u] taskset '%s' is set"
,taskid,str);
}
} else {
error("task/cgroup: task[%u] unable to build "
"taskset '%s'",taskid,str);
fstatus = SLURM_ERROR;
}
free(str);
}
/* Destroy hwloc objects */
#if HWLOC_API_VERSION <= 0x00010000
hwloc_cpuset_free(cpuset);
#else
hwloc_bitmap_free(cpuset);
#endif
hwloc_topology_destroy(topology);
return fstatus;
#endif
}
static void
look_rset(int sdl, hwloc_obj_type_t type, struct hwloc_topology *topology, int level)
{
rsethandle_t rset, rad;
int i,maxcpus,j;
int nbnodes;
struct hwloc_obj *obj;
if ((topology->flags & HWLOC_TOPOLOGY_FLAG_WHOLE_SYSTEM))
rset = rs_alloc(RS_ALL);
else
rset = rs_alloc(RS_PARTITION);
rad = rs_alloc(RS_EMPTY);
nbnodes = rs_numrads(rset, sdl, 0);
if (nbnodes == -1) {
perror("rs_numrads");
return;
}
for (i = 0; i < nbnodes; i++) {
hwloc_bitmap_t cpuset;
unsigned os_index = (unsigned) -1; /* no os_index except for PU and NUMANODE below */
if (rs_getrad(rset, rad, sdl, i, 0)) {
fprintf(stderr,"rs_getrad(%d) failed: %s\n", i, strerror(errno));
continue;
}
if (!rs_getinfo(rad, R_NUMPROCS, 0))
continue;
maxcpus = rs_getinfo(rad, R_MAXPROCS, 0);
cpuset = hwloc_bitmap_alloc();
for (j = 0; j < maxcpus; j++) {
if (rs_op(RS_TESTRESOURCE, rad, NULL, R_PROCS, j))
hwloc_bitmap_set(cpuset, j);
}
if (type == HWLOC_OBJ_PU) {
os_index = hwloc_bitmap_first(cpuset);
hwloc_debug("Found PU #%u inside node %d for sdl %d\n", os_index, i, sdl);
assert(hwloc_bitmap_weight(cpuset) == 1);
} else if (type == HWLOC_OBJ_NUMANODE) {
/* NUMA node os_index isn't used for binding, just use the rad number to get unique values.
* Note that we'll use that fact in hwloc_aix_prepare_membind(). */
os_index = i;
hwloc_debug("Using os_index #%u for NUMA node inside node %d for sdl %d\n", os_index, i, sdl);
}
obj = hwloc_alloc_setup_object(type, os_index);
obj->cpuset = cpuset;
obj->os_level = sdl;
switch(type) {
case HWLOC_OBJ_NUMANODE:
obj->nodeset = hwloc_bitmap_alloc();
hwloc_bitmap_set(obj->nodeset, i);
obj->memory.local_memory = 0; /* TODO: odd, rs_getinfo(rad, R_MEMSIZE, 0) << 10 returns the total memory ... */
obj->memory.page_types_len = 2;
obj->memory.page_types = malloc(2*sizeof(*obj->memory.page_types));
memset(obj->memory.page_types, 0, 2*sizeof(*obj->memory.page_types));
obj->memory.page_types[0].size = hwloc_getpagesize();
#ifdef HAVE__SC_LARGE_PAGESIZE
obj->memory.page_types[1].size = sysconf(_SC_LARGE_PAGESIZE);
#endif
/* TODO: obj->memory.page_types[1].count = rs_getinfo(rset, R_LGPGFREE, 0) / hugepagesize */
break;
case HWLOC_OBJ_CACHE:
obj->attr->cache.size = _system_configuration.L2_cache_size;
obj->attr->cache.associativity = _system_configuration.L2_cache_asc;
obj->attr->cache.linesize = 0; /* unknown by default */
if (__power_pc())
if (__power_4() || __power_5() || __power_6() || __power_7())
obj->attr->cache.linesize = 128;
obj->attr->cache.depth = 2;
obj->attr->cache.type = HWLOC_OBJ_CACHE_UNIFIED; /* OK for power[4567], unknown for others */
break;
case HWLOC_OBJ_GROUP:
obj->attr->group.depth = level;
break;
case HWLOC_OBJ_CORE:
{
hwloc_obj_t obj2, obj3;
obj2 = hwloc_alloc_setup_object(HWLOC_OBJ_CACHE, i);
obj2->cpuset = hwloc_bitmap_dup(obj->cpuset);
obj2->attr->cache.size = _system_configuration.dcache_size;
obj2->attr->cache.associativity = _system_configuration.dcache_asc;
obj2->attr->cache.linesize = _system_configuration.dcache_line;
obj2->attr->cache.depth = 1;
if (_system_configuration.cache_attrib & (1<<30)) {
/* Unified cache */
obj2->attr->cache.type = HWLOC_OBJ_CACHE_UNIFIED;
hwloc_debug("Adding an L1u cache for core %d\n", i);
hwloc_insert_object_by_cpuset(topology, obj2);
} else {
/* Separate Instruction and Data caches */
obj2->attr->cache.type = HWLOC_OBJ_CACHE_DATA;
hwloc_debug("Adding an L1d cache for core %d\n", i);
hwloc_insert_object_by_cpuset(topology, obj2);
obj3 = hwloc_alloc_setup_object(HWLOC_OBJ_CACHE, i);
obj3->cpuset = hwloc_bitmap_dup(obj->cpuset);
obj3->attr->cache.size = _system_configuration.icache_size;
obj3->attr->cache.associativity = _system_configuration.icache_asc;
obj3->attr->cache.linesize = _system_configuration.icache_line;
obj3->attr->cache.depth = 1;
obj3->attr->cache.type = HWLOC_OBJ_CACHE_INSTRUCTION;
hwloc_debug("Adding an L1i cache for core %d\n", i);
hwloc_insert_object_by_cpuset(topology, obj3);
}
break;
}
default:
break;
}
hwloc_debug_2args_bitmap("%s %d has cpuset %s\n",
hwloc_obj_type_string(type),
i, obj->cpuset);
hwloc_insert_object_by_cpuset(topology, obj);
}
rs_free(rset);
rs_free(rad);
}
int orte_ess_base_proc_binding(void)
{
hwloc_obj_t node, obj;
hwloc_cpuset_t cpus, nodeset;
hwloc_obj_type_t target;
unsigned int cache_level = 0;
struct hwloc_topology_support *support;
char *map;
int ret;
char *error=NULL;
hwloc_cpuset_t mycpus;
/* Determine if we were pre-bound or not */
if (NULL != getenv(OPAL_MCA_PREFIX"orte_bound_at_launch")) {
orte_proc_is_bound = true;
if (NULL != (map = getenv(OPAL_MCA_PREFIX"orte_base_applied_binding"))) {
orte_proc_applied_binding = hwloc_bitmap_alloc();
if (0 != (ret = hwloc_bitmap_list_sscanf(orte_proc_applied_binding, map))) {
error = "applied_binding parse";
goto error;
}
}
}
/* see if we were bound when launched */
if (!orte_proc_is_bound) {
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Not bound at launch",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
/* we were not bound at launch */
if (NULL == opal_hwloc_topology) {
/* there is nothing we can do, so just return */
return ORTE_SUCCESS;
}
support = (struct hwloc_topology_support*)hwloc_topology_get_support(opal_hwloc_topology);
/* get our node object */
node = hwloc_get_root_obj(opal_hwloc_topology);
nodeset = opal_hwloc_base_get_available_cpus(opal_hwloc_topology, node);
