Re: [PATCH bpf-next v5] selftests/bpf: Add benchmark for bpf memory allocator

[Andrii Nakryiko <andrii.nakryiko@gmail.com>]

On Mon, Jun 12, 2023 at 4:38 PM Alexei Starovoitov
<alexei.starovoitov@gmail.com> wrote:
>
> On Mon, Jun 12, 2023 at 4:17 PM Andrii Nakryiko
> <andrii.nakryiko@gmail.com> wrote:
> >
> > On Fri, Jun 9, 2023 at 9:12 AM Alexei Starovoitov
> > <alexei.starovoitov@gmail.com> wrote:
> > >
> > > On Thu, Jun 8, 2023 at 11:32 PM Hou Tao <houtao@huaweicloud.com> wrote:
> > > >
> > > > >
> > > > >> +                    --producers=8 --prod-affinity=0-7 "$@")"
> > > > > -a -p 8 should just work.
> > > > > No need to pick specific cpus.
> > > > No. For VM with only 8 CPUs, the affinity of the first producer will be
> > > > CPU 1 and the affinity of the last producer will be CPU 8, so the
> > > > benchmark will fail to run. But I think I can fix it, so the affinity of
> > > > the last producer will be 0 instead.
> > >
> > > Right. Noticed that too.
> > > That should probably be a separate patch to fix this cpu assignment
> > > issue in bench for all benchs.
> > >
> > > Andrii,
> > > when you wrote it did you really mean to start assigning cpus from 1
> > > or that was just an oversight?
> >
> >   616 ▸       /* unless explicit producer CPU list is specified, continue after¬
> >   617 ▸        * last consumer CPU¬
> >   618 ▸        */¬
> >
> >
> > It's been a while, but it seems like each consumer gets its CPU first,
> > then each producer. So yeah, seems intentional.
> >
> > For context, this was done for BPF ringbuf benchmarking, so by default
> > I wanted to separate a single consumer from multiple producers.
>
> I see. In this cas Hou's bench has empty consumer:
> +static void *htab_mem_consumer(void *arg)
> +{
> +       return NULL;
> +}
>
> but setup_benchmark() still creates an instant 'return NULL' thread
> and pins it to cpu 0.
>
> I guess the fix is for htab-mem bench to set env.consumer_cnt = 0
> at init time and don't supply empty consumer at all.
>
> May be consumer_cnt=0 can a default and ringbuf bench will just the
> default to 1?

makes sense