On 10/8/07, erik quanstrom <[EMAIL PROTECTED]> wrote:
> > I've been thinking a lot about this, particularly as we've been diving
> > deep into tracing performance of our network paths as part of the Blue
> > Gene work. As of our preliminary results, it would seem that Plan 9
> > attempts to take the most general approach to things with an emphasis
> > on keeping everything simple. Unfortunately we end up paying a heavy
> > price in raw performance (at least in the networking case).
>
> i have found plan 9 ip networking does not fair well pushing ip networking
> at 10gbps, especially tcp, and especially with standard ethernet frames.
>
> i didn't spend enough time this spring with the tcp stack to understand what
> it's doing, but i got the impression that it could make the simple case
> of receiving packets in order simplier at the expence of misorderd packets.
>
> fwiw, we get much greater performance pushing AoE. so i don't think
> that network queues themselves are the problem.
>
I think most of the issues we are facing are due to somewhat difficult
esoteric network interfaces combined with a single 700 MHz simple
embedded in-order processor. That being said however, any advantages
we can get on the relatively slow hardware should benefit the
efficiency (if not the latency and performance) of the faster
platforms. Even if we can push peak on our platform it won't help us
if we are chewing up cpu cycles that we need for computation.
-eric