/* get our bindings */
cpus = hwloc_bitmap_alloc();
if (hwloc_get_cpubind(opal_hwloc_topology, cpus, HWLOC_CPUBIND_PROCESS) < 0) {
/* we are NOT bound if get_cpubind fails, nor can we be bound - the
* environment does not support it
*/
hwloc_bitmap_free(cpus);
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Binding not supported",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
goto MOVEON;
}
/* we are bound if the two cpusets are not equal,
* or if there is only ONE cpu available to us
*/
if (0 != hwloc_bitmap_compare(cpus, nodeset) ||
opal_hwloc_base_single_cpu(nodeset) ||
opal_hwloc_base_single_cpu(cpus)) {
/* someone external set it - indicate it is set
* so that we know
*/
orte_proc_is_bound = true;
hwloc_bitmap_list_asprintf(&orte_process_info.cpuset, cpus);
hwloc_bitmap_free(cpus);
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Process was externally bound",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
} else if (support->cpubind->set_thisproc_cpubind &&
OPAL_BINDING_POLICY_IS_SET(opal_hwloc_binding_policy) &&
OPAL_BIND_TO_NONE != OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
/* the system is capable of doing processor affinity, but it
* has not yet been set - see if a slot_list was given
*/
hwloc_bitmap_zero(cpus);
if (OPAL_BIND_TO_CPUSET == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
if (OPAL_SUCCESS != (ret = opal_hwloc_base_slot_list_parse(opal_hwloc_base_slot_list,
opal_hwloc_topology,
OPAL_HWLOC_LOGICAL, cpus))) {
error = "Setting processor affinity failed";
hwloc_bitmap_free(cpus);
goto error;
}
if (0 > hwloc_set_cpubind(opal_hwloc_topology, cpus, 0)) {
error = "Setting processor affinity failed";
hwloc_bitmap_free(cpus);
goto error;
}
hwloc_bitmap_list_asprintf(&orte_process_info.cpuset, cpus);
hwloc_bitmap_free(cpus);
orte_proc_is_bound = true;
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Process bound according to slot_list",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
} else {
/* cleanup */
hwloc_bitmap_free(cpus);
/* get the node rank */
if (ORTE_NODE_RANK_INVALID == orte_process_info.my_node_rank) {
/* this is not an error - could be due to being
* direct launched - so just ignore and leave
* us unbound
*/
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Process not bound - no node rank available",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
goto MOVEON;
}
/* if the binding policy is hwthread, then we bind to the nrank-th
* hwthread on this node
*/
if (OPAL_BIND_TO_HWTHREAD == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
if (NULL == (obj = opal_hwloc_base_get_obj_by_type(opal_hwloc_topology, HWLOC_OBJ_PU,
0, orte_process_info.my_node_rank, OPAL_HWLOC_LOGICAL))) {
ret = ORTE_ERR_NOT_FOUND;
error = "Getting hwthread object";
goto error;
}
cpus = opal_hwloc_base_get_available_cpus(opal_hwloc_topology, obj);
if (0 > hwloc_set_cpubind(opal_hwloc_topology, cpus, 0)) {
ret = ORTE_ERROR;
error = "Setting processor affinity failed";
goto error;
}
hwloc_bitmap_list_asprintf(&orte_process_info.cpuset, cpus);
hwloc_bitmap_free(cpus);
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Process bound to hwthread",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
} else if (OPAL_BIND_TO_CORE == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
/* if the binding policy is core, then we bind to the nrank-th
* core on this node
*/
if (NULL == (obj = opal_hwloc_base_get_obj_by_type(opal_hwloc_topology, HWLOC_OBJ_CORE,
0, orte_process_info.my_node_rank, OPAL_HWLOC_LOGICAL))) {
ret = ORTE_ERR_NOT_FOUND;
error = "Getting core object";
goto error;
}
cpus = opal_hwloc_base_get_available_cpus(opal_hwloc_topology, obj);
if (0 > hwloc_set_cpubind(opal_hwloc_topology, cpus, 0)) {
error = "Setting processor affinity failed";
ret = ORTE_ERROR;
goto error;
}
hwloc_bitmap_list_asprintf(&orte_process_info.cpuset, cpus);
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Process bound to core",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
} else {
/* for all higher binding policies, we bind to the specified
* object that the nrank-th core belongs to
*/
if (NULL == (obj = opal_hwloc_base_get_obj_by_type(opal_hwloc_topology, HWLOC_OBJ_CORE,
0, orte_process_info.my_node_rank, OPAL_HWLOC_LOGICAL))) {
ret = ORTE_ERR_NOT_FOUND;
error = "Getting core object";
goto error;
}
if (OPAL_BIND_TO_L1CACHE == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
target = HWLOC_OBJ_CACHE;
cache_level = 1;
} else if (OPAL_BIND_TO_L2CACHE == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
target = HWLOC_OBJ_CACHE;
cache_level = 2;
} else if (OPAL_BIND_TO_L3CACHE == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
target = HWLOC_OBJ_CACHE;
cache_level = 3;
} else if (OPAL_BIND_TO_SOCKET == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
target = HWLOC_OBJ_SOCKET;
} else if (OPAL_BIND_TO_NUMA == OPAL_GET_BINDING_POLICY(opal_hwloc_binding_policy)) {
target = HWLOC_OBJ_NODE;
} else {
ret = ORTE_ERR_NOT_FOUND;
error = "Binding policy not known";
goto error;
}
for (obj = obj->parent; NULL != obj; obj = obj->parent) {
if (target == obj->type) {
if (HWLOC_OBJ_CACHE == target && cache_level != obj->attr->cache.depth) {
continue;
}
/* this is the place! */
cpus = opal_hwloc_base_get_available_cpus(opal_hwloc_topology, obj);
if (0 > hwloc_set_cpubind(opal_hwloc_topology, cpus, 0)) {
ret = ORTE_ERROR;
error = "Setting processor affinity failed";
goto error;
}
hwloc_bitmap_list_asprintf(&orte_process_info.cpuset, cpus);
orte_proc_is_bound = true;
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Process bound to %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
