Re: [PATCH 2/2 v2] perf bench futex: add NUMA support

From: Davidlohr Bueso
Date: Thu Oct 20 2016 - 22:34:28 EST

On Wed, 19 Oct 2016, Sebastian Andrzej Siewior wrote:

On 2016-10-19 11:16:16 [-0700], Davidlohr Bueso wrote:
On Mon, 17 Oct 2016, Sebastian Andrzej Siewior wrote:

> By default the application uses malloc() and all available CPUs. This
> patch introduces NUMA support which means:
> - memory is allocated node local via numa_alloc_local()
> - all CPUs of the specified NUMA node are used. This is also true if the
> number of threads set is greater than the number of CPUs available on
> this node.

Can't we just use numactl to bind cpus and memory to be node-local?

something like
numactl --cpunodebind=$NODE --membind=$NODE perf ???


This should work for memory however since we use
pthread_attr_setaffinity_np(&thread_attr, sizeof(cpu_set_t), &cpu);
we would need to query the affinity mask, and deploy threads based on
that mask.

Ah right. I also considered getting rid of the affinity, but that would
probably hurt more than help (or at least alter) for non-numa options.

Using NUMA support within this bench-tool has also the side effect that
the output gives all the option used.

So if we are going to support the numa option for the benchmark, could you please
move the new code into futex.h instead of futex-hash.c? That way we can integrate
the support for the other futex programs as well.