/******************* FUNCTION *********************/
int TopoHwloc::getCurrentId ( int level, int depth ) const
{
//vars
int res = -1;
//get absolute depth
int absDepth = getAbsDepth(level,depth);
//search current thread binding
hwloc_cpuset_t cpuset = hwloc_bitmap_alloc();
int status = hwloc_get_cpubind (topology, cpuset, 0);
allocAssume(status >= 0,"Failed to get the current thread CPU binding with hwloc.");
//count overlap over current level
int cnt = hwloc_get_nbobjs_inside_cpuset_by_depth(topology,cpuset,absDepth);
//if get only one return its ID
if (cnt == 1)
{
hwloc_obj_t obj = hwloc_get_obj_inside_cpuset_by_depth(topology,cpuset,absDepth,0);
//TODO maybe their is an accessor for this
res = obj->logical_index;
}
//clean memory
hwloc_bitmap_free(cpuset);
return res;
}
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);
}
int hmonitor_eventset_add_named_event(void * monitor_eventset, const char * event)
{
struct accumulate_eventset * set = (struct accumulate_eventset *) monitor_eventset;
harray child_events = NULL;
hwloc_obj_t obj;
unsigned depth,index,i, n_events = 0;
hmon m;
if(event==NULL || monitor_eventset == NULL)
return -1;
/* Decend depth by depth to look if event exists */
int topo_depth = hwloc_topology_get_depth(hmon_topology);
for(depth=set->location->depth; depth<topo_depth; depth++){
/* Traverse depth to find matching event */
for(index=0; index<hwloc_get_nbobjs_inside_cpuset_by_depth(hmon_topology, set->location->cpuset, depth); index++){
obj = hwloc_get_obj_inside_cpuset_by_depth(hmon_topology, set->location->cpuset, depth, index);
child_events = obj->userdata;
if(child_events != NULL){
/* Walk monitor list at this location to find matching event */
for(i=0; i<harray_length(child_events); i++){
m = harray_get(child_events,i);
if(!strcmp(m->id, event)){
harray_push(set->child_events, m);
n_events=m->n_samples;
}
}
}
}
}
if(n_events == 0){
/* Exit failure */
fprintf(stderr, "Unrecognized event name %s, expected defined monitor name, deeper than %s.\n", event, hwloc_type_name(set->location->type));
return -1;
}
return n_events;
}
void chpl_topo_init(void) {
//
// We only load hwloc topology information in configurations where
// the locale model is other than "flat" or the tasking is based on
// Qthreads (which will use the topology we load). We don't use
// it otherwise (so far) because loading it is somewhat expensive.
//
if (strcmp(CHPL_LOCALE_MODEL, "flat") != 0
|| strcmp(CHPL_TASKS, "qthreads") == 0) {
haveTopology = true;
} else {
haveTopology = false;
return;
}
// Check hwloc API version.
// Require at least hwloc version 1.11 (we need 1.11.5 to not crash
// in some NUMA configurations).
// Check both at build time and run time.
#define REQUIRE_HWLOC_VERSION 0x00010b00
#if HWLOC_API_VERSION < REQUIRE_HWLOC_VERSION
#error hwloc version 1.11.5 or newer is required
#endif
CHK_ERR(hwloc_get_api_version() >= REQUIRE_HWLOC_VERSION);
//
// Allocate and initialize topology object.
//
CHK_ERR_ERRNO(hwloc_topology_init(&topology) == 0);
//
// Perform the topology detection.
//
CHK_ERR_ERRNO(hwloc_topology_load(topology) == 0);
//
// What is supported?
//
topoSupport = hwloc_topology_get_support(topology);
//
// TODO: update comment
// For now, don't support setting memory locality when comm=ugni or
// comm=gasnet, seg!=everything. Those are the two configurations in
// which we use hugepages and/or memory registered with the comm
// interface, both of which may be a problem for the set-membind call.
// We will have other ways to achieve locality for these configs in
// the future.