hwloc_obj_type_string(target)));
break;
}
}
if (!orte_proc_is_bound) {
ret = ORTE_ERROR;
error = "Setting processor affinity failed";
goto error;
}
}
}
}
} else {
OPAL_OUTPUT_VERBOSE((5, orte_ess_base_framework.framework_output,
"%s Process bound at launch",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
}
MOVEON:
/* get or update our local cpuset - it will get used multiple
* times, so it's more efficient to keep a global copy
*/
opal_hwloc_base_get_local_cpuset();
/* get the cpus we are bound to */
mycpus = hwloc_bitmap_alloc();
if (hwloc_get_cpubind(opal_hwloc_topology,
mycpus,
HWLOC_CPUBIND_PROCESS) < 0) {
if (NULL != orte_process_info.cpuset) {
free(orte_process_info.cpuset);
orte_process_info.cpuset = NULL;
}
if (opal_hwloc_report_bindings || 4 < opal_output_get_verbosity(orte_ess_base_framework.framework_output)) {
opal_output(0, "MCW rank %d is not bound",
ORTE_PROC_MY_NAME->vpid);
}
} else {
/* store/update the string representation of our local binding */
if (NULL != orte_process_info.cpuset) {
free(orte_process_info.cpuset);
orte_process_info.cpuset = NULL;
}
hwloc_bitmap_list_asprintf(&orte_process_info.cpuset, mycpus);
/* report the binding, if requested */
if (opal_hwloc_report_bindings || 4 < opal_output_get_verbosity(orte_ess_base_framework.framework_output)) {
char tmp1[1024], tmp2[1024];
if (OPAL_ERR_NOT_BOUND == opal_hwloc_base_cset2str(tmp1, sizeof(tmp1), opal_hwloc_topology, mycpus)) {
opal_output(0, "MCW rank %d is not bound (or bound to all available processors)", ORTE_PROC_MY_NAME->vpid);
} else {
opal_hwloc_base_cset2mapstr(tmp2, sizeof(tmp2), opal_hwloc_topology, mycpus);
opal_output(0, "MCW rank %d bound to %s: %s",
ORTE_PROC_MY_NAME->vpid, tmp1, tmp2);
}
}
}
hwloc_bitmap_free(mycpus);
/* push our cpuset so others can calculate our locality */
if (NULL != orte_process_info.cpuset) {
OPAL_MODEX_SEND_VALUE(ret, OPAL_PMIX_GLOBAL, OPAL_PMIX_CPUSET,
orte_process_info.cpuset, OPAL_STRING);
}
return ORTE_SUCCESS;
error:
if (ORTE_ERR_SILENT != ret) {
orte_show_help("help-orte-runtime",
"orte_init:startup:internal-failure",
true, error, ORTE_ERROR_NAME(ret), ret);
}
return ORTE_ERR_SILENT;
}
int main(int argc, char *argv[])
{
netloc_map_t map;
netloc_map_server_t srcserver, dstserver;
hwloc_topology_t srctopo, dsttopo;
hwloc_obj_t srcobj, dstobj;
netloc_map_paths_t paths;
unsigned nr_paths, nr_edges, i, j;
struct netloc_map_edge_s *edges;
unsigned flags = 0x3;
char *path;
int err;
if (argc > 2) {
if (!strcmp(argv[1], "--flags")) {
flags = (unsigned) strtoul(argv[2], NULL, 0);
argc -= 2;
argv += 2;
}
}
if (argc < 6) {
fprintf(stderr, "%s [options] <datadir> <srcserver> <srcpu> <dstserver> <dstpu>\n", argv[0]);
fprintf(stderr, "Example: %s mynetlocdata server2 1 server3 7\n", argv[0]);
fprintf(stderr, " Loads netloc map from 'mynetlocdata' directory and display all paths\n");
fprintf(stderr, " from server 'server2' PU #1 to server 'server3' PU #7.\n");
fprintf(stderr, "Options:\n");
fprintf(stderr, " --flags N Use value N as map paths flags. Default is 3 which displays all edges.\n");
exit(EXIT_FAILURE);
}
err = netloc_map_create(&map);
if (err) {
fprintf(stderr, "Failed to create the map\n");
exit(EXIT_FAILURE);
}
asprintf(&path, "%s/hwloc", argv[1]);
err = netloc_map_load_hwloc_data(map, path);
free(path);
if (err) {
fprintf(stderr, "Failed to load hwloc data\n");
return -1;
}
asprintf(&path, "file://%s/netloc", argv[1]);
err = netloc_map_load_netloc_data(map, path);
free(path);
if (err) {
fprintf(stderr, "Failed to load netloc data\n");
return -1;
}
err = netloc_map_build(map, 0);
if (err) {
fprintf(stderr, "Failed to build map data\n");
return -1;
}
err = netloc_map_name2server(map, argv[2], &srcserver);
if (err) {
fprintf(stderr, "Could not find src server %s\n", argv[2]);
return -1;
}
err = netloc_map_server2hwloc(srcserver, &srctopo);
if (err) {
fprintf(stderr, "Could not find src server %s hwloc topology\n", argv[2]);
return -1;
}
srcobj = hwloc_get_obj_by_type(srctopo, HWLOC_OBJ_PU, atoi(argv[3]));
if (!srcobj) {
fprintf(stderr, "Could not find src server %s PU #%s\n", argv[2], argv[3]);
return -1;
}
err = netloc_map_name2server(map, argv[4], &dstserver);
if (err) {
fprintf(stderr, "Could not find dst server %s\n", argv[4]);
return -1;
}
err = netloc_map_server2hwloc(dstserver, &dsttopo);
if (err) {
fprintf(stderr, "Could not find dst server %s hwloc topology\n", argv[4]);
return -1;
}
dstobj = hwloc_get_obj_by_type(dsttopo, HWLOC_OBJ_PU, atoi(argv[5]));
if (!dstobj) {
fprintf(stderr, "Could not find dst server %s PU #%s\n", argv[4], argv[5]);
return -1;
}
err = netloc_map_paths_build(map,
srctopo, srcobj,
dsttopo, dstobj,
flags,
&paths, &nr_paths);
if (err < 0) {
fprintf(stderr, "Failed to build paths\n");
return -1;
}
printf("got %u paths\n", nr_paths);
for(i=0; i<nr_paths; i++) {
printf(" path #%u:\n", i);
err = netloc_map_paths_get(paths, i, &edges, &nr_edges);
assert(!err);
printf(" %u edges\n", nr_edges);
for(j=0; j<nr_edges; j++) {
struct netloc_map_edge_s *edge = &edges[j];
printf(" edge #%u type %d: ", j, edge->type);
switch (edge->type) {
case NETLOC_MAP_EDGE_TYPE_NETLOC:
printf("netloc from %s to %s in subnet type %s id %s\n",
edge->netloc.edge->src_node_id,
edge->netloc.edge->dest_node_id,
netloc_decode_network_type(netloc_access_network_ref(edge->netloc.topology)->network_type),
netloc_access_network_ref(edge->netloc.topology)->subnet_id);
break;
case NETLOC_MAP_EDGE_TYPE_HWLOC_PARENT:
printf("hwloc UP from %s:%u (%s) to parent %s:%u (%s) weight %u\n",
hwloc_obj_type_string(edge->hwloc.src_obj->type), edge->hwloc.src_obj->logical_index, edge->hwloc.src_obj->name ? : "<unnamed>",
