[Bioclusters] bio/life science cluster switch

[LAI Loong Fong]

Hi Joe,
May I know what is the reason for avoiding Cisco (beside being expensive)?
We are using the Cisco 3750, 4503 and 4506 and do not have much problems
with them. Of cause the HP Procurve are nice switches too, we are using the
2848 and it has a 96Gbps backplace. We also use the Foundry Big Iron, but
that will be an over kill.

Regards
LAI Loong-Fong

On 6/8/04 7:58 PM, "Joe Landman" <landman@scalableinformatics.com> wrote:

> Hi Steve:
> 
>  The choice of switch fabric is largely determined by the application
> and use case mix.  For codes requiring low latency network fabric such
> as electronic structure codes, various molecular dynamics, and whatnot
> else, you would be looking at one of the low latency high bandwidth
> systems.  For more general applications, and trivial/embarrassingly
> parallel work, you could go with a higher latency fabric.
> 
> Fabric topology is also important, though for this size, most folks
> aim towards fat trees or simple switch connected with no hierarchies.
> 
>  Since you will be going pure gigabit, I might recommend you look at
> some of the switches with better backplane bandwidth and latency
> characteristics.  The HP Procurve 4108gl are modular, and you can add
> the J4908A 20 port gigabit module (6 to cover your needs).  It wont be
> cheap, but it will work nicely.  The Foundry's are good choices as
> well.  I would avoid putting a Cisco in a cluster.
> 
> One of the nicer aspects of the low latency gigabit connections is the
> ability to use TCP offload cards.  If you really need low latency
> interconnects over all or a portion of your system, the offload adapters
> may benefit some workloads.  As always, it is a cost-benefit analysis.
> 
> (disclosure:  my company designs/integrates clusters like this, so my
> comments reflect our biases)
> 
> Joe
> 
> Steve wrote:
> 
>> I'm putting together a general lifescience based cluster (software
>> ranging from docking apps. to blast). Its going to consist of
>> approximately 100 nodes, including 2 head nodes. All nodes will be
>> connected through GIGE, and we will not be using high speed
>> interconnects.
>> 
>> In the past I've used Cisco switches (4000 and 6000 series), stable,
>> reliable and costly.
>> Does anyone have any recommendations? I've had mixed reviews about
>> Extreme, so I'd be interested in Foundy, Procurve or any systems
>> people may have had success with.
>> 
>> Cheers,
>> 
>> Steve