Add numa-balanced mode to hpx::bind, spread cores over numa domains #2900
Conversation
| std::size_t num_threads = affinities.size(); | ||
| num_pus.resize(num_threads); | ||
| // numa nodes | ||
| std::size_t num_numas = (std::max)(std::size_t(1),t.get_number_of_numa_nodes()); |
There was a problem hiding this comment.
Should we add a fallback to the number of sockets here if the kernel misses NUMA support? We do that already in a couple of places.
There was a problem hiding this comment.
Not really any point. if the kernel lacks numa support then asking for numa-balanced is meaningless anyway. Let it default to 1 domain.
There was a problem hiding this comment.
What about reporting an error at startup if the user asks for NUMA-domain awareness on a machine which does not have NUMA domains.
| } | ||
| // how many threads should go on each domain | ||
| for (std::size_t n=0; n<num_numas; ++n) { | ||
| num_threads_numa[n] = num_threads*num_cores_numa[n]/cores_t; |
There was a problem hiding this comment.
This needs to be rounded up, in the case where, for example, 1 thread is requested, we end up with 0 threads per numa domain.
|
That fixes the 1 thread and any odd number requested - and also ensures that the rounding up does not oversubscribe the node |
biddisco
moved this from Work in Progress
to completed
in resource partitioner/manager work
This fixes #2899