numa_preferred returns unexpected result

[angainor]

On two systems I tested, numa_preferred always returns 0 if not explicitly initialized. That means that even when I start a process explicitly on some processor on NUMA node ~= 0, using the default memory allocation policy (local alloc), numa_preferred will return 0. The memory will be allocated correctly on the local numa node.

https://gist.github.com/angainor/d09c367631880ae99e1acf7f6653128b

I suggest we add the following function to numa_tools:

numa_tools::index_type
numa_tools::local_node() const noexcept
{
    return numa_node_of_cpu(sched_getcpu());
}

While this will not return the preferred node, it will return the thread-local node id, which is often what we want.

[boeschf]

Good catch. I see in the gist that you use a multi-threaded test. does it work for single-threaded process? I.e. does it has something to do with the way we implement the initialization of libnuma?

[angainor]

@boeschf I think it's the same - in the gist I print

  printf("MAIN preferred %d\n", numa_preferred());

before the threaded section also.

[boeschf]

oh yes, my bad