[prev in list] [next in list] [prev in thread] [next in thread] 

List:       linux-smp
Subject:    Re: recap of Linux-SMP CPU and memory bandwidth bench
From:       alan () lxorguk ! ukuu ! org ! uk (Alan Cox)
Date:       1997-06-27 23:51:07
[Download RAW message or body]

> > You should be seeing the full 100baseT jammed. You may be seeing capture
> > effect. On SMP our networking performance in 2.0 is zapped by the SMP
> > locks measurably. You may want to see if adding "window 16384" or
> > "window 14000" improves your 100baseT performance by trying to avoid
> > capture.
> Could you elaborate on this a bit - I am very interested in using
> 100baseT to connect SMP boxes - as a parallel computer.

Ok two things. The SMP locks in 2.0 can cause interrupts to be processed
much more slowly than uniprocessor 2.0. That can have a real performance hit
2.1 already cures this. It will never be cured in 2.0.x (its too big a 
change)

Second item "capture effect" - this only applies in half duplex mode and
generally mostly on 10baseT networks. What happens is that someone sends
continual data. The other end tries to transmit acks and fails, and fails
and maybe collides. But even if it collides because it has backed off
a lot already the sender of the large amounts of data will normally be
the person who sends data next. Thus a sender of continual data "captures"
the ethernet [remember ethernet is not and isnt meant to be fair...]

On a full duplex 100baseT this just doesnt occur. 

> >Its also heavily card dependant. The tulips have bus mastering DMA
> >the 3c90x has sort of working DMA but not for a lot of cards (if you see
> >"rx pacing bug exists" then your 3c90x is being run in a slower mode)
> So the tulips are better?

Generally yes

[prev in list] [next in list] [prev in thread] [next in thread]