Re: Testing Bandwidth performance

["todd glassey" <todd.glassey () worldnet att net>]

----- Original Message -----
From: "David G. Andersen" <dga () lcs mit edu>
To: "todd glassey" <todd.glassey () worldnet att net>
Cc: "Jared Mauch" <jared () puck Nether net>; "Martin Hannigan"
<hannigan () fugawi net>; "Wojtek Zlobicki" <wojtekz () idirect com>; "Alan Sato"
<asato () altrio net>; <nanog () trapdoor merit edu>
Sent: Wednesday, June 26, 2002 6:30 AM
Subject: Re: Testing Bandwidth performance


> On Wed, Jun 26, 2002 at 06:18:00AM -0700, todd glassey mooed:

I have never been referred to as bovine before. I usually describe myself as
a small polar bear... Hmmm.

> > Oh and use something like a SNIFFER to generate the traffic. Most of
what we
> > know of as commercial computer's cannot generate more than 70% to 80%
> > capacity on whatever network they are on because of driver overhead and
OS
> > latency etc etc etc. It was funny, but I remember testing FDDI on a
UnixWARE
> > based platform and watching the driver suck 79% of the system into the
> > floor.
>
>   Btw, if you've got a bit of time on your hands, the Click router
> components have some extremely-low-overhead drivers (for specific
> ethernet cards under Linux).

Good Point.

> They can generate traffic at pretty
> impressive rates.  They used them for testing DOS traffic for a while.
>
>   http://pdos.lcs.mit.edu/click/

Still there are very few parametric engines that will generate more than
100Mb/S traffic - continuously.

>   (Most of the driver overhead you see is interrupt latency;

Depends on which OS you are running, and what encapsulation or other
packaging/unpackaging is done in the Driver also accounts for a substantial
amount of the compute model. If those services are done mostly in HW then
the systems I have played with will give you up to about 80% capacity. And
on ethernet that is not Collision-Free (i.e.run as full duplex), then you
have to deal with the line characteristics so with both engines competing to
flood the net you may actually get less than 80%  total performance...


> click
> uses an optimized polling style to really cram things through).  Also,
> the new FreeBSD polling patches should make it so you can get more
> throughput from your drivers when doing tests.  I understand there are
> similar things for Linux.

The Linux Router Project has similar features.

>   -Dave
>
> --
> work: dga () lcs mit edu                          me:  dga () pobox com
>       MIT Laboratory for Computer Science           http://www.angio.net/