static void check(unsigned nbnodes, unsigned nbcores, unsigned nbpus)
{
unsigned depth;
unsigned nb;
unsigned long long total_memory;
/* sanity checks */
depth = hwloc_topology_get_depth(topology);
assert(depth == 4);
depth = hwloc_get_type_depth(topology, HWLOC_OBJ_NODE);
assert(depth == 1);
depth = hwloc_get_type_depth(topology, HWLOC_OBJ_CORE);
assert(depth == 2);
depth = hwloc_get_type_depth(topology, HWLOC_OBJ_PU);
assert(depth == 3);
/* actual checks */
nb = hwloc_get_nbobjs_by_type(topology, HWLOC_OBJ_NODE);
assert(nb == nbnodes);
nb = hwloc_get_nbobjs_by_type(topology, HWLOC_OBJ_CORE);
assert(nb == nbcores);
nb = hwloc_get_nbobjs_by_type(topology, HWLOC_OBJ_PU);
assert(nb == nbpus);
total_memory = hwloc_get_root_obj(topology)->memory.total_memory;
assert(total_memory == nbnodes * 1024*1024*1024); /* synthetic topology puts 1GB per node */
}
int main(void)
{
hwloc_topology_t topology;
hwloc_obj_t obj;
hwloc_topology_init(&topology);
hwloc_topology_set_io_types_filter(topology, HWLOC_TYPE_FILTER_KEEP_ALL);
hwloc_topology_load(topology);
printf("Found %d bridges\n", hwloc_get_nbobjs_by_type(topology, HWLOC_OBJ_BRIDGE));
obj = NULL;
while ((obj = hwloc_get_next_bridge(topology, obj)) != NULL) {
assert(obj->type == HWLOC_OBJ_BRIDGE);
/* only host->pci and pci->pci bridge supported so far */
if (obj->attr->bridge.upstream_type == HWLOC_OBJ_BRIDGE_HOST) {
assert(obj->attr->bridge.downstream_type == HWLOC_OBJ_BRIDGE_PCI);
printf(" Found host->PCI bridge for domain %04x bus %02x-%02x\n",
obj->attr->bridge.downstream.pci.domain,
obj->attr->bridge.downstream.pci.secondary_bus,
obj->attr->bridge.downstream.pci.subordinate_bus);
} else {
assert(obj->attr->bridge.upstream_type == HWLOC_OBJ_BRIDGE_PCI);
assert(obj->attr->bridge.downstream_type == HWLOC_OBJ_BRIDGE_PCI);
printf(" Found PCI->PCI bridge [%04x:%04x] for domain %04x bus %02x-%02x\n",
obj->attr->bridge.upstream.pci.vendor_id,
obj->attr->bridge.upstream.pci.device_id,
obj->attr->bridge.downstream.pci.domain,
obj->attr->bridge.downstream.pci.secondary_bus,
obj->attr->bridge.downstream.pci.subordinate_bus);
}
}
printf("Found %d PCI devices\n", hwloc_get_nbobjs_by_type(topology, HWLOC_OBJ_PCI_DEVICE));
obj = NULL;
while ((obj = hwloc_get_next_pcidev(topology, obj)) != NULL) {
assert(obj->type == HWLOC_OBJ_PCI_DEVICE);
printf(" Found PCI device class %04x vendor %04x model %04x\n",
obj->attr->pcidev.class_id, obj->attr->pcidev.vendor_id, obj->attr->pcidev.device_id);
}
printf("Found %d OS devices\n", hwloc_get_nbobjs_by_type(topology, HWLOC_OBJ_OS_DEVICE));
obj = NULL;
while ((obj = hwloc_get_next_osdev(topology, obj)) != NULL) {
assert(obj->type == HWLOC_OBJ_OS_DEVICE);
printf(" Found OS device %s subtype %d\n", obj->name, obj->attr->osdev.type);
}
assert(HWLOC_TYPE_DEPTH_BRIDGE == hwloc_get_type_depth(topology, HWLOC_OBJ_BRIDGE));
assert(HWLOC_TYPE_DEPTH_PCI_DEVICE == hwloc_get_type_depth(topology, HWLOC_OBJ_PCI_DEVICE));
assert(HWLOC_TYPE_DEPTH_OS_DEVICE == hwloc_get_type_depth(topology, HWLOC_OBJ_OS_DEVICE));
assert(hwloc_compare_types(HWLOC_OBJ_BRIDGE, HWLOC_OBJ_PCI_DEVICE) < 0);
assert(hwloc_compare_types(HWLOC_OBJ_BRIDGE, HWLOC_OBJ_OS_DEVICE) < 0);
assert(hwloc_compare_types(HWLOC_OBJ_PCI_DEVICE, HWLOC_OBJ_OS_DEVICE) < 0);
hwloc_topology_destroy(topology);
return 0;
}
int
hwloc_type_sscanf_as_depth(const char *string, hwloc_obj_type_t *typep,
hwloc_topology_t topology, int *depthp)
{
union hwloc_obj_attr_u attr;
hwloc_obj_type_t type;
int depth;
int err;
err = hwloc_type_sscanf(string, &type, &attr, sizeof(attr));
if (err < 0)
return err;
depth = hwloc_get_type_depth(topology, type);
if (type == HWLOC_OBJ_GROUP
&& depth == HWLOC_TYPE_DEPTH_MULTIPLE
&& attr.group.depth != (unsigned)-1) {
unsigned l;
depth = HWLOC_TYPE_DEPTH_UNKNOWN;
for(l=0; l<topology->nb_levels; l++) {
if (topology->levels[l][0]->type == HWLOC_OBJ_GROUP
&& topology->levels[l][0]->attr->group.depth == attr.group.depth) {
depth = l;
break;
}
}
}
if (typep)
*typep = type;
*depthp = (unsigned) depth;
return 0;
}
static int
hwloc_solaris_get_sth_membind(hwloc_topology_t topology, idtype_t idtype, id_t id, hwloc_nodeset_t nodeset, hwloc_membind_policy_t *policy, int flags __hwloc_attribute_unused)
{
int depth = hwloc_get_type_depth(topology, HWLOC_OBJ_NUMANODE);
int n;
int i;
if (depth < 0) {
errno = ENOSYS;
return -1;
}
hwloc_bitmap_zero(nodeset);
n = hwloc_get_nbobjs_by_depth(topology, depth);
for (i = 0; i < n; i++) {
hwloc_obj_t obj = hwloc_get_obj_by_depth(topology, depth, i);
lgrp_affinity_t aff = lgrp_affinity_get(idtype, id, obj->os_index);
if (aff == LGRP_AFF_STRONG)
hwloc_bitmap_set(nodeset, obj->os_index);
}
if (hwloc_bitmap_iszero(nodeset))
hwloc_bitmap_copy(nodeset, hwloc_topology_get_complete_nodeset(topology));
*policy = HWLOC_MEMBIND_BIND;
return 0;
}
static void
output_compute_pu_min_textwidth(struct lstopo_output *output)
{
unsigned fontsize = output->fontsize;
char text[64];
int n;
hwloc_topology_t topology = output->topology;
hwloc_obj_t lastpu;
if (!output->methods->textsize) {
output->min_pu_textwidth = 0;
return;
}
if (output->logical) {
int depth = hwloc_get_type_depth(topology, HWLOC_OBJ_PU);
lastpu = hwloc_get_obj_by_depth(topology, depth, hwloc_get_nbobjs_by_depth(topology, depth)-1);
} else {
unsigned lastidx = hwloc_bitmap_last(hwloc_topology_get_topology_cpuset(topology));
lastpu = hwloc_get_pu_obj_by_os_index(topology, lastidx);
}
n = lstopo_obj_snprintf(output, text, sizeof(text), lastpu);
output->min_pu_textwidth = get_textwidth(output, text, n, fontsize);
}
/*
* Distribute cpus to the task using block distribution
*/
static int _task_cgroup_cpuset_dist_block(
hwloc_topology_t topology, hwloc_obj_type_t hwtype,
hwloc_obj_type_t req_hwtype, uint32_t nobj,
slurmd_job_t *job, int bind_verbose, hwloc_bitmap_t cpuset)
{
hwloc_obj_t obj;
uint32_t i, pfirst,plast;
uint32_t taskid = job->envtp->localid;
int hwdepth;
if (bind_verbose)
info("task/cgroup: task[%u] using block distribution, "
"task_dist %u", taskid, job->task_dist);
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);
_add_cpuset(hwtype, req_hwtype, obj, taskid, bind_verbose,
cpuset);
}
return XCGROUP_SUCCESS;
}
int PlatformTopology::num_domain(int domain_type) const
{
int result = 0;
try {
result = hwloc_get_nbobjs_by_type(m_topo, hwloc_domain(domain_type));
}
catch (const Exception &ex) {
if (ex.err_value() != GEOPM_ERROR_INVALID) {
throw ex;
}
if (domain_type == GEOPM_DOMAIN_TILE) {
/// @todo This assumes that tiles are just below
/// package in hwloc hierarchy. If tiles are
/// at L2 cache, but processor has an L3 cache,
/// this may not be correct.
