AMBER Archive (2000)

Subject: amber on AMDs cluster

From: Xavier Deupi (xavi_at_yogi.uab.es)
Date: Wed Jun 14 2000 - 07:35:41 CDT

Hi all,

sometime ago I posted a message asking about beowulf-type clusters to run
AMBER. I had some feedback with some of you (thanks to all!), so I
realized there's some groups working on it.

Now we're really on it, and we have a couple of questions...

We've got a couple of AMDs K7 650 to test, and the scaling is really
desapointing.
Running sander with a system of ~9000 atoms for 1ps takes ~15min. When we
use the 2 processors, it takes... 19 min!

It seems to be a communication problem, because each node works under 50%
of its power, probably waiting too much for the communication.
The communication between the nodes looks ok (ping time of 0,1 ms), so
this should not be the problem.

We're using LAM-MPI as communication protocol.

So our questions are:

1) Do people usually use PVM or MPI?

2) Using MPI, what are the usual scaling rates, going from 1 to 2 nodes?

3) Anyone has tested this king of computation on AMD K6 ?

4) Anything against the following command line of mpi :
        "mpirun -v -c2c -c 2 -w -O sander input_file..." ??

5) Any suggestions ???

Thanks in advance,
xavi

________________________________________________________

  Xavier Deupi-Corral
  Computational Medicine Lab
  Biostatistics Unit. School of Medicine
  Universitat Autonoma de Barcelona. CATALUNYA (Spain)
  Phone : (3493)-581.23.48
  E-mail: xavi_at_yogi.uab.es
________________________________________________________