[4062] in linux-net channel archive
Re: Question on performance
daemon@ATHENA.MIT.EDU (Chris Albertson)
Fri Aug 16 20:11:08 1996
From: chris@topdog.pas1.logicon.com (Chris Albertson)
Date: Fri, 16 Aug 96 10:24:32 PDT
To: alan@cymru.net
Cc: linux-net@vger.rutgers.edu
>
> > > Our Sun technical rep said that Suns can push data close to the ethernet
> > > "wire speed" but PCs only get about 1/2 this rate. This advice was in the
>
> The Sun rep is wrong. This is not unusual with Sun reps.
>
You cut my quote off to soon. "the advice was in the -- the context of setting
up a Sun server. The _typical_ DOS/Windows 486 PC with a low end ISA ne2000
card running TCP as a stack of TSRs is only half as fast as a Sun SPARC running
SunOS. The numbers I see are closer to 1/3 as fast. This is for a _typical_
office PC. Not a PCI based state of the art box running Linux. The paper
went on to say that _typical_ PC users tend not to hit the network as hard as
_typical_ workstation users. For this reason it was argued that one SPARC server
could serve 100 typical PCs but not nearly so many UNIX workstations. It then
want on to talk about CPU and RAM requirements then how many SCSI busses and the
number of drives per SCSI bus. All in all I think advice worked out well.
> > (e.g. NCR53c810) and a busmaster Ethernet controller (e.g. PCNet-PCI
> > or Tulip) will probably _not_ find its network or SCSI interfaces to
> > be the bottleneck - the bottlenecks will be in the filesystem / block
> > cache and in the network stack, I think
>
> The bottleneck becomes memory speed and disk speed. On a 100baseT ethernet
> you need to sustain about 10Mbytes/second from disk to memory and
> 10Mbytes/second from memory to network card, and do a checksum (in our
> case a copy/checksum) between. Thats already 30Mbytes/second memory
> bandwidth just to send the data out.
>
> Alan
>