int depth = hwloc_get_type_depth(m_topo, hwloc_domain(GEOPM_DOMAIN_PACKAGE)) + 1;
result = hwloc_get_nbobjs_by_depth(m_topo, depth);
}
else {
throw ex;
}
if (result == 0) {
throw ex;
}
}
return result;
}
static int
hwloc_solaris_get_sth_cpubind(hwloc_topology_t topology, idtype_t idtype, id_t id, hwloc_bitmap_t hwloc_set, int flags __hwloc_attribute_unused)
{
processorid_t binding;
int depth = hwloc_get_type_depth(topology, HWLOC_OBJ_NUMANODE);
int n;
int i;
if (depth < 0) {
errno = ENOSYS;
return -1;
}
/* first check if processor_bind() was used to bind to a single processor rather than to an lgroup */
if ( processor_bind(idtype, id, PBIND_QUERY, &binding) == 0 && binding != PBIND_NONE ) {
hwloc_bitmap_only(hwloc_set, binding);
return 0;
}
/* if not, check lgroups */
hwloc_bitmap_zero(hwloc_set);
n = hwloc_get_nbobjs_by_depth(topology, depth);
for (i = 0; i < n; i++) {
hwloc_obj_t obj = hwloc_get_obj_by_depth(topology, depth, i);
lgrp_affinity_t aff = lgrp_affinity_get(idtype, id, obj->os_index);
if (aff == LGRP_AFF_STRONG)
hwloc_bitmap_or(hwloc_set, hwloc_set, obj->cpuset);
}
if (hwloc_bitmap_iszero(hwloc_set))
hwloc_bitmap_copy(hwloc_set, hwloc_topology_get_complete_cpuset(topology));
return 0;
}
int
hwloc_get_obj_depth_by_name(hwloc_topology_t topology, char * obj_name){
/* find hwloc obj depth */
hwloc_obj_type_t type;
int depthattrp;
hwloc_obj_cache_type_t cache_type;
if(hwloc_obj_type_sscanf(obj_name,&type,&depthattrp,&cache_type,sizeof(cache_type))==-1){
fprintf(stderr,"type \"%s\" was not recognized\n",obj_name);
return -1;
}
int depth = hwloc_get_type_depth(topology,type);
if(depth==HWLOC_TYPE_DEPTH_MULTIPLE){
if(type==HWLOC_OBJ_CACHE){
depth = hwloc_get_cache_type_depth(topology,depthattrp,cache_type);
if(depth == HWLOC_TYPE_DEPTH_UNKNOWN){
fprintf(stderr,"type %s cannot be found, level=%d\n",obj_name,depthattrp);
return -1;
}
if(depth == HWLOC_TYPE_DEPTH_MULTIPLE){
fprintf(stderr,"type %s multiple caches match for\n",obj_name);
return -1;
}
}
else{
fprintf(stderr,"type \"%s\" isn't handled...\n",obj_name);
return -1;
}
}
return depth;
}
int compute_context_nbr(int * ctxnbr, int threadnbr, int verbose)
{
#if (defined WITH_STARPU && defined STARPU_CONTEXT)
if (*ctxnbr == -1)
{
int depth;
unsigned i, n;
int ncpu_per_socket, nsocket, ncpu;
hwloc_topology_t topology;
hwloc_obj_t obj;
/* Allocate and initialize topology object. */
hwloc_topology_init(&topology);
/* ... Optionally, put detection configuration here to ignore
some objects types, define a synthetic topology, etc....
The default is to detect all the objects of the machine that
the caller is allowed to access. See Configure Topology
Detection. */
/* Perform the topology detection. */
hwloc_topology_load(topology);
depth = hwloc_get_type_depth(topology, HWLOC_OBJ_SOCKET);
if (depth == HWLOC_TYPE_DEPTH_UNKNOWN) {
/* number of socket is unknow, let say we have quadcore... */
ncpu_per_socket = 4;
} else {
ncpu = hwloc_get_nbobjs_by_type(topology, HWLOC_OBJ_CORE);
nsocket = hwloc_get_nbobjs_by_depth(topology, depth);
ncpu_per_socket = ncpu/nsocket;
}
if (verbose > API_VERBOSE_NO)
fprintf(stdout, "ncpu_per_socket %d\n", ncpu_per_socket);
nsocket = threadnbr/ncpu_per_socket;
if (threadnbr%ncpu_per_socket)
nsocket ++;
*ctxnbr = nsocket + 1;
hwloc_topology_destroy(topology);
}
if (threadnbr + 1 < *ctxnbr)
*ctxnbr = threadnbr+1;
if (*ctxnbr == 1)
*ctxnbr = 2;
/* can't have more than STARPU_NMAX_SCHED_CTXS CTX */
{
PASTIX_INT nctx_bot = *ctxnbr - 1;
while (*ctxnbr > STARPU_NMAX_SCHED_CTXS)
{
nctx_bot /= 2;
*ctxnbr = nctx_bot + 1;
}
}
#endif /* WITH_STARPU */
return 0;
}
magma_int_t magma_get_parallel_numthreads()
{
// query number of cores
magma_int_t ncores = 0;
#ifdef HAVE_HWLOC
// hwloc gives physical cores, not hyperthreads
// from http://stackoverflow.com/questions/12483399/getting-number-of-cores-not-ht-threads
hwloc_topology_t topology;
hwloc_topology_init( &topology );
hwloc_topology_load( topology );
magma_int_t depth = hwloc_get_type_depth( topology, HWLOC_OBJ_CORE );
if (depth != HWLOC_TYPE_DEPTH_UNKNOWN) {
ncores = hwloc_get_nbobjs_by_depth( topology, depth );
}
hwloc_topology_destroy( topology );
#endif
if ( ncores == 0 ) {
#ifdef _MSC_VER // Windows
SYSTEM_INFO sysinfo;
GetSystemInfo( &sysinfo );
ncores = sysinfo.dwNumberOfProcessors;
#else
ncores = sysconf( _SC_NPROCESSORS_ONLN );
#endif
}
// query MAGMA_NUM_THREADS or OpenMP
const char *threads_str = getenv("MAGMA_NUM_THREADS");
magma_int_t threads = 0;
if ( threads_str != NULL ) {
char* endptr;
threads = strtol( threads_str, &endptr, 10 );
if ( threads < 1 || *endptr != '\0' ) {
threads = 1;
fprintf( stderr, "$MAGMA_NUM_THREADS='%s' is an invalid number; using %d thread.\n",
threads_str, (int) threads );
}
}
else {
#if defined(_OPENMP)
#pragma omp parallel
{
threads = omp_get_num_threads();
}
#else
threads = ncores;
#endif
}
// limit to range [1, number of cores]
threads = max( 1, min( ncores, threads ));
return threads;
}
void
check_map_event_obj(hwloc_topology_t topology, char * obj_name, char * event_name)
{
PAPI_event_info_t info;
int err, eventcode;
hwloc_obj_t obj;
unsigned depth = hwloc_get_obj_depth_by_name(topology,obj_name);
if((err = PAPI_event_name_to_code(event_name,&eventcode))!=PAPI_OK){
handle_error(err);
fprintf(stderr,"could not get \"%s\" event infos\n",event_name);
}
