[Bioclusters] Direct connect infiniband/quadrics?

[Farul M. Ghazali]

On Tue, 15 Mar 2005, John Van Workum wrote:

> You also should ask yourself what are you going to determine using point to
> point IB/Quadrics? Assuming you have some benchmark code that is well written
> and works on both IB/Quadrics and Ethernet, your results will show Ethernet is
> slower. But you might not be able to determine if using IB/Quadrics justifies
> the cost. Perhaps a better test case is needed. How well will your application
> scale? You might want to spend the money on a ~16 node cluster with GB
> Ethernet for the test bed. If your tests run well on GB Ethernet you could
> then extrapolate the benefits of upgrading to IB/Quadrics.

John,

The lab already does run amber and gromacs on a 32-CPU Xeon cluster so
there is a baseline for benchmarking. Unfortunately the cluster belongs to
the physics department so we can't just upgrade that cluster.

The question is which one provides the best performance for the money,
either a dual-cpu with gigabit or quad-cpu with IB/Quadrics or something
else altogether. The budget and the actual number of nodes we can afford
with that budget will come into play when talking about expensive
interconnects as well. From my point of view, a two-node quad-Opteron
system with IB/Quadrics is a good starting point for comparing the
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.