Re: [PATCH 1/5] soreuseport: infrastructure

[Eric Dumazet <eric.dumazet@gmail.com>]

On Wed, 2013-01-16 at 10:37 -0800, Tom Herbert wrote:
> > Ideally, we should chose a target given by the current CPU number, in
> > case SYN messages are spread on all cpus or a set of cpus.
> >
> It is an ideal, but I don't readily see a practical way to do this
> given the available information, the fact that number of sockets
> created is up to the application, and the fact that there is no fixed
> binding of a socket to CPU.
> 
> Consider the simple "cpu % num" algorithm in packet fanout.  Suppose
> is a 16 CPU system, and RX queues for the NIC are processed on CPUS
> 0,3,7,11 and user creates 4 sockets.  In this configuration, on the
> first socket would ever be selected!

Sure, any hand coded 'optimization' should be correctly done.

On a 16 cpus system, I would create 16 queues, if we stick to the "cpu %
nr_queues" simple selection.

If some queues are never used, thats not a big deal, unless you have a
crazy spin polling of the queues. A blocked thread consumes almost
nothing.

It would be a rather straightforward patch to add mask capability to
af_packet (a la rps_cpus ), if we really wanted 4 queues, served by cpus
0,3,7,11 but I don't think there is an urgent need.

Another way would be to let user land to declare a preferred cpu for the
queue, even if not related to process scheduler affinity.

Anyway, this can be addressed later, in followup patches.