PAPI_get_event_info(eventcode,&info);
switch(info.location){
case PAPI_LOCATION_UNCORE:
if(depth >= hwloc_get_type_depth(topology, HWLOC_OBJ_CORE)){
fprintf(stderr,"event %s is an UNCORE event and is actually mapped on %s which is under HWLOC_OBJ_CORE\n",
event_name, obj_name);
}
break;
case PAPI_LOCATION_CPU:
if(depth > hwloc_get_type_depth(topology, HWLOC_OBJ_MACHINE)){
fprintf(stderr,"event %s is a CPU event and is actually mapped on %s which is strictly under HWLOC_OBJ_MACHINE\n",
event_name, obj_name);
}
break;
case PAPI_LOCATION_PACKAGE:
if(depth > hwloc_get_type_depth(topology, HWLOC_OBJ_SOCKET)){
fprintf(stderr,"event %s is a PACKAGE event and is actually mapped on %s which is strictly under HWLOC_OBJ_SOCKET\n",
event_name, obj_name);
}
break;
default:
break;
}
}
static int _get_ldom_sched_cpuset(hwloc_topology_t topology,
hwloc_obj_type_t hwtype, hwloc_obj_type_t req_hwtype,
uint32_t ldom, cpu_set_t *mask)
{
hwloc_obj_t obj;
hwloc_bitmap_t cpuset;
int hwdepth;
cpuset = hwloc_bitmap_alloc();
hwdepth = hwloc_get_type_depth(topology, hwtype);
obj = hwloc_get_obj_by_depth(topology, hwdepth, ldom);
_add_hwloc_cpuset(hwtype, req_hwtype, obj, 0, 0, cpuset);
hwloc_cpuset_to_glibc_sched_affinity(topology, cpuset,
mask, sizeof(cpu_set_t));
hwloc_bitmap_free(cpuset);
return true;
}
int get_max_objs_inside_cpuset_by_type(hwloc_topology_t topology, hwloc_cpuset_t cpuset, hwloc_obj_type_t type){
int depth = hwloc_get_type_depth(topology, type);
if(depth == HWLOC_TYPE_DEPTH_UNKNOWN){
fprintf(stderr, "Cannot find depth %s\n", hwloc_type_name(type));
return -1;
}
if(depth == HWLOC_TYPE_DEPTH_MULTIPLE){
hwloc_obj_t deepest_of_type = hwloc_get_obj_inside_cpuset_by_type(topology, cpuset, HWLOC_OBJ_PU,0);
while(deepest_of_type !=NULL && deepest_of_type->type != type)
deepest_of_type = deepest_of_type->parent;
if(deepest_of_type == NULL)
return -1;
else
depth = deepest_of_type->depth;
}
return hwloc_get_nbobjs_inside_cpuset_by_depth(topology, cpuset, depth);
}
static int
hwloc_aix_get_sth_membind(hwloc_topology_t topology, rstype_t what, rsid_t who, hwloc_bitmap_t nodeset, hwloc_membind_policy_t *policy, int flags __hwloc_attribute_unused)
{
hwloc_bitmap_t hwloc_set;
rsethandle_t rset;
unsigned cpu, maxcpus;
int res = -1;
int depth, n, i;
depth = hwloc_get_type_depth(topology, HWLOC_OBJ_NUMANODE);
if (depth < 0) {
errno = EXDEV;
return -1;
}
n = hwloc_get_nbobjs_by_depth(topology, depth);
rset = rs_alloc(RS_EMPTY);
if (ra_getrset(what, who, 0, rset) == -1)
goto out;
hwloc_set = hwloc_bitmap_alloc();
maxcpus = rs_getinfo(rset, R_MAXPROCS, 0);
for (cpu = 0; cpu < maxcpus; cpu++)
if (rs_op(RS_TESTRESOURCE, rset, NULL, R_PROCS, cpu) == 1)
hwloc_bitmap_set(hwloc_set, cpu);
hwloc_bitmap_and(hwloc_set, hwloc_set, hwloc_topology_get_complete_cpuset(topology));
hwloc_bitmap_zero(nodeset);
for (i = 0; i < n; i++) {
hwloc_obj_t obj = hwloc_get_obj_by_depth(topology, depth, i);
if (hwloc_bitmap_isincluded(obj->cpuset, hwloc_set))
hwloc_bitmap_set(nodeset, obj->os_index);
}
hwloc_bitmap_free(hwloc_set);
*policy = HWLOC_MEMBIND_BIND;
res = 0;
out:
rs_free(rset);
return res;
}
static int
dss_topo_init()
{
hwloc_topology_init(&dss_topo);
hwloc_topology_load(dss_topo);
dss_core_depth = hwloc_get_type_depth(dss_topo, HWLOC_OBJ_CORE);
dss_core_nr = hwloc_get_nbobjs_by_type(dss_topo, HWLOC_OBJ_CORE);
dss_tgt_nr = dss_tgt_nr_get(dss_core_nr, nr_threads);
if (dss_core_offset < 0 || dss_core_offset >= dss_core_nr) {
D_ERROR("invalid dss_core_offset %d (set by \"-f\" option), "
"should within range [0, %d]", dss_core_offset,
dss_core_nr - 1);
return -DER_INVAL;
}
return 0;
}
static int
hwloc_solaris_set_sth_membind(hwloc_topology_t topology, idtype_t idtype, id_t id, hwloc_const_nodeset_t nodeset, hwloc_membind_policy_t policy, int flags)
{
int depth;
int n, i;
switch (policy) {
case HWLOC_MEMBIND_DEFAULT:
case HWLOC_MEMBIND_BIND:
break;
default:
errno = ENOSYS;
return -1;
}
if (flags & HWLOC_MEMBIND_NOCPUBIND) {
errno = ENOSYS;
return -1;
}
depth = hwloc_get_type_depth(topology, HWLOC_OBJ_NUMANODE);
if (depth < 0) {
errno = EXDEV;
return -1;
}
n = hwloc_get_nbobjs_by_depth(topology, depth);
for (i = 0; i < n; i++) {
hwloc_obj_t obj = hwloc_get_obj_by_depth(topology, depth, i);
if (hwloc_bitmap_isset(nodeset, obj->os_index)) {
lgrp_affinity_set(idtype, id, obj->os_index, LGRP_AFF_STRONG);
} else {
if (flags & HWLOC_CPUBIND_STRICT)
lgrp_affinity_set(idtype, id, obj->os_index, LGRP_AFF_NONE);
else
lgrp_affinity_set(idtype, id, obj->os_index, LGRP_AFF_WEAK);
}
}
return 0;
}
void Hwloc::getNumSockets(unsigned int &allowedNodes, int &numSockets, unsigned int &hwThreads) {
#ifdef HWLOC
numSockets = 0;
// Nodes that can be seen by hwloc
allowedNodes = 0;
// Hardware threads
hwThreads = 0;
int depth = hwloc_get_type_depth( _hwlocTopology, HWLOC_OBJ_NODE );
// If there are NUMA nodes in this machine
if ( depth != HWLOC_TYPE_DEPTH_UNKNOWN ) {
//hwloc_const_cpuset_t cpuset = hwloc_topology_get_online_cpuset( _hwlocTopology );
//allowedNodes = hwloc_get_nbobjs_inside_cpuset_by_type( _hwlocTopology, cpuset, HWLOC_OBJ_NODE );
//hwThreads = hwloc_get_nbobjs_inside_cpuset_by_type( _hwlocTopology, cpuset, HWLOC_OBJ_PU );
unsigned nodes = hwloc_get_nbobjs_by_depth( _hwlocTopology, depth );
//hwloc_cpuset_t set = i
// For each node, count how many hardware threads there are below.