hwloc_obj_type_string(edge->hwloc.dest_obj->type), edge->hwloc.dest_obj->logical_index, edge->hwloc.dest_obj->name ? : "<unnamed>",
edge->hwloc.weight);
break;
case NETLOC_MAP_EDGE_TYPE_HWLOC_HORIZONTAL:
printf("hwloc HORIZONTAL from %s:%u (%s) to cousin %s:%u (%s) weight %u\n",
hwloc_obj_type_string(edge->hwloc.src_obj->type), edge->hwloc.src_obj->logical_index, edge->hwloc.src_obj->name ? : "<unnamed>",
hwloc_obj_type_string(edge->hwloc.dest_obj->type), edge->hwloc.dest_obj->logical_index, edge->hwloc.dest_obj->name ? : "<unnamed>",
edge->hwloc.weight);
break;
case NETLOC_MAP_EDGE_TYPE_HWLOC_CHILD:
printf("hwloc DOWN from %s:%u (%s) to child %s:%u (%s) weight %u\n",
hwloc_obj_type_string(edge->hwloc.src_obj->type), edge->hwloc.src_obj->logical_index, edge->hwloc.src_obj->name ? : "<unnamed>",
hwloc_obj_type_string(edge->hwloc.dest_obj->type), edge->hwloc.dest_obj->logical_index, edge->hwloc.dest_obj->name ? : "<unnamed>",
edge->hwloc.weight);
break;
case NETLOC_MAP_EDGE_TYPE_HWLOC_PCI:
printf("hwloc PCI from %s:%u (%s) to child %s:%u (%s) weight %u\n",
hwloc_obj_type_string(edge->hwloc.src_obj->type), edge->hwloc.src_obj->logical_index, edge->hwloc.src_obj->name ? : "<unnamed>",
hwloc_obj_type_string(edge->hwloc.dest_obj->type), edge->hwloc.dest_obj->logical_index, edge->hwloc.dest_obj->name ? : "<unnamed>",
edge->hwloc.weight);
break;
}
}
}
netloc_map_paths_destroy(paths);
netloc_map_put_hwloc(map, srctopo);
netloc_map_put_hwloc(map, dsttopo);
netloc_map_destroy(map);
return 0;
}
void output_console(struct lstopo_output *loutput, const char *filename)
{
hwloc_topology_t topology = loutput->topology;
unsigned topodepth;
int verbose_mode = loutput->verbose_mode;
int logical = loutput->logical;
FILE *output;
output = open_output(filename, loutput->overwrite);
if (!output) {
fprintf(stderr, "Failed to open %s for writing (%s)\n", filename, strerror(errno));
return;
}
topodepth = hwloc_topology_get_depth(topology);
/*
* if verbose_mode == 0, only print the summary.
* if verbose_mode == 1, only print the topology tree.
* if verbose_mode > 1, print both.
*/
if (lstopo_show_only != (hwloc_obj_type_t)-1) {
if (verbose_mode > 1)
fprintf(output, "Only showing %s objects\n", hwloc_obj_type_string(lstopo_show_only));
output_only (topology, hwloc_get_root_obj(topology), output, logical, verbose_mode);
} else if (verbose_mode >= 1) {
output_topology (topology, hwloc_get_root_obj(topology), NULL, output, 0, logical, verbose_mode);
fprintf(output, "\n");
}
if ((verbose_mode > 1 || !verbose_mode) && lstopo_show_only == (hwloc_obj_type_t)-1) {
hwloc_lstopo_show_summary(output, topology);
}
if (verbose_mode > 1 && lstopo_show_only == (hwloc_obj_type_t)-1) {
const struct hwloc_distances_s * distances;
unsigned depth;
for (depth = 0; depth < topodepth; depth++) {
distances = hwloc_get_whole_distance_matrix_by_depth(topology, depth);
if (!distances || !distances->latency)
continue;
fprintf(output, "relative latency matrix between %ss (depth %u) by %s indexes:\n",
hwloc_obj_type_string(hwloc_get_depth_type(topology, depth)),
depth,
logical ? "logical" : "physical");
hwloc_utils_print_distance_matrix(output, topology, hwloc_get_root_obj(topology), distances->nbobjs, depth, distances->latency, logical);
}
}
if (verbose_mode > 1 && lstopo_show_only == (hwloc_obj_type_t)-1) {
hwloc_const_bitmap_t complete = hwloc_topology_get_complete_cpuset(topology);
hwloc_const_bitmap_t topo = hwloc_topology_get_topology_cpuset(topology);
hwloc_const_bitmap_t allowed = hwloc_topology_get_allowed_cpuset(topology);
if (!hwloc_bitmap_isequal(topo, complete)) {
hwloc_bitmap_t unknown = hwloc_bitmap_alloc();
char *unknownstr;
hwloc_bitmap_copy(unknown, complete);
hwloc_bitmap_andnot(unknown, unknown, topo);
hwloc_bitmap_asprintf(&unknownstr, unknown);
fprintf (output, "%d processors not represented in topology: %s\n", hwloc_bitmap_weight(unknown), unknownstr);
free(unknownstr);
hwloc_bitmap_free(unknown);
}
if (!hwloc_bitmap_isequal(topo, allowed)) {
hwloc_bitmap_t disallowed = hwloc_bitmap_alloc();
char *disallowedstr;
hwloc_bitmap_copy(disallowed, topo);
hwloc_bitmap_andnot(disallowed, disallowed, allowed);
hwloc_bitmap_asprintf(&disallowedstr, disallowed);
fprintf(output, "%d processors represented but not allowed: %s\n", hwloc_bitmap_weight(disallowed), disallowedstr);
free(disallowedstr);
hwloc_bitmap_free(disallowed);
}
if (!hwloc_topology_is_thissystem(topology))
fprintf (output, "Topology not from this system\n");
}
if (output != stdout)
fclose(output);
}
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 void
hwloc_synthetic_process_level_indexes(struct hwloc_synthetic_backend_data_s *data,
unsigned curleveldepth,
int verbose)
{
struct hwloc_synthetic_level_data_s *curlevel = &data->level[curleveldepth];
unsigned long total = curlevel->totalwidth;
const char *attr = curlevel->index_string;
unsigned long length = curlevel->index_string_length;
unsigned *array = NULL;
struct hwloc_synthetic_intlv_loop_s * loops = NULL;
size_t i;
if (!attr)
return;
array = calloc(total, sizeof(*array));
if (!array) {
if (verbose)
fprintf(stderr, "Failed to allocate synthetic index array of size %lu\n", total);
goto out;
}
i = strspn(attr, "0123456789,");
if (i == length) {
/* explicit array of indexes */
for(i=0; i<total; i++) {
const char *next;
unsigned idx = strtoul(attr, (char **) &next, 10);
if (next == attr) {
if (verbose)
fprintf(stderr, "Failed to read synthetic index #%lu at '%s'\n", (unsigned long) i, attr);
goto out_with_array;
}
array[i] = idx;
if (i != total-1) {
if (*next != ',') {
if (verbose)
fprintf(stderr, "Missing comma after synthetic index #%lu at '%s'\n", (unsigned long) i, attr);
goto out_with_array;
