[Beowulf] Clos networking and myrinet

[Mikhail Kuzminsky]

In message from Greg Lindahl <lindahl at pathscale.com> (Wed, 8 Jun 2005 
11:57:17 -0700):
>> >In a Linux cluster, each node talks to multiple other nodes. So 
>>they
>> >always interfere with each other.
>
>> Sorry, why is it right ?
>> - for typical example, if I have a pair of nodes w/MPI_Send one 1st
>> and MPI_Recv on second ?
>
>For only a pair of nodes, a switch is always nonblocking. Do you 
>build
>a lot of 2 node clusters? ;-) You could save some money by getting 
>rid
>of the switch...
>
>Seriously, as I said, if you only have each node talking to one other
>node, then the statement about Clos networks being non-blocking is
>true.  But this is not typical for Linux clusters.

I'm developing my MPI application for quantum chemistry, and I'm 
taking
into account the possibility to perform a set of simultaneous 
MPI_Send/MPI_Recv "data exchanges". It's good for wide spectrum
of clusters w/variable interconnetcs, beginning from smallest clusters 
having only 1 switch.

Our application is bandwidth-limited, so it's extremally important for 
us - to optimize "MPI- exchanges" of messages.

Yours
Mikhail

>
>> It looks for me that both Myrinet and Infiniband topologies
>> specially support "Clos-network configurations" ...
>
>Yes. The InfiniBand people use different jargon (CBB: constant
>bisection bandwidth) but it boils down to the same thing.
>
>-- greg
>
>_______________________________________________
>Beowulf mailing list, Beowulf at beowulf.org
>To change your subscription (digest mode or unsubscribe) visit 
>http://www.beowulf.org/mailman/listinfo/beowulf