good, cheap switches?

[Felix Rauch]

On Sat, 11 May 2002, Eray Ozkural wrote:
> On Saturday 11 May 2002 15:24, Andrew Leahy wrote:
> > Recently, I've locked horns with the computer support people about what
> > sort of a switch I should use in my 20-node beowulf cluster.  I'm looking
> > for a 24-port 100Mbit switch, and they're pushing a Cisco 3350-24 switch
> > (around $1900) apparently for the simple reason that that's the switch they
> > use on the rest of the campus.  As I look at the product specification
> > sheet, their switch is loaded with features a campus-wide network manager
> > is bound to love, but would be useless on an isolated beowulf cluster. 
> > Since cost is an issue here, I'm curious if anybody has any recommendations
> > for a cheaper (unmanaged?) switch with a switching fabric of comparable
> > quality.
> 
> I think you've already solved your problem. You don't need Cisco. You'd be 
> paying your dollars to their marketing and advertising people instead of the 
> engineering.

I'd agree with this. I have not much experience with Cisco switches,
but we once tested one of the Cisco switches (I think 2900XL or
something similar with 24 ports) and it failed miserably when using
more than 12 ports at full bidirectional speed [1]. Other Cisco models
might be better, but most likely also more expensive.

> I had looked at performance comparisons of switches when we were to buy one 
> and I ended up with a 3com switch (which wasn't the cheapest). You'd need 
> something like that, don't trust what the companies themselves claim about 
> the performance of the switches.

Again, I can second the part aboutw what companies claim. When we
specified the requirements for our 128 node cluster [2], we put in
"full bisection bandwidth". So our assembler bought us an Enterasys
Matrix E7 switch, for which the data sheet specified enough
performance for full bisection bandwidth. However, our tests proofed
the opposite.

For small unmanaged switches we made good experiences with small ATI
(Allied Telesyn) switches. As for our (bandwidth oriented) tests, we
didn't find any problems [1].

- Felix

[1] Some switch measurements results:
    http://www.cs.inf.ethz.ch/~rauch/tmp/FE.Catalyst2900XL.Stream.pdf
    http://www.cs.inf.ethz.ch/~rauch/tmp/FE.Catalyst2900XL.Agg.pdf
    http://www.cs.inf.ethz.ch/~rauch/tmp/FE.CentreCom742i.Stream.pdf
    http://www.cs.inf.ethz.ch/~rauch/tmp/FE.CentreCom742i.Agg.pdf
[2] http://www.xibalba.inf.ethz.ch/
---
Felix Rauch                      | Email: rauch at inf.ethz.ch
Institute for Computer Systems   | Homepage: http://www.cs.inf.ethz.ch/~rauch/
ETH Zentrum / RZ H18             | Phone: ++41 1 632 7489
CH - 8092 Zuerich / Switzerland  | Fax:   ++41 1 632 1307