}
attr = next+1;
} else {
attr = next;
}
}
curlevel->index_array = array;
} else {
/* interleaving */
unsigned nr_loops = 1, cur_loop;
unsigned minstep = total;
unsigned long nbs = 1;
unsigned j, mul;
const char *tmp;
tmp = attr;
while (tmp) {
tmp = strchr(tmp, ':');
if (!tmp || tmp >= attr+length)
break;
nr_loops++;
tmp++;
}
/* nr_loops colon-separated fields, but we may need one more at the end */
loops = malloc((nr_loops+1)*sizeof(*loops));
if (!loops) {
if (verbose)
fprintf(stderr, "Failed to allocate synthetic index interleave loop array of size %u\n", nr_loops);
goto out_with_array;
}
if (*attr >= '0' && *attr <= '9') {
/* interleaving as x*y:z*t:... */
unsigned step, nb;
tmp = attr;
cur_loop = 0;
while (tmp) {
char *tmp2, *tmp3;
step = (unsigned) strtol(tmp, &tmp2, 0);
if (tmp2 == tmp || *tmp2 != '*') {
if (verbose)
fprintf(stderr, "Failed to read synthetic index interleaving loop '%s' without number before '*'\n", tmp);
goto out_with_loops;
}
if (!step) {
if (verbose)
fprintf(stderr, "Invalid interleaving loop with step 0 at '%s'\n", tmp);
goto out_with_loops;
}
tmp2++;
nb = (unsigned) strtol(tmp2, &tmp3, 0);
if (tmp3 == tmp2 || (*tmp3 && *tmp3 != ':' && *tmp3 != ')' && *tmp3 != ' ')) {
if (verbose)
fprintf(stderr, "Failed to read synthetic index interleaving loop '%s' without number between '*' and ':'\n", tmp);
goto out_with_loops;
}
if (!nb) {
if (verbose)
fprintf(stderr, "Invalid interleaving loop with number 0 at '%s'\n", tmp2);
goto out_with_loops;
}
loops[cur_loop].step = step;
loops[cur_loop].nb = nb;
if (step < minstep)
minstep = step;
nbs *= nb;
cur_loop++;
if (*tmp3 == ')' || *tmp3 == ' ')
break;
tmp = (const char*) (tmp3+1);
}
} else {
/* interleaving as type1:type2:... */
hwloc_obj_type_t type;
hwloc_obj_cache_type_t cachetypeattr;
int depthattr;
int err;
/* find level depths for each interleaving loop */
tmp = attr;
cur_loop = 0;
while (tmp) {
err = hwloc_obj_type_sscanf(tmp, &type, &depthattr, &cachetypeattr, sizeof(cachetypeattr));
if (err < 0) {
if (verbose)
fprintf(stderr, "Failed to read synthetic index interleaving loop type '%s'\n", tmp);
goto out_with_loops;
}
if (type == HWLOC_OBJ_MISC || type == HWLOC_OBJ_BRIDGE || type == HWLOC_OBJ_PCI_DEVICE || type == HWLOC_OBJ_OS_DEVICE) {
if (verbose)
fprintf(stderr, "Misc object type disallowed in synthetic index interleaving loop type '%s'\n", tmp);
goto out_with_loops;
}
for(i=0; i<curleveldepth; i++) {
if (type != data->level[i].type)
continue;
if ((type == HWLOC_OBJ_GROUP || type == HWLOC_OBJ_CACHE)
&& depthattr != -1
&& (unsigned) depthattr != data->level[i].depth)
continue;
if (type == HWLOC_OBJ_CACHE
&& cachetypeattr != (hwloc_obj_cache_type_t) -1
&& cachetypeattr != data->level[i].cachetype)
continue;
loops[cur_loop].level_depth = (unsigned)i;
break;
}
if (i == curleveldepth) {
if (verbose)
fprintf(stderr, "Failed to find level for synthetic index interleaving loop type '%s' above '%s'\n",
tmp, hwloc_obj_type_string(curlevel->type));
goto out_with_loops;
}
tmp = strchr(tmp, ':');
if (!tmp || tmp > attr+length)
break;
tmp++;
cur_loop++;
}
/* compute actual loop step/nb */
for(cur_loop=0; cur_loop<nr_loops; cur_loop++) {
unsigned mydepth = loops[cur_loop].level_depth;
unsigned prevdepth = 0;
unsigned step, nb;
for(i=0; i<nr_loops; i++) {
if (loops[i].level_depth == mydepth && i != cur_loop) {
if (verbose)
fprintf(stderr, "Invalid duplicate interleaving loop type in synthetic index '%s'\n", attr);
goto out_with_loops;
}
if (loops[i].level_depth < mydepth
&& loops[i].level_depth > prevdepth)
prevdepth = loops[i].level_depth;
}
step = curlevel->totalwidth / data->level[mydepth].totalwidth; /* number of objects below us */
nb = data->level[mydepth].totalwidth / data->level[prevdepth].totalwidth; /* number of us within parent */
loops[cur_loop].step = step;
loops[cur_loop].nb = nb;
assert(nb);
assert(step);
if (step < minstep)
minstep = step;
nbs *= nb;
}
}
assert(nbs);
if (nbs != total) {
/* one loop of total/nbs steps is missing, add it if it's just the smallest one */
if (minstep == total/nbs) {
loops[nr_loops].step = 1;
loops[nr_loops].nb = total/nbs;
nr_loops++;
} else {
if (verbose)
fprintf(stderr, "Invalid index interleaving total width %lu instead of %lu\n", nbs, total);
goto out_with_loops;
}
}
/* generate the array of indexes */
mul = 1;
for(i=0; i<nr_loops; i++) {
unsigned step = loops[i].step;
unsigned nb = loops[i].nb;
for(j=0; j<total; j++)
array[j] += ((j / step) % nb) * mul;
mul *= nb;
}
/* check that we have the right values (cannot pass total, cannot give duplicate 0) */
for(j=0; j<total; j++) {
if (array[j] >= total) {
if (verbose)
fprintf(stderr, "Invalid index interleaving generates out-of-range index %u\n", array[j]);
goto out_with_loops;
}
if (!array[j] && j) {
if (verbose)
fprintf(stderr, "Invalid index interleaving generates duplicate index values\n");
goto out_with_loops;
}
}
free(loops);
curlevel->index_array = array;
}
return;
out_with_loops:
free(loops);
out_with_array:
free(array);
out:
return;
}
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,
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;
}
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;
}
void output_console(hwloc_topology_t topology, const char *filename, int logical, int legend __hwloc_attribute_unused, int verbose_mode)
{
unsigned topodepth;
FILE *output;
if (!filename || !strcmp(filename, "-"))
output = stdout;
else {
output = open_file(filename, "w");
if (!output) {
fprintf(stderr, "Failed to open %s for writing (%s)\n", filename, strerror(errno));
return;
}
}
topodepth = hwloc_topology_get_depth(topology);
/*
* if verbose_mode == 0, only print the summary.