for ( unsigned nodeIdx = 0; nodeIdx < nodes; ++nodeIdx )
{
hwloc_obj_t node = hwloc_get_obj_by_depth( _hwlocTopology, depth, nodeIdx );
int localThreads = hwloc_get_nbobjs_inside_cpuset_by_type( _hwlocTopology, node->cpuset, HWLOC_OBJ_PU );
// Increase hw thread count
hwThreads += localThreads;
// If this node has hw threads beneath, increase the number of viewable nodes
if ( localThreads > 0 ) ++allowedNodes;
}
numSockets = nodes;
}
// Otherwise, set it to 1
else {
allowedNodes = 1;
numSockets = 1;
}
#else
numSockets = 0;
allowedNodes = 0;
#endif
}
int main(void)
{
hwloc_topology_t topology;
hwloc_obj_t obj;
unsigned indexes[5];
float distances[5*5];
unsigned depth;
unsigned width;
/* group 3 numa nodes as 1 group of 2 and 1 on the side */
hwloc_topology_init(&topology);
hwloc_topology_set_synthetic(topology, "node:3 pu:1");
indexes[0] = 0; indexes[1] = 1; indexes[2] = 2;
distances[0] = 1; distances[1] = 4; distances[2] = 4;
distances[3] = 4; distances[4] = 1; distances[5] = 2;
distances[6] = 4; distances[7] = 2; distances[8] = 1;
hwloc_topology_set_distance_matrix(topology, HWLOC_OBJ_PU, 3, indexes, distances);
hwloc_topology_load(topology);
/* 2 groups at depth 1 */
depth = hwloc_get_type_depth(topology, HWLOC_OBJ_GROUP);
assert(depth == 1);
width = hwloc_get_nbobjs_by_depth(topology, depth);
assert(width == 1);
/* 3 node at depth 2 */
depth = hwloc_get_type_depth(topology, HWLOC_OBJ_NUMANODE);
assert(depth == 2);
width = hwloc_get_nbobjs_by_depth(topology, depth);
assert(width == 3);
/* find the root obj */
obj = hwloc_get_root_obj(topology);
assert(obj->arity == 2);
/* check its children */
assert(obj->children[0]->type == HWLOC_OBJ_NUMANODE);
assert(obj->children[0]->depth == 2);
assert(obj->children[0]->arity == 1);
assert(obj->children[1]->type == HWLOC_OBJ_GROUP);
assert(obj->children[1]->depth == 1);
assert(obj->children[1]->arity == 2);
hwloc_topology_destroy(topology);
/* group 5 packages as 2 group of 2 and 1 on the side, all of them below a common node object */
hwloc_topology_init(&topology);
hwloc_topology_set_synthetic(topology, "node:1 pack:5 pu:1");
indexes[0] = 0; indexes[1] = 1; indexes[2] = 2; indexes[3] = 3; indexes[4] = 4;
distances[ 0] = 1; distances[ 1] = 2; distances[ 2] = 4; distances[ 3] = 4; distances[ 4] = 4;
distances[ 5] = 2; distances[ 6] = 1; distances[ 7] = 4; distances[ 8] = 4; distances[ 9] = 4;
distances[10] = 4; distances[11] = 4; distances[12] = 1; distances[13] = 4; distances[14] = 4;
distances[15] = 4; distances[16] = 4; distances[17] = 4; distances[18] = 1; distances[19] = 2;
distances[20] = 4; distances[21] = 4; distances[22] = 4; distances[23] = 2; distances[24] = 1;
hwloc_topology_set_distance_matrix(topology, HWLOC_OBJ_PACKAGE, 5, indexes, distances);
hwloc_topology_load(topology);
/* 1 node at depth 1 */
depth = hwloc_get_type_depth(topology, HWLOC_OBJ_NUMANODE);
assert(depth == 1);
width = hwloc_get_nbobjs_by_depth(topology, depth);
assert(width == 1);
/* 2 groups at depth 2 */
depth = hwloc_get_type_depth(topology, HWLOC_OBJ_GROUP);
assert(depth == 2);
width = hwloc_get_nbobjs_by_depth(topology, depth);
assert(width == 2);
/* 5 packages at depth 3 */
depth = hwloc_get_type_depth(topology, HWLOC_OBJ_PACKAGE);
assert(depth == 3);
width = hwloc_get_nbobjs_by_depth(topology, depth);
assert(width == 5);
/* find the node obj */
obj = hwloc_get_root_obj(topology);
assert(obj->arity == 1);
obj = obj->children[0];
assert(obj->type == HWLOC_OBJ_NUMANODE);
assert(obj->arity == 3);
/* check its children */
assert(obj->children[0]->type == HWLOC_OBJ_GROUP);
assert(obj->children[0]->depth == 2);
assert(obj->children[0]->arity == 2);
assert(obj->children[1]->type == HWLOC_OBJ_PACKAGE);
assert(obj->children[1]->depth == 3);
assert(obj->children[1]->arity == 1);
assert(obj->children[2]->type == HWLOC_OBJ_GROUP);
assert(obj->children[2]->depth == 2);
assert(obj->children[2]->arity == 2);
hwloc_topology_destroy(topology);
return 0;
}
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;
}
int
pocl_topology_detect_device_info(cl_device_id device)
{
hwloc_topology_t pocl_topology;
int ret = 0;
#ifdef HWLOC_API_2
if (hwloc_get_api_version () < 0x20000)
POCL_MSG_ERR ("pocl was compiled against libhwloc 2.x but is"
"actually running against libhwloc 1.x \n");
#else
if (hwloc_get_api_version () >= 0x20000)
POCL_MSG_ERR ("pocl was compiled against libhwloc 1.x but is"
"actually running against libhwloc 2.x \n");
#endif
/*
* hwloc's OpenCL backend causes problems at the initialization stage
* because it reloads libpocl.so via the ICD loader.
*
* See: https://github.com/pocl/pocl/issues/261
*
* The only trick to stop hwloc from initializing the OpenCL plugin
* I could find is to point the plugin search path to a place where there
* are no plugins to be found.
