Switch to avoid (was Re: NIC to avoid...?)
rauch at inf.ethz.ch
Tue Aug 28 09:47:44 PDT 2001
On Tue, 28 Aug 2001, Michael Sowka wrote:
> Performance is the key as we're blowing a fair amount of money and a
> Cisco Catalist 3500 Switch....
I certainly don't want to stop you blowing money and I don't know
anything about Catalist 3500 switches, but since you said performance
is the key I thought I'd share some experiences we had with a Catalyst
We had to compare the performance of two switches, one of them beeing
the Catalyst 2900XL. We basically beat that thing into pieces. The
experiment was performed with an extended version of our cloning tool
Dolly . This version just sends data as fast as it can from one
node to the next and from there at the same time to the next and so
on. It sends data around in a virtual TCP ring, so every used port of
the switch has to send and receive data with full speed. Pretty tough
for some switches, it seems, but a realistic scenario in our case.
The result was that the Catalyst 2900XL (which was a 24 port switch)
could sustain full performance at about 11 MB/s for the data stream up
to 12 nodes [2,3]. For more than 12 nodes it went to its knees, beging
for mercy. For 15 nodes it sustained only 2.5 MB/s!
An Allied Telesyn Centrecom 742i 24 port switch had no problems handling the
load [4,5] up to 15 nodes (we don't have more).
As I said, I don't know the differences between the Catalists 3500 and
2900XL, but if you need full bandwith on many ports, then you'd
probably better benchmark that switch (reading the specifications
doesn't necessarily help. We recently benchmarked an Enterasys Matrix E7
with 128 nodes and this thing also didn't fulfill the specs, although
it didn't fail as bad as the 2900XL).
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
More information about the Beowulf