* if verbose_mode == 1, only print the topology tree.
* if verbose_mode > 1, print both.
*/
if (lstopo_show_only != (hwloc_obj_type_t)-1) {
if (verbose_mode > 1)
fprintf(output, "Only showing %s objects\n", hwloc_obj_type_string(lstopo_show_only));
output_only (topology, hwloc_get_root_obj(topology), output, logical, verbose_mode);
} else if (verbose_mode >= 1) {
output_topology (topology, hwloc_get_root_obj(topology), NULL, output, 0, logical, verbose_mode);
fprintf(output, "\n");
}
if ((verbose_mode > 1 || !verbose_mode) && lstopo_show_only == (hwloc_obj_type_t)-1) {
hwloc_lstopo_show_summary(output, topology);
}
if (verbose_mode > 1 && lstopo_show_only == (hwloc_obj_type_t)-1) {
const struct hwloc_distances_s * distances;
unsigned depth;
for (depth = 0; depth < topodepth; depth++) {
distances = hwloc_get_whole_distance_matrix_by_depth(topology, depth);
if (!distances || !distances->latency)
continue;
printf("latency matrix between %ss (depth %u) by %s indexes:\n",
hwloc_obj_type_string(hwloc_get_depth_type(topology, depth)),
depth,
logical ? "logical" : "physical");
hwloc_utils_print_distance_matrix(topology, hwloc_get_root_obj(topology), distances->nbobjs, depth, distances->latency, logical);
}
}
if (verbose_mode > 1 && lstopo_show_only == (hwloc_obj_type_t)-1) {
hwloc_const_bitmap_t complete = hwloc_topology_get_complete_cpuset(topology);
hwloc_const_bitmap_t topo = hwloc_topology_get_topology_cpuset(topology);
hwloc_const_bitmap_t online = hwloc_topology_get_online_cpuset(topology);
hwloc_const_bitmap_t allowed = hwloc_topology_get_allowed_cpuset(topology);
if (complete && !hwloc_bitmap_isequal(topo, complete)) {
hwloc_bitmap_t unknown = hwloc_bitmap_alloc();
char *unknownstr;
hwloc_bitmap_copy(unknown, complete);
hwloc_bitmap_andnot(unknown, unknown, topo);
hwloc_bitmap_asprintf(&unknownstr, unknown);
fprintf (output, "%d processors not represented in topology: %s\n", hwloc_bitmap_weight(unknown), unknownstr);
free(unknownstr);
hwloc_bitmap_free(unknown);
}
if (complete && !hwloc_bitmap_isequal(online, complete)) {
hwloc_bitmap_t offline = hwloc_bitmap_alloc();
char *offlinestr;
hwloc_bitmap_copy(offline, complete);
hwloc_bitmap_andnot(offline, offline, online);
hwloc_bitmap_asprintf(&offlinestr, offline);
fprintf (output, "%d processors offline: %s\n", hwloc_bitmap_weight(offline), offlinestr);
free(offlinestr);
hwloc_bitmap_free(offline);
}
if (complete && !hwloc_bitmap_isequal(allowed, online)) {
if (!hwloc_bitmap_isincluded(online, allowed)) {
hwloc_bitmap_t forbidden = hwloc_bitmap_alloc();
char *forbiddenstr;
hwloc_bitmap_copy(forbidden, online);
hwloc_bitmap_andnot(forbidden, forbidden, allowed);
hwloc_bitmap_asprintf(&forbiddenstr, forbidden);
fprintf(output, "%d processors online but not allowed: %s\n", hwloc_bitmap_weight(forbidden), forbiddenstr);
free(forbiddenstr);
hwloc_bitmap_free(forbidden);
}
if (!hwloc_bitmap_isincluded(allowed, online)) {
hwloc_bitmap_t potential = hwloc_bitmap_alloc();
char *potentialstr;
hwloc_bitmap_copy(potential, allowed);
hwloc_bitmap_andnot(potential, potential, online);
hwloc_bitmap_asprintf(&potentialstr, potential);
fprintf(output, "%d processors allowed but not online: %s\n", hwloc_bitmap_weight(potential), potentialstr);
free(potentialstr);
hwloc_bitmap_free(potential);
}
}
if (!hwloc_topology_is_thissystem(topology))
fprintf (output, "Topology not from this system\n");
}
if (output != stdout)
fclose(output);
}
static void
look_rset(int sdl, hwloc_obj_type_t type, struct hwloc_topology *topology, int level)
{
rsethandle_t rset, rad;
int i,maxcpus,j;
int nbnodes;
struct hwloc_obj *obj;
if ((topology->flags & HWLOC_TOPOLOGY_FLAG_WHOLE_SYSTEM))
rset = rs_alloc(RS_ALL);
else
rset = rs_alloc(RS_PARTITION);
rad = rs_alloc(RS_EMPTY);
nbnodes = rs_numrads(rset, sdl, 0);
if (nbnodes == -1) {
perror("rs_numrads");
return;
}
for (i = 0; i < nbnodes; i++) {
if (rs_getrad(rset, rad, sdl, i, 0)) {
fprintf(stderr,"rs_getrad(%d) failed: %s\n", i, strerror(errno));
continue;
}
if (!rs_getinfo(rad, R_NUMPROCS, 0))
continue;
/* It seems logical processors are numbered from 1 here, while the
* bindprocessor functions numbers them from 0... */
obj = hwloc_alloc_setup_object(type, i - (type == HWLOC_OBJ_PU));
obj->cpuset = hwloc_bitmap_alloc();
obj->os_level = sdl;
maxcpus = rs_getinfo(rad, R_MAXPROCS, 0);
for (j = 0; j < maxcpus; j++) {
if (rs_op(RS_TESTRESOURCE, rad, NULL, R_PROCS, j))
hwloc_bitmap_set(obj->cpuset, j);
}
switch(type) {
case HWLOC_OBJ_NODE:
obj->nodeset = hwloc_bitmap_alloc();
hwloc_bitmap_set(obj->nodeset, i);
obj->memory.local_memory = 0; /* TODO: odd, rs_getinfo(rad, R_MEMSIZE, 0) << 10 returns the total memory ... */
obj->memory.page_types_len = 2;
obj->memory.page_types = malloc(2*sizeof(*obj->memory.page_types));
memset(obj->memory.page_types, 0, 2*sizeof(*obj->memory.page_types));
obj->memory.page_types[0].size = hwloc_getpagesize();
#ifdef HAVE__SC_LARGE_PAGESIZE
obj->memory.page_types[1].size = sysconf(_SC_LARGE_PAGESIZE);
#endif
/* TODO: obj->memory.page_types[1].count = rs_getinfo(rset, R_LGPGFREE, 0) / hugepagesize */
