charmm scalability on 2.4 kernels

[Steve Fellini]

Hi All,

Many of our users run charmm dynamics using Ewald/PME calculations,
which require 3D FFTs.  When run in parallel on a cluster, these
calculations scale moderately well to around 2-4 nodes (4-8
processors) when run on either nodes on a Myrinet network, or on nodes
running on fast ethernet with Linux 2.2.16 kernels with Josip
Loncaric's tcp fix (see
http://biowulf.nih.gov/charmm-bench/image003.gif).

However for nodes running on ethernet with Linux 2.4.12 (i.e., without
the tcp fix) scalability is very poor - so much so that there's no
point in running on more than one node (see
http://biowulf.nih.gov/charmm-bench/image004.gif and image005.gif).

Has anyone successfully configured/tuned the 2.4 kernel to improve
scalability of parallel jobs with non-trivial communications?

Thanks,
Steve

--								 
Steven Fellini
Center for Information Technology
National Institutes of Health 
steven.fellini at nih.gov