[Coin-discuss] Clustering COIN?

[Ted Ralphs]

anthony at resolution.com wrote:

> I hope this finds you well.  Great information about the parallel solver.
> 
> I've downloaded COPS and am looking into installing and setting it up.  It
> looks like it relies on MPI to handle the messaging?  We'd be happy to
> share our experience w/ building a cluster out with COPS if it works for
> us!

Yes, it uses MPI.

> A few questions before we go too far in the right / wrong direction:
> 
> 1.  We're using CBC right now to solve an MPS file.  Does COPS read MPS
> files?

Yes, BLIS will read MPS files.

> 2.  Our model runs around 3 hours at the moment, with CBC.  Do you have
> any idea what type of performance we could see with COPS?

It's difficult to say. We haven't done any direct comparisons between 
the two and it would depend on problem structure, etc. Generally 
speaking, CBC should outperform BLIS on a single processor, but for any 
given instance or class of models, of course, anything goes. There are 
lots of hooks for customizing BLIS (similar to those in CBC), so it 
might be possible to improve the performance by customizing if you are 
solving the same classes of models repeatedly....

> 3.  I've got an MPS file .... any chance I could get you to run it on your
> cluster and see what type of performance we're looking at .... before we
> spend several days rolling out a solution which might not work?

Sure, I could give it a try. We just got a new cluster here that we are 
playing with, so I can see what happens.

> Again, if this is a valid solution for us, we'll probably use EC2 as our
> architecture, and I'd be happy to share the results.

It would be great to see it running on something like EC2, so we're 
happy to give you whatever help you need.

Cheers,

Ted
-- 
Dr. Ted Ralphs
Associate Professor
Industrial and Systems Engineering
Lehigh University
(610)758-4784
ted at lehigh dot edu
www.lehigh.edu/~tkr2