Le 29/09/2014 19:01, Aulwes, Rob a écrit :
> Hi,
>
> I'm trying to allocate and bind memory on the same NUMA domain as the
> calling thread. The code I use is as follows.
>
> /* retrieve the single PU where the current thread actually
> runs within this process binding */
>
>
> int err = hwloc_get_last_cpu_location(topo, set,
> HWLOC_CPUBIND_THREAD);
>
> if (err < 0) {
>
> fprintf(stderr, "failed to get last cpu location\n");
>
> return NULL;
>
> }
>
>
> int idx = hwloc_bitmap_first(set);
>
> auto obj = hwloc_get_pu_obj_by_os_index(topo, idx);
>
>
> // get NUMA node containing PU
>
> obj = hwloc_get_ancestor_obj_by_type(topo, HWLOC_OBJ_NODE, obj);
>
> void * buffer = hwloc_alloc_membind_nodeset(topo, len,
> obj->nodeset,
>
> HWLOC_MEMBIND_BIND,
>
> HWLOC_MEMBIND_THREAD |
> HWLOC_MEMBIND_STRICT);
>
> if (!buffer) {
>
> fprintf(stderr, "failed to manually allocate memory on
> node %u\n", obj->os_index);
>
> return NULL;
>
> }
>
>
> However, for some allocations, if I print out where the allocation was
> bound, it doesn't match the node of the thread. Any ideas about what
> I'm doing wrong?
>
Hello
I don't see any problem with your code.
Are you sure the threads are bound within a single node?
Random optimization idea: instead of
bitmap_first()+get_pu_obj_by_os_index()+get_ancestor_by_type(), you can
directly get the first NUMA node involved in cpuset "set" by calling
hwloc_get_next_obj_covering_cpuset_by_type(topo, set, HWLOC_OBJ_NODE, NULL)
Brice