break;
case HWLOC_OBJ_CACHE:
obj->attr->cache.size = _system_configuration.L2_cache_size;
obj->attr->cache.associativity = _system_configuration.L2_cache_asc;
obj->attr->cache.linesize = 0; /* TODO: ? */
obj->attr->cache.depth = 2;
obj->attr->cache.type = HWLOC_OBJ_CACHE_UNIFIED; /* FIXME? */
break;
case HWLOC_OBJ_GROUP:
obj->attr->group.depth = level;
break;
case HWLOC_OBJ_CORE:
{
hwloc_obj_t obj2, obj3;
obj2 = hwloc_alloc_setup_object(HWLOC_OBJ_CACHE, i);
obj2->cpuset = hwloc_bitmap_dup(obj->cpuset);
obj2->attr->cache.size = _system_configuration.dcache_size;
obj2->attr->cache.associativity = _system_configuration.dcache_asc;
obj2->attr->cache.linesize = _system_configuration.dcache_line;
obj2->attr->cache.depth = 1;
if (_system_configuration.cache_attrib & (1<<30)) {
/* Unified cache */
obj2->attr->cache.type = HWLOC_OBJ_CACHE_UNIFIED;
hwloc_debug("Adding an L1u cache for core %d\n", i);
hwloc_insert_object_by_cpuset(topology, obj2);
} else {
/* Separate Instruction and Data caches */
obj2->attr->cache.type = HWLOC_OBJ_CACHE_DATA;
hwloc_debug("Adding an L1d cache for core %d\n", i);
hwloc_insert_object_by_cpuset(topology, obj2);
obj3 = hwloc_alloc_setup_object(HWLOC_OBJ_CACHE, i);
obj3->cpuset = hwloc_bitmap_dup(obj->cpuset);
obj3->attr->cache.size = _system_configuration.icache_size;
obj3->attr->cache.associativity = _system_configuration.icache_asc;
obj3->attr->cache.linesize = _system_configuration.icache_line;
obj3->attr->cache.depth = 1;
obj3->attr->cache.type = HWLOC_OBJ_CACHE_INSTRUCTION;
hwloc_debug("Adding an L1i cache for core %d\n", i);
hwloc_insert_object_by_cpuset(topology, obj3);
}
break;
}
default:
break;
}
hwloc_debug_2args_bitmap("%s %d has cpuset %s\n",
hwloc_obj_type_string(type),
i, obj->cpuset);
hwloc_insert_object_by_cpuset(topology, obj);
}
rs_free(rset);
rs_free(rad);
}
static int
hwloc_look_windows(struct hwloc_backend *backend)
{
struct hwloc_topology *topology = backend->topology;
hwloc_bitmap_t groups_pu_set = NULL;
SYSTEM_INFO SystemInfo;
DWORD length;
if (topology->levels[0][0]->cpuset)
/* somebody discovered things */
return 0;
hwloc_alloc_obj_cpusets(topology->levels[0][0]);
GetSystemInfo(&SystemInfo);
if (!GetLogicalProcessorInformationExProc && GetLogicalProcessorInformationProc) {
PSYSTEM_LOGICAL_PROCESSOR_INFORMATION procInfo, tmpprocInfo;
unsigned id;
unsigned i;
struct hwloc_obj *obj;
hwloc_obj_type_t type;
length = 0;
procInfo = NULL;
while (1) {
if (GetLogicalProcessorInformationProc(procInfo, &length))
break;
if (GetLastError() != ERROR_INSUFFICIENT_BUFFER)
return -1;
tmpprocInfo = realloc(procInfo, length);
if (!tmpprocInfo) {
free(procInfo);
goto out;
}
procInfo = tmpprocInfo;
}
assert(!length || procInfo);
for (i = 0; i < length / sizeof(*procInfo); i++) {
/* Ignore unknown caches */
if (procInfo->Relationship == RelationCache
&& procInfo->Cache.Type != CacheUnified
&& procInfo->Cache.Type != CacheData
&& procInfo->Cache.Type != CacheInstruction)
continue;
id = -1;
switch (procInfo[i].Relationship) {
case RelationNumaNode:
type = HWLOC_OBJ_NUMANODE;
id = procInfo[i].NumaNode.NodeNumber;
break;
case RelationProcessorPackage:
type = HWLOC_OBJ_PACKAGE;
break;
case RelationCache:
type = HWLOC_OBJ_CACHE;
break;
case RelationProcessorCore:
type = HWLOC_OBJ_CORE;
break;
case RelationGroup:
default:
type = HWLOC_OBJ_GROUP;
break;
}
obj = hwloc_alloc_setup_object(type, id);
obj->cpuset = hwloc_bitmap_alloc();
hwloc_debug("%s#%u mask %lx\n", hwloc_obj_type_string(type), id, procInfo[i].ProcessorMask);
/* ProcessorMask is a ULONG_PTR */
hwloc_bitmap_set_ith_ULONG_PTR(obj->cpuset, 0, procInfo[i].ProcessorMask);
hwloc_debug_2args_bitmap("%s#%u bitmap %s\n", hwloc_obj_type_string(type), id, obj->cpuset);
switch (type) {
case HWLOC_OBJ_NUMANODE:
{
ULONGLONG avail;
obj->nodeset = hwloc_bitmap_alloc();
hwloc_bitmap_set(obj->nodeset, id);
if ((GetNumaAvailableMemoryNodeExProc && GetNumaAvailableMemoryNodeExProc(id, &avail))
|| (GetNumaAvailableMemoryNodeProc && GetNumaAvailableMemoryNodeProc(id, &avail)))
obj->memory.local_memory = avail;
obj->memory.page_types_len = 2;
obj->memory.page_types = malloc(2 * sizeof(*obj->memory.page_types));
memset(obj->memory.page_types, 0, 2 * sizeof(*obj->memory.page_types));
obj->memory.page_types_len = 1;
obj->memory.page_types[0].size = SystemInfo.dwPageSize;
#ifdef HAVE__SC_LARGE_PAGESIZE
obj->memory.page_types_len++;
obj->memory.page_types[1].size = sysconf(_SC_LARGE_PAGESIZE);
#endif
break;
}
case HWLOC_OBJ_CACHE:
obj->attr->cache.size = procInfo[i].Cache.Size;
obj->attr->cache.associativity = procInfo[i].Cache.Associativity == CACHE_FULLY_ASSOCIATIVE ? -1 : procInfo[i].Cache.Associativity ;
obj->attr->cache.linesize = procInfo[i].Cache.LineSize;
obj->attr->cache.depth = procInfo[i].Cache.Level;
switch (procInfo->Cache.Type) {
case CacheUnified:
obj->attr->cache.type = HWLOC_OBJ_CACHE_UNIFIED;
break;
case CacheData:
obj->attr->cache.type = HWLOC_OBJ_CACHE_DATA;
break;
case CacheInstruction:
obj->attr->cache.type = HWLOC_OBJ_CACHE_INSTRUCTION;
break;
default:
hwloc_free_unlinked_object(obj);
continue;
}