*/
setenv ("HWLOC_PLUGINS_PATH", "/dev/null", 1);
ret = hwloc_topology_init (&pocl_topology);
if (ret == -1)
{
POCL_MSG_ERR ("Cannot initialize the topology.\n");
return ret;
}
#ifdef HWLOC_API_2
hwloc_topology_set_io_types_filter(pocl_topology, HWLOC_TYPE_FILTER_KEEP_NONE);
hwloc_topology_set_type_filter (pocl_topology, HWLOC_OBJ_SYSTEM, HWLOC_TYPE_FILTER_KEEP_NONE);
hwloc_topology_set_type_filter (pocl_topology, HWLOC_OBJ_GROUP, HWLOC_TYPE_FILTER_KEEP_NONE);
hwloc_topology_set_type_filter (pocl_topology, HWLOC_OBJ_BRIDGE, HWLOC_TYPE_FILTER_KEEP_NONE);
hwloc_topology_set_type_filter (pocl_topology, HWLOC_OBJ_MISC, HWLOC_TYPE_FILTER_KEEP_NONE);
hwloc_topology_set_type_filter (pocl_topology, HWLOC_OBJ_PCI_DEVICE, HWLOC_TYPE_FILTER_KEEP_NONE);
hwloc_topology_set_type_filter (pocl_topology, HWLOC_OBJ_OS_DEVICE, HWLOC_TYPE_FILTER_KEEP_NONE);
#else
hwloc_topology_ignore_type (pocl_topology, HWLOC_TOPOLOGY_FLAG_WHOLE_IO);
hwloc_topology_ignore_type (pocl_topology, HWLOC_OBJ_SYSTEM);
hwloc_topology_ignore_type (pocl_topology, HWLOC_OBJ_GROUP);
hwloc_topology_ignore_type (pocl_topology, HWLOC_OBJ_BRIDGE);
hwloc_topology_ignore_type (pocl_topology, HWLOC_OBJ_MISC);
hwloc_topology_ignore_type (pocl_topology, HWLOC_OBJ_PCI_DEVICE);
hwloc_topology_ignore_type (pocl_topology, HWLOC_OBJ_OS_DEVICE);
#endif
ret = hwloc_topology_load (pocl_topology);
if (ret == -1)
{
POCL_MSG_ERR ("Cannot load the topology.\n");
goto exit_destroy;
}
#ifdef HWLOC_API_2
device->global_mem_size =
hwloc_get_root_obj(pocl_topology)->total_memory;
#else
device->global_mem_size =
hwloc_get_root_obj(pocl_topology)->memory.total_memory;
#endif
// Try to get the number of CPU cores from topology
int depth = hwloc_get_type_depth(pocl_topology, HWLOC_OBJ_PU);
if(depth != HWLOC_TYPE_DEPTH_UNKNOWN)
device->max_compute_units = hwloc_get_nbobjs_by_depth(pocl_topology, depth);
/* Find information about global memory cache by looking at the first
* cache covering the first PU */
do {
size_t cache_size = 0, cacheline_size = 0;
hwloc_obj_t core
= hwloc_get_next_obj_by_type (pocl_topology, HWLOC_OBJ_CORE, NULL);
if (core)
{
hwloc_obj_t cache
= hwloc_get_shared_cache_covering_obj (pocl_topology, core);
if ((cache) && (cache->attr))
{
cacheline_size = cache->attr->cache.linesize;
cache_size = cache->attr->cache.size;
}
else
core = NULL; /* fallback to L1 cache size */
}
hwloc_obj_t pu
= hwloc_get_next_obj_by_type (pocl_topology, HWLOC_OBJ_PU, NULL);
if (!core && pu)
{
hwloc_obj_t cache
= hwloc_get_shared_cache_covering_obj (pocl_topology, pu);
if ((cache) && (cache->attr))
{
cacheline_size = cache->attr->cache.linesize;
cache_size = cache->attr->cache.size;
}
}
if (!cache_size || !cacheline_size)
break;
device->global_mem_cache_type
= 0x2; // CL_READ_WRITE_CACHE, without including all of CL/cl.h
device->global_mem_cacheline_size = cacheline_size;
device->global_mem_cache_size = cache_size;
} while (0);
// Destroy topology object and return
exit_destroy:
hwloc_topology_destroy (pocl_topology);
return ret;
}
/* 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 int bind_in_place(orte_job_t *jdata,
hwloc_obj_type_t target,
unsigned cache_level)
{
/* traverse the hwloc topology tree on each node downwards
* until we find an unused object of type target - and then bind
* the process to that target
*/
int i, j;
orte_job_map_t *map;
orte_node_t *node;
orte_proc_t *proc;
hwloc_cpuset_t cpus;
unsigned int idx, ncpus;
struct hwloc_topology_support *support;
opal_hwloc_obj_data_t *data;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"mca:rmaps: bind in place for job %s with bindings %s",
ORTE_JOBID_PRINT(jdata->jobid),
opal_hwloc_base_print_binding(jdata->map->binding));
/* initialize */
map = jdata->map;
for (i=0; i < map->nodes->size; i++) {
if (NULL == (node = (orte_node_t*)opal_pointer_array_get_item(map->nodes, i))) {
continue;
}
if (!orte_do_not_launch) {
/* if we don't want to launch, then we are just testing the system,
* so ignore questions about support capabilities
*/
support = (struct hwloc_topology_support*)hwloc_topology_get_support(node->topology);
/* check if topology supports cpubind - have to be careful here
* as Linux doesn't currently support thread-level binding. This
* may change in the future, though, and it isn't clear how hwloc
* interprets the current behavior. So check both flags to be sure.
*/
if (!support->cpubind->set_thisproc_cpubind &&
!support->cpubind->set_thisthread_cpubind) {
if (!OPAL_BINDING_REQUIRED(map->binding) ||
!OPAL_BINDING_POLICY_IS_SET(map->binding)) {
/* we are not required to bind, so ignore this */
continue;
}
orte_show_help("help-orte-rmaps-base.txt", "rmaps:cpubind-not-supported", true, node->name);
return ORTE_ERR_SILENT;
}
/* check if topology supports membind - have to be careful here
* as hwloc treats this differently than I (at least) would have
* expected. Per hwloc, Linux memory binding is at the thread,
* and not process, level. Thus, hwloc sets the "thisproc" flag
* to "false" on all Linux systems, and uses the "thisthread" flag
* to indicate binding capability - don't warn if the user didn't
* specifically request binding
*/
if (!support->membind->set_thisproc_membind &&
!support->membind->set_thisthread_membind &&
OPAL_BINDING_POLICY_IS_SET(map->binding)) {
if (OPAL_HWLOC_BASE_MBFA_WARN == opal_hwloc_base_mbfa && !membind_warned) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:membind-not-supported", true, node->name);
membind_warned = true;
} else if (OPAL_HWLOC_BASE_MBFA_ERROR == opal_hwloc_base_mbfa) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:membind-not-supported-fatal", true, node->name);
return ORTE_ERR_SILENT;
}
}
}
/* some systems do not report cores, and so we can get a situation where our
* default binding policy will fail for no necessary reason. So if we are
* computing a binding due to our default policy, and no cores are found
* on this node, just silently skip it - we will not bind
*/
if (!OPAL_BINDING_POLICY_IS_SET(map->binding) &&
HWLOC_TYPE_DEPTH_UNKNOWN == hwloc_get_type_depth(node->topology, HWLOC_OBJ_CORE)) {
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"Unable to bind-to core by default on node %s as no cores detected",
node->name);
continue;
}
/* we share topologies in order
* to save space, so we need to reset the usage info to reflect
* our own current state
*/
reset_usage(node, jdata->jobid);
/* cycle thru the procs */
for (j=0; j < node->procs->size; j++) {
if (NULL == (proc = (orte_proc_t*)opal_pointer_array_get_item(node->procs, j))) {
continue;
}
/* ignore procs from other jobs */
if (proc->name.jobid != jdata->jobid) {
continue;
}
/* ignore procs that have already been bound - should
* never happen, but safer
*/
if (NULL != proc->cpu_bitmap) {
continue;
}
/* get the index of this location */
if (UINT_MAX == (idx = opal_hwloc_base_get_obj_idx(node->topology, proc->locale, OPAL_HWLOC_AVAILABLE))) {
ORTE_ERROR_LOG(ORTE_ERR_BAD_PARAM);
return ORTE_ERR_SILENT;
}
/* track the number bound */
data = (opal_hwloc_obj_data_t*)proc->locale->userdata;
data->num_bound++;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"BINDING PROC %s TO %s NUMBER %u",
ORTE_NAME_PRINT(&proc->name),
hwloc_obj_type_string(proc->locale->type), idx);
/* get the number of cpus under this location */
