Re: [PATCH 1/2] rps: core implementation

[Jarek Poplawski <jarkao2@gmail.com>]

On Fri, Nov 20, 2009 at 09:08:10AM -0800, Tom Herbert wrote:
> > The description reads: "This solution queues packets early on in the
> > receive path on the backlog queues of other CPUs.", so I'm not sure
> > it's intended.
> 
> That is precisely the intent.  Getting packets quickly distributed to
> the target cpus maximizes parallelism and reduces latency.

Then precisely the intent is "the backlog queues of target CPUs".
"The backlog queues of other CPUs" may suggest that e.g. one cpu is
only doing distribution etc.

> 
> Did you test it like this (and it was visibly worse)?:
> >
> >        if (cpu < 0 || cpu == smp_processor_id())
> >
> >> +             return __netif_receive_skb(skb);
> >> +     else
> >> +             return enqueue_to_backlog(skb, cpu);
> >> +}
> >> +
> This increases overall latency due to head of line blocking which will
> outweigh the benefits of optimizing for this one case.

The way I asked should suggest I "suspected" it's on purpose, and was
curious about "digits", but thanks for confirming this.

Jarek P.