break;
case HWLOC_OBJ_GROUP:
obj->attr->group.depth = procInfo[i].Relationship == RelationGroup;
break;
default:
break;
}
hwloc_insert_object_by_cpuset(topology, obj);
}
free(procInfo);
}
if (GetLogicalProcessorInformationExProc) {
PSYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX procInfoTotal, tmpprocInfoTotal, procInfo;
unsigned id;
struct hwloc_obj *obj;
hwloc_obj_type_t type;
length = 0;
procInfoTotal = NULL;
while (1) {
if (GetLogicalProcessorInformationExProc(RelationAll, procInfoTotal, &length))
break;
if (GetLastError() != ERROR_INSUFFICIENT_BUFFER)
return -1;
tmpprocInfoTotal = realloc(procInfoTotal, length);
if (!tmpprocInfoTotal) {
free(procInfoTotal);
goto out;
}
procInfoTotal = tmpprocInfoTotal;
}
for (procInfo = procInfoTotal;
(void*) procInfo < (void*) ((uintptr_t) procInfoTotal + length);
procInfo = (void*) ((uintptr_t) procInfo + procInfo->Size)) {
unsigned num, i;
GROUP_AFFINITY *GroupMask;
/* Ignore unknown caches */
if (procInfo->Relationship == RelationCache
&& procInfo->Cache.Type != CacheUnified
&& procInfo->Cache.Type != CacheData
&& procInfo->Cache.Type != CacheInstruction)
continue;
id = -1;
switch (procInfo->Relationship) {
case RelationNumaNode:
type = HWLOC_OBJ_NUMANODE;
num = 1;
GroupMask = &procInfo->NumaNode.GroupMask;
id = procInfo->NumaNode.NodeNumber;
break;
case RelationProcessorPackage:
type = HWLOC_OBJ_PACKAGE;
num = procInfo->Processor.GroupCount;
GroupMask = procInfo->Processor.GroupMask;
break;
case RelationCache:
type = HWLOC_OBJ_CACHE;
num = 1;
GroupMask = &procInfo->Cache.GroupMask;
break;
case RelationProcessorCore:
type = HWLOC_OBJ_CORE;
num = procInfo->Processor.GroupCount;
GroupMask = procInfo->Processor.GroupMask;
break;
case RelationGroup:
/* So strange an interface... */
for (id = 0; id < procInfo->Group.ActiveGroupCount; id++) {
KAFFINITY mask;
obj = hwloc_alloc_setup_object(HWLOC_OBJ_GROUP, id);
obj->cpuset = hwloc_bitmap_alloc();
mask = procInfo->Group.GroupInfo[id].ActiveProcessorMask;
hwloc_debug("group %u %d cpus mask %lx\n", id,
procInfo->Group.GroupInfo[id].ActiveProcessorCount, mask);
/* KAFFINITY is ULONG_PTR */
hwloc_bitmap_set_ith_ULONG_PTR(obj->cpuset, id, mask);
hwloc_debug_2args_bitmap("group %u %d bitmap %s\n", id, procInfo->Group.GroupInfo[id].ActiveProcessorCount, obj->cpuset);
/* save the set of PUs so that we can create them at the end */
if (!groups_pu_set)
groups_pu_set = hwloc_bitmap_alloc();
hwloc_bitmap_or(groups_pu_set, groups_pu_set, obj->cpuset);
hwloc_insert_object_by_cpuset(topology, obj);
}
continue;
default:
/* Don't know how to get the mask. */
hwloc_debug("unknown relation %d\n", procInfo->Relationship);
continue;
}
obj = hwloc_alloc_setup_object(type, id);
obj->cpuset = hwloc_bitmap_alloc();
for (i = 0; i < num; i++) {
hwloc_debug("%s#%u %d: mask %d:%lx\n", hwloc_obj_type_string(type), id, i, GroupMask[i].Group, GroupMask[i].Mask);
/* GROUP_AFFINITY.Mask is KAFFINITY, which is ULONG_PTR */
hwloc_bitmap_set_ith_ULONG_PTR(obj->cpuset, GroupMask[i].Group, GroupMask[i].Mask);
}
hwloc_debug_2args_bitmap("%s#%u bitmap %s\n", hwloc_obj_type_string(type), id, obj->cpuset);
switch (type) {
case HWLOC_OBJ_NUMANODE:
{
ULONGLONG avail;
obj->nodeset = hwloc_bitmap_alloc();
hwloc_bitmap_set(obj->nodeset, id);
if ((GetNumaAvailableMemoryNodeExProc && GetNumaAvailableMemoryNodeExProc(id, &avail))
|| (GetNumaAvailableMemoryNodeProc && GetNumaAvailableMemoryNodeProc(id, &avail)))
obj->memory.local_memory = avail;
obj->memory.page_types = malloc(2 * sizeof(*obj->memory.page_types));
memset(obj->memory.page_types, 0, 2 * sizeof(*obj->memory.page_types));
obj->memory.page_types_len = 1;
obj->memory.page_types[0].size = SystemInfo.dwPageSize;
#ifdef HAVE__SC_LARGE_PAGESIZE
obj->memory.page_types_len++;
obj->memory.page_types[1].size = sysconf(_SC_LARGE_PAGESIZE);
#endif
break;
}
case HWLOC_OBJ_CACHE:
obj->attr->cache.size = procInfo->Cache.CacheSize;
obj->attr->cache.associativity = procInfo->Cache.Associativity == CACHE_FULLY_ASSOCIATIVE ? -1 : procInfo->Cache.Associativity ;
obj->attr->cache.linesize = procInfo->Cache.LineSize;
obj->attr->cache.depth = procInfo->Cache.Level;
switch (procInfo->Cache.Type) {
case CacheUnified:
obj->attr->cache.type = HWLOC_OBJ_CACHE_UNIFIED;
break;
case CacheData:
obj->attr->cache.type = HWLOC_OBJ_CACHE_DATA;
break;
case CacheInstruction:
obj->attr->cache.type = HWLOC_OBJ_CACHE_INSTRUCTION;
break;
default:
hwloc_free_unlinked_object(obj);
continue;
}
break;
default:
break;
}
hwloc_insert_object_by_cpuset(topology, obj);
}
free(procInfoTotal);
}
if (groups_pu_set) {
/* the system supports multiple Groups.
* PU indexes may be discontiguous, especially if Groups contain less than 64 procs.
*/
hwloc_obj_t obj;
unsigned idx;
hwloc_bitmap_foreach_begin(idx, groups_pu_set) {
obj = hwloc_alloc_setup_object(HWLOC_OBJ_PU, idx);
obj->cpuset = hwloc_bitmap_alloc();
hwloc_bitmap_only(obj->cpuset, idx);
hwloc_debug_1arg_bitmap("cpu %u has cpuset %s\n",
idx, obj->cpuset);
hwloc_insert_object_by_cpuset(topology, obj);
} hwloc_bitmap_foreach_end();
hwloc_bitmap_free(groups_pu_set);
} else {