if (0 == (ncpus = opal_hwloc_base_get_npus(node->topology, proc->locale))) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:no-available-cpus", true, node->name);
return ORTE_ERR_SILENT;
}
/* error out if adding a proc would cause overload and that wasn't allowed,
* and it wasn't a default binding policy (i.e., the user requested it)
*/
if (ncpus < data->num_bound &&
!OPAL_BIND_OVERLOAD_ALLOWED(jdata->map->binding) &&
OPAL_BINDING_POLICY_IS_SET(jdata->map->binding)) {
orte_show_help("help-orte-rmaps-base.txt", "rmaps:binding-overload", true,
opal_hwloc_base_print_binding(map->binding), node->name,
data->num_bound, ncpus);
return ORTE_ERR_SILENT;
}
/* bind the proc here */
cpus = opal_hwloc_base_get_available_cpus(node->topology, proc->locale);
hwloc_bitmap_list_asprintf(&proc->cpu_bitmap, cpus);
/* record the location */
proc->bind_location = proc->locale;
opal_output_verbose(5, orte_rmaps_base_framework.framework_output,
"%s BOUND PROC %s TO %s[%s:%u] on node %s",
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
ORTE_NAME_PRINT(&proc->name),
proc->cpu_bitmap,
hwloc_obj_type_string(proc->locale->type),
idx, node->name);
}
}
return ORTE_SUCCESS;
}
static int bind_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;
}
/* 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
}
static int _task_cgroup_cpuset_dist_block(
hwloc_topology_t topology, hwloc_obj_type_t hwtype,
hwloc_obj_type_t req_hwtype, uint32_t nobj,
stepd_step_rec_t *job, int bind_verbose, hwloc_bitmap_t cpuset)
{
hwloc_obj_t obj;
uint32_t core_loop, ntskip, npdist;
uint32_t i, j, pfirst, plast;
uint32_t taskid = job->envtp->localid;
int hwdepth;
uint32_t npus, ncores, nsockets;
int spec_thread_cnt = 0;
bitstr_t *spec_threads = NULL;
uint32_t *thread_idx;
uint32_t core_idx;
bool core_fcyclic, core_block;
nsockets = (uint32_t) hwloc_get_nbobjs_by_type(topology,
HWLOC_OBJ_SOCKET);
ncores = (uint32_t) hwloc_get_nbobjs_by_type(topology,
HWLOC_OBJ_CORE);
npus = (uint32_t) hwloc_get_nbobjs_by_type(topology,
HWLOC_OBJ_PU);
core_block = (job->task_dist & SLURM_DIST_COREMASK) ==
SLURM_DIST_COREBLOCK ? true : false;
core_fcyclic = (job->task_dist & SLURM_DIST_COREMASK) ==
SLURM_DIST_CORECFULL ? true : false;
thread_idx = xmalloc(ncores * sizeof(uint32_t));
if (bind_verbose) {
info("task/cgroup: task[%u] using block distribution, "
"task_dist 0x%x", taskid, job->task_dist);
}
if ((hwloc_compare_types(hwtype, HWLOC_OBJ_PU) == 0) && !core_block) {
thread_idx = xmalloc(ncores * sizeof(uint32_t));
ntskip = taskid;
npdist = job->cpus_per_task;
i = 0; j = 0;
core_idx = 0;
core_loop = 0;
while (i < ntskip + 1 && core_loop < npdist + 1) {
while ((core_idx < ncores) && (j < npdist)) {
obj = hwloc_get_obj_below_by_type(
topology, HWLOC_OBJ_CORE, core_idx,
hwtype, thread_idx[core_idx]);
if (obj != NULL) {
thread_idx[core_idx]++;
j++;
if (i == ntskip)
_add_hwloc_cpuset(hwtype,
req_hwtype, obj, taskid,
bind_verbose, cpuset);
if ((j < npdist) && core_fcyclic)
core_idx++;
} else {
core_idx++;
}
}
if (j == npdist) {
i++; j = 0;
core_idx++; // no validity check, handled by the while
core_loop = 0;
} else {
core_loop++;
core_idx = 0;
}
}
xfree(thread_idx);
/* should never happen in normal scenario */
if (core_loop > npdist) {
error("task/cgroup: task[%u] infinite loop broken while "
"trying to provision compute elements using %s",
taskid, format_task_dist_states(job->task_dist));
return XCGROUP_ERROR;
} else
return XCGROUP_SUCCESS;
}
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);
if ((job->job_core_spec != (uint16_t) NO_VAL) &&
(job->job_core_spec & CORE_SPEC_THREAD) &&
(job->job_core_spec != CORE_SPEC_THREAD) &&
(nsockets != 0)) {
/* Skip specialized threads as needed */
int i, t, c, s;
int cores = MAX(1, (ncores / nsockets));
int threads = npus / cores;
spec_thread_cnt = job->job_core_spec & (~CORE_SPEC_THREAD);
spec_threads = bit_alloc(npus);
for (t = threads - 1;
((t >= 0) && (spec_thread_cnt > 0)); t--) {
for (c = cores - 1;
((c >= 0) && (spec_thread_cnt > 0)); c--) {
for (s = nsockets - 1;
((s >= 0) && (spec_thread_cnt > 0)); s--) {
i = s * cores + c;
i = (i * threads) + t;
bit_set(spec_threads, i);
spec_thread_cnt--;
}
}
}
if (hwtype == HWLOC_OBJ_PU) {
for (i = 0; i <= pfirst && i < npus; i++) {
if (bit_test(spec_threads, i))
pfirst++;
};
}
}
for (i = pfirst; i <= plast && i < nobj ; i++) {
obj = hwloc_get_obj_by_depth(topology, hwdepth, (int)i);
_add_hwloc_cpuset(hwtype, req_hwtype, obj, taskid,
bind_verbose, cpuset);
}
if (spec_threads) {
for (i = 0; i < npus; i++) {
if (bit_test(spec_threads, i)) {
hwloc_bitmap_clr(cpuset, i);
}
};
FREE_NULL_BITMAP(spec_threads);
}
return XCGROUP_SUCCESS;
}
int main(void)
{
int depth;
unsigned i, n;
unsigned long size;
int levels;
char string[128];
int topodepth;
void *m;
hwloc_topology_t topology;
hwloc_cpuset_t cpuset;
hwloc_obj_t obj;
/* Allocate and initialize topology object. */
hwloc_topology_init(&topology);
/* ... Optionally, put detection configuration here to ignore
some objects types, define a synthetic topology, etc....
The default is to detect all the objects of the machine that
the caller is allowed to access. See Configure Topology
Detection. */
/* Perform the topology detection. */
hwloc_topology_load(topology);
/* Optionally, get some additional topology information
in case we need the topology depth later. */
topodepth = hwloc_topology_get_depth(topology);
/*****************************************************************
* First example:
* Walk the topology with an array style, from level 0 (always
* the system level) to the lowest level (always the proc level).
*****************************************************************/
for (depth = 0; depth < topodepth; depth++) {
printf("*** Objects at level %d\n", depth);
for (i = 0; i < hwloc_get_nbobjs_by_depth(topology, depth);
i++) {
hwloc_obj_type_snprintf(string, sizeof(string),
hwloc_get_obj_by_depth(topology, depth, i), 0);
printf("Index %u: %s\n", i, string);
}
}
/*****************************************************************
* Second example:
* Walk the topology with a tree style.
*****************************************************************/
printf("*** Printing overall tree\n");
print_children(topology, hwloc_get_root_obj(topology), 0);
/*****************************************************************
* Third example:
* Print the number of packages.
*****************************************************************/
depth = hwloc_get_type_depth(topology, HWLOC_OBJ_PACKAGE);
if (depth == HWLOC_TYPE_DEPTH_UNKNOWN) {
printf("*** The number of packages is unknown\n");
} else {
printf("*** %u package(s)\n",
hwloc_get_nbobjs_by_depth(topology, depth));
}
/*****************************************************************
* Fourth example:
* Compute the amount of cache that the first logical processor
* has above it.
*****************************************************************/
levels = 0;
size = 0;
for (obj = hwloc_get_obj_by_type(topology, HWLOC_OBJ_PU, 0);
obj;
obj = obj->parent)
if (obj->type == HWLOC_OBJ_CACHE) {
levels++;
size += obj->attr->cache.size;
}
printf("*** Logical processor 0 has %d caches totaling %luKB\n",
levels, size / 1024);
/*****************************************************************
* Fifth example:
* Bind to only one thread of the last core of the machine.
*
* First find out where cores are, or else smaller sets of CPUs if
* the OS doesn't have the notion of a "core".