//
do_set_area_membind = true;
if ((strcmp(CHPL_COMM, "gasnet") == 0
&& strcmp(CHPL_GASNET_SEGMENT, "everything") != 0)) {
do_set_area_membind = false;
}
//
// We need depth information.
//
topoDepth = hwloc_topology_get_depth(topology);
//
// How many NUMA domains do we have?
//
{
int level;
//
// Note: If there are multiple levels with NUMA nodes, this finds
// only the uppermost.
//
for (level = 0, numaLevel = -1;
level < topoDepth && numaLevel == -1;
level++) {
if (hwloc_get_depth_type(topology, level) == HWLOC_OBJ_NUMANODE) {
numaLevel = level;
}
}
}
//
// Find the NUMA nodes, that is, the objects at numaLevel that also
// have CPUs. This is as opposed to things like Xeon Phi HBM, which
// is memory-only, no CPUs.
//
{
const hwloc_cpuset_t cpusetAll = hwloc_get_root_obj(topology)->cpuset;
numNumaDomains =
hwloc_get_nbobjs_inside_cpuset_by_depth(topology, cpusetAll, numaLevel);
}
}
void chpl_topo_init(void) {
//
// For now we don't load topology information for locModel=flat, since
// we won't use it in that case and loading it is somewhat expensive.
// Eventually we will probably load it even for locModel=flat and use
// it as the information source for what's currently in chplsys, and
// also pass it to Qthreads when we use that (so it doesn't load it
// again), but that's work for the future.
//
haveTopology = (strcmp(CHPL_LOCALE_MODEL, "flat") != 0) ? true : false;
if (!haveTopology) {
return;
}
// Check hwloc API version.
// Require at least hwloc version 1.11 (we need 1.11.5 to not crash
// in some NUMA configurations).
// Check both at build time and run time.
#define REQUIRE_HWLOC_VERSION 0x00010b00
#if HWLOC_API_VERSION < REQUIRE_HWLOC_VERSION
#error hwloc version 1.11.5 or newer is required
#else
{
unsigned version = hwloc_get_api_version();
// check that the version is at least REQUIRE_HWLOC_VERSION
if (version < REQUIRE_HWLOC_VERSION)
chpl_internal_error("hwloc version 1.11.5 or newer is required");
}
#endif
//
// Allocate and initialize topology object.
//
if (hwloc_topology_init(&topology)) {
report_error("hwloc_topology_init()", errno);
}
//
// Perform the topology detection.
//
if (hwloc_topology_load(topology)) {
report_error("hwloc_topology_load()", errno);
}
//
// What is supported?
//
topoSupport = hwloc_topology_get_support(topology);
//
// TODO: update comment
// For now, don't support setting memory locality when comm=ugni or
// comm=gasnet, seg!=everything. Those are the two configurations in
// which we use hugepages and/or memory registered with the comm
// interface, both of which may be a problem for the set-membind call.
// We will have other ways to achieve locality for these configs in
// the future.
//
do_set_area_membind = true;
if ((strcmp(CHPL_COMM, "gasnet") == 0
&& strcmp(CHPL_GASNET_SEGMENT, "everything") != 0)) {
do_set_area_membind = false;
}
//
// We need depth information.
//
topoDepth = hwloc_topology_get_depth(topology);
//
// How many NUMA domains do we have?
//
{
int level;
//
// Note: If there are multiple levels with NUMA nodes, this finds
// only the uppermost.
//
for (level = 0, numaLevel = -1;
level < topoDepth && numaLevel == -1;
level++) {
if (hwloc_get_depth_type(topology, level) == HWLOC_OBJ_NUMANODE) {
numaLevel = level;
}
}
}
//
// Find the NUMA nodes, that is, the objects at numaLevel that also
// have CPUs. This is as opposed to things like Xeon Phi HBM, which
// is memory-only, no CPUs.
//
{
const hwloc_cpuset_t cpusetAll = hwloc_get_root_obj(topology)->cpuset;
numNumaDomains =
hwloc_get_nbobjs_inside_cpuset_by_depth(topology, cpusetAll, numaLevel);
}
}
