[Bioclusters] Direct connect infiniband/quadrics?

Guy Coates gmpc at sanger.ac.uk
Thu Mar 17 12:44:16 EST 2005


>existing system with a proposed new system. I'm going to try IB/Quadrics
>because from what I can tell, the molecular dynamics applications that run
>on the system will definitely take advantage of the faster interconnect.

The devil is in the details when you say the code "can take advantage of"
a faster interconnect. What the interconnect give you is an increase
scalability (ie better parallel efficiency) as you increase the node
count, rather than making it "go faster".

When I last used gromacs in anger, (a few years ago)  I was seeing pretty
good parallel efficiency out to 4 machine (8 CPUs) on 100Mbit ethernet.

There is no point in putting a fast interconnect between two machines; you
won't see any better performance than if you'd just stuck an ethernet
cross-over cable between them.

If you are going to be running large simulations over more than 10
machines at a time, then yes, think about the interconnect, but benchmark
first.

Guy

-- 
Dr. Guy Coates,  Informatics System Group
The Wellcome Trust Sanger Institute, Hinxton, Cambridge, CB10 1SA, UK
Tel: +44 (0)1223 834244 ex 7199







More information about the Bioclusters mailing list