*****************************************************************/
depth = hwloc_get_type_or_below_depth(topology, HWLOC_OBJ_CORE);
/* Get last core. */
obj = hwloc_get_obj_by_depth(topology, depth,
hwloc_get_nbobjs_by_depth(topology, depth) - 1);
if (obj) {
/* Get a copy of its cpuset that we may modify. */
cpuset = hwloc_bitmap_dup(obj->cpuset);
/* Get only one logical processor (in case the core is
SMT/hyper-threaded). */
hwloc_bitmap_singlify(cpuset);
/* And try to bind ourself there. */
if (hwloc_set_cpubind(topology, cpuset, 0)) {
char *str;
int error = errno;
hwloc_bitmap_asprintf(&str, obj->cpuset);
printf("Couldn't bind to cpuset %s: %s\n", str, strerror(error));
free(str);
}
/* Free our cpuset copy */
hwloc_bitmap_free(cpuset);
}
/*****************************************************************
* Sixth example:
* Allocate some memory on the last NUMA node, bind some existing
* memory to the last NUMA node.
*****************************************************************/
/* Get last node. There's always at least one. */
n = hwloc_get_nbobjs_by_type(topology, HWLOC_OBJ_NUMANODE);
obj = hwloc_get_obj_by_type(topology, HWLOC_OBJ_NUMANODE, n - 1);
size = 1024*1024;
m = hwloc_alloc_membind_nodeset(topology, size, obj->nodeset,
HWLOC_MEMBIND_BIND, 0);
hwloc_free(topology, m, size);
m = malloc(size);
hwloc_set_area_membind_nodeset(topology, m, size, obj->nodeset,
HWLOC_MEMBIND_BIND, 0);
free(m);
/* Destroy topology object. */
hwloc_topology_destroy(topology);
return 0;
}
/* 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
}
//Initializes HWLOC and load the machine architecture
int hw_topology_init (struct arch_topology *topo)
{
hwloc_obj_t obj, core1, core2;
int count, i, j, error;
//Create the machine representation
error = hwloc_topology_init(&topology);
//Go throught the topology only if HWLOC is
//successifully initialized
if(!error)
{
hwloc_topology_load(topology);
local_topo = malloc(sizeof(struct arch_topology));
#if defined (__DBCSR_ACC) || defined (__PW_CUDA)
int nDev;
ma_get_ndevices_cu(&nDev);
#endif
//Extract number of NUMA nodes
if (hwloc_get_type_depth (topology, HWLOC_OBJ_NODE))
topo->nnodes = hwloc_get_nbobjs_by_depth (topology, hwloc_get_type_depth (topology, HWLOC_OBJ_NODE));
else
topo->nnodes = 0;
//Get number of cores, sockets and processing units
topo->ncores = hwloc_get_nbobjs_by_depth (topology, hwloc_get_type_depth (topology, HWLOC_OBJ_CORE));
topo->nsockets = hwloc_get_nbobjs_by_depth (topology, hwloc_get_type_depth (topology, HWLOC_OBJ_SOCKET));
topo->npus = hwloc_get_nbobjs_by_depth (topology, hwloc_get_type_depth (topology, HWLOC_OBJ_PU));
//Compute number of memory controlers per socket
//basically the number of NUMA nodes per socket
if (topo->nnodes > topo->nsockets)
topo->nmemcontroller = topo->nnodes/topo->nsockets;
else
topo->nmemcontroller = 1;
count = 0;
topo->nshared_caches = 0;
//Get derivate information - get number of cache per PU
for(obj = hwloc_get_obj_by_type(topology,HWLOC_OBJ_PU,0);
obj; obj = obj->parent)
{
if (obj->type == HWLOC_OBJ_CACHE)
{
if (obj->arity>1)
topo->nshared_caches++;
else
{ count++; topo->ncaches = count; }
}
}
//Number of direct siblings
//Siblings cores are the ones that share at least one component
//level of the architecture
count = 0;
core1 = hwloc_get_obj_by_type(topology, HWLOC_OBJ_CORE, 0);
core2 = hwloc_get_obj_by_type(topology, HWLOC_OBJ_CORE, 1);
obj = hwloc_get_common_ancestor_obj(topology, core1, core2);
if (obj)
topo->nsiblings = obj->arity;
//Machine node and core representation
machine_nodes = (struct node*) malloc (topo->nnodes*sizeof(struct node));
machine_cores = (struct core*) malloc (topo->ncores*sizeof(struct core));
phys_cpus = malloc (topo->ncores*sizeof(int));
get_phys_id(topology, topo->ncores, 0);
//Get the caches sizes and other information for each core
for (i = 0; i < topo->ncores ; i++)
{
machine_cores[i].caches = malloc (topo->ncaches*sizeof(size_t));
machine_cores[i].shared_caches = malloc (topo->ncaches*sizeof(int));
for (j = 0; j < topo->ncaches; j++)
machine_cores[i].shared_caches[j] = 0;
for (j = topo->ncaches ; j > topo->ncaches - topo->nshared_caches; j--)
machine_cores[i].shared_caches[j-1] = 1;
machine_cores[i].nsiblings = topo->nsiblings;
machine_cores[i].siblings_id = malloc (topo->nsiblings*sizeof(unsigned));
if(topo->ncores == topo->npus){
core1 = hwloc_get_obj_by_type(topology, HWLOC_OBJ_PU, i);
machine_cores[i].id = core1->os_index;
count = 0;
for(obj = hwloc_get_obj_by_type(topology,HWLOC_OBJ_PU,i);
obj; obj = obj->parent) {
if (obj->type == HWLOC_OBJ_CACHE){
machine_cores[i].caches[count] = obj->attr->cache.size / 1024;
count++;
}
if (obj->type == HWLOC_OBJ_NODE)
machine_cores[i].numaNode = obj->logical_index;
}
}
else{
core1 = hwloc_get_obj_by_type(topology, HWLOC_OBJ_CORE, i);
machine_cores[i].id = core1->os_index;
count = 0;
for(obj = hwloc_get_obj_by_type(topology,HWLOC_OBJ_CORE,i);
obj; obj = obj->parent) {
if (obj->type == HWLOC_OBJ_CACHE) {
machine_cores[i].caches[count] = obj->attr->cache.size / 1024;
count++;
}
if (obj->type == HWLOC_OBJ_NODE)
machine_cores[i].numaNode = obj->logical_index;
}
}
}
//Get siblings id - so each core knows its siblings
for (i = 0; i < topo->ncores ; i++)
{
if(topo->ncores == topo->npus){
core1 = hwloc_get_obj_by_type(topology, HWLOC_OBJ_PU, i);
set_phys_siblings(i,machine_cores[i].id,core1,topo->ncores,topo->nsiblings,HWLOC_OBJ_PU);
}
else{
core1 = hwloc_get_obj_by_type(topology, HWLOC_OBJ_CORE, i);
set_phys_siblings(i,machine_cores[i].id,core1,topo->ncores,topo->nsiblings,HWLOC_OBJ_CORE);
}
}
int ncore_node = topo->ncores/topo->nnodes;
int count_cores;
//Get the information for each NUMAnode
for (i = 0; i < topo->nnodes ; i++)
{
obj = hwloc_get_obj_by_type(topology, HWLOC_OBJ_NODE, i);
machine_nodes[i].id = obj->os_index;
machine_nodes[i].memory = obj->memory.total_memory;
machine_nodes[i].ncores = ncore_node;
machine_nodes[i].mycores = malloc (ncore_node*sizeof(unsigned));
//Get the cores id of each NUMAnode
count_cores = 0;
set_node_cores(topology, obj, i, &count_cores);
//GPU support
#if defined (__DBCSR_ACC) || defined (__PW_CUDA)
int *devIds;
devIds = malloc (nDev*sizeof(int));
topo->ngpus = nDev;
