Re: NAMD and multi-rail IB for improved latency

From: Axel Kohlmeyer (
Date: Fri Mar 27 2009 - 09:28:40 CDT

On Fri, Mar 27, 2009 at 1:05 AM, Dow Hurst <> wrote:
> Just found this link which doesn't make sense to me as latency seems more
> important than bandwidth for NAMD.
> Any thoughts?


thanks to using the charm++ and the fact that the real space non-bonded and
the k-space coulomb contributions can be computed independently, NAMD can
do a significant amount of latency hiding. if you want to build a high
machine (many jobs but each run somewhat slowish) then you can even get
away with tcp/ip and gigE. in the limit of scaling you'll first notice
the intrinsic
overhead of MPI (as charm++ has to do a lot of polling), this can be avoided
but compiling charm++ to run on the infiniband layer directly which
has communication
primitives that are much more suitable for charm++ than for MPI.

not sure what state that part of charm++ is currently in, though. last
time i tried
(quite a while back), it didn't work reliably on our local hardware.
i'd thus ask
the charm++ folks for help on this subject.

latency _is_ an issue for _other_ MD codes that cannot hide them.


> Best wishes,
> Dow
> Dow Hurst wrote:
>> We've got a couple of weeks max to finalize spec'ing a new cluster. Has
>> anyone knowledge of lowering latency for NAMD by implementing a multi-rail
>> IB solution using MVAPICH or Intel's MPI? My research tells me latency is
>> key to scaling effectively. Has anyone cut down latency to below 1.0us
>> using multi-rail IB?
>> If so, what was the configuration of cards and software? Any caveats
>> involved?
>> Anyone run NAMD on top of Appro's enhanced version of Intel's MPI in a
>> multi-rail configuration?
>> Thanks,
>> Dow

Axel Kohlmeyer
  Center for Molecular Modeling   --   University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
If you make something idiot-proof, the universe creates a better idiot.

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:52:31 CST