ma_get_cu(i,devIds);
machine_nodes[i].mygpus = devIds;
#endif
}
//counting network cards
count = 0;
hwloc_topology_t topo_net;
error = hwloc_topology_init(&topo_net);
hwloc_topology_set_flags(topo_net, HWLOC_TOPOLOGY_FLAG_IO_DEVICES);
if (!error){
hwloc_topology_load(topo_net);
for (obj = hwloc_get_obj_by_type(topo_net, HWLOC_OBJ_OS_DEVICE, 0);
obj;
obj = hwloc_get_next_osdev(topo_net,obj))
if (obj->attr->osdev.type == HWLOC_OBJ_OSDEV_NETWORK ||
obj->attr->osdev.type == HWLOC_OBJ_OSDEV_OPENFABRICS)
count++;
topo->nnetcards = count;
}
else //if can not load I/O devices
topo->nnetcards = 0;
hwloc_topology_destroy(topo_net);
/*Local copy of the machine topology components*/
local_topo->nnodes = topo->nnodes;
local_topo->nsockets = topo->nsockets;
local_topo->ncores = topo->ncores;
local_topo->npus = topo->npus;
local_topo->ngpus = topo->ngpus;
local_topo->ncaches = topo->ncaches;
local_topo->nshared_caches = topo->nshared_caches;
local_topo->nsiblings = topo->nsiblings;
local_topo->nmemcontroller = topo->nmemcontroller;
local_topo->nnetcards = topo->nnetcards;
}
return error;
}
extern int
get_cpuinfo(uint16_t *p_cpus, uint16_t *p_boards,
uint16_t *p_sockets, uint16_t *p_cores, uint16_t *p_threads,
uint16_t *p_block_map_size,
uint16_t **p_block_map, uint16_t **p_block_map_inv)
{
enum { SOCKET=0, CORE=1, PU=2, LAST_OBJ=3 };
hwloc_topology_t topology;
hwloc_obj_t obj;
hwloc_obj_type_t objtype[LAST_OBJ];
unsigned idx[LAST_OBJ];
int nobj[LAST_OBJ];
int actual_cpus;
int macid;
int absid;
int actual_boards = 1, depth;
int i;
debug2("hwloc_topology_init");
if (hwloc_topology_init(&topology)) {
/* error in initialize hwloc library */
debug("hwloc_topology_init() failed.");
return 1;
}
/* parse all system */
hwloc_topology_set_flags(topology, HWLOC_TOPOLOGY_FLAG_WHOLE_SYSTEM);
/* ignores cache, misc */
hwloc_topology_ignore_type (topology, HWLOC_OBJ_CACHE);
hwloc_topology_ignore_type (topology, HWLOC_OBJ_MISC);
/* load topology */
debug2("hwloc_topology_load");
if (hwloc_topology_load(topology)) {
/* error in load hardware topology */
debug("hwloc_topology_load() failed.");
hwloc_topology_destroy(topology);
return 2;
}
/* Some processors (e.g. AMD Opteron 6000 series) contain multiple
* NUMA nodes per socket. This is a configuration which does not map
* into the hardware entities that Slurm optimizes resource allocation
* for (PU/thread, core, socket, baseboard, node and network switch).
* In order to optimize resource allocations on such hardware, Slurm
* will consider each NUMA node within the socket as a separate socket.
* You can disable this configuring "SchedulerParameters=Ignore_NUMA",
* in which case Slurm will report the correct socket count on the node,
* but not be able to optimize resource allocations on the NUMA nodes.
*/
objtype[SOCKET] = HWLOC_OBJ_SOCKET;
objtype[CORE] = HWLOC_OBJ_CORE;
objtype[PU] = HWLOC_OBJ_PU;
if (hwloc_get_type_depth(topology, HWLOC_OBJ_NODE) >
hwloc_get_type_depth(topology, HWLOC_OBJ_SOCKET)) {
char *sched_params = slurm_get_sched_params();
if (sched_params &&
strcasestr(sched_params, "Ignore_NUMA")) {
info("Ignoring NUMA nodes within a socket");
} else {
info("Considering each NUMA node as a socket");
objtype[SOCKET] = HWLOC_OBJ_NODE;
}
xfree(sched_params);
}
/* number of objects */
depth = hwloc_get_type_depth(topology, HWLOC_OBJ_GROUP);
if (depth != HWLOC_TYPE_DEPTH_UNKNOWN) {
actual_boards = MAX(hwloc_get_nbobjs_by_depth(topology, depth),
1);
}
nobj[SOCKET] = hwloc_get_nbobjs_by_type(topology, objtype[SOCKET]);
nobj[CORE] = hwloc_get_nbobjs_by_type(topology, objtype[CORE]);
/*
* Workaround for hwloc
* hwloc_get_nbobjs_by_type() returns 0 on some architectures.
*/
if ( nobj[SOCKET] == 0 ) {
debug("get_cpuinfo() fudging nobj[SOCKET] from 0 to 1");
nobj[SOCKET] = 1;
}
if ( nobj[CORE] == 0 ) {
debug("get_cpuinfo() fudging nobj[CORE] from 0 to 1");
nobj[CORE] = 1;
}
if ( nobj[SOCKET] == -1 )
fatal("get_cpuinfo() can not handle nobj[SOCKET] = -1");
if ( nobj[CORE] == -1 )
fatal("get_cpuinfo() can not handle nobj[CORE] = -1");
actual_cpus = hwloc_get_nbobjs_by_type(topology, objtype[PU]);
#if 0
/* Used to find workaround above */
info("CORE = %d SOCKET = %d actual_cpus = %d nobj[CORE] = %d",
CORE, SOCKET, actual_cpus, nobj[CORE]);
#endif
nobj[PU] = actual_cpus/nobj[CORE]; /* threads per core */
nobj[CORE] /= nobj[SOCKET]; /* cores per socket */
debug("CPUs:%d Boards:%u Sockets:%d CoresPerSocket:%d ThreadsPerCore:%d",
actual_cpus, actual_boards, nobj[SOCKET], nobj[CORE], nobj[PU]);
/* allocate block_map */
*p_block_map_size = (uint16_t)actual_cpus;
if (p_block_map && p_block_map_inv) {
*p_block_map = xmalloc(actual_cpus * sizeof(uint16_t));
*p_block_map_inv = xmalloc(actual_cpus * sizeof(uint16_t));
/* initialize default as linear mapping */
for (i = 0; i < actual_cpus; i++) {
(*p_block_map)[i] = i;
(*p_block_map_inv)[i] = i;
}
/* create map with hwloc */
for (idx[SOCKET]=0; idx[SOCKET]<nobj[SOCKET]; ++idx[SOCKET]) {
for (idx[CORE]=0; idx[CORE]<nobj[CORE]; ++idx[CORE]) {
for (idx[PU]=0; idx[PU]<nobj[PU]; ++idx[PU]) {
/* get hwloc_obj by indexes */
obj=hwloc_get_obj_below_array_by_type(
topology, 3, objtype, idx);
if (!obj)
continue;
macid = obj->os_index;
absid = idx[SOCKET]*nobj[CORE]*nobj[PU]
+ idx[CORE]*nobj[PU]
+ idx[PU];
if ((macid >= actual_cpus) ||
(absid >= actual_cpus)) {
/* physical or logical ID are
* out of range */
continue;
}
debug4("CPU map[%d]=>%d", absid, macid);
(*p_block_map)[absid] = macid;
(*p_block_map_inv)[macid] = absid;
}
}
}
}
hwloc_topology_destroy(topology);
/* update output parameters */
*p_cpus = actual_cpus;
*p_boards = actual_boards;
*p_sockets = nobj[SOCKET];
*p_cores = nobj[CORE];
*p_threads = nobj[PU];
#if DEBUG_DETAIL
/*** Display raw data ***/
debug("CPUs:%u Boards:%u Sockets:%u CoresPerSocket:%u ThreadsPerCore:%u",
*p_cpus, *p_boards, *p_sockets, *p_cores, *p_threads);
/* Display the mapping tables */
if (p_block_map && p_block_map_inv) {
debug("------");
debug("Abstract -> Machine logical CPU ID block mapping:");
debug("AbstractId PhysicalId Inverse");
for (i = 0; i < *p_cpus; i++) {
debug3(" %4d %4u %4u",
i, (*p_block_map)[i], (*p_block_map_inv)[i]);
}
debug("------");
}
#endif
return 0;
}
