Re: NAMD infiniband performance

From: Axel Kohlmeyer (akohlmey_at_cmm.chem.upenn.edu)
Date: Tue May 19 2009 - 17:13:50 CDT

On Tue, 2009-05-19 at 12:04 -0400, Guanglei Cui wrote:
> Dear Alex,
>
> I guess it's difficult to justify the cost of infiniband without
> infiniband :-) Anyways, I'll give MPI a try. Is there any particular
> implementation you recommend? And a related question ... how does

my personal preference is OpenMPI because of its flexibility
and convenience, but there also is MVAPICH. if you have an OFED
installation, you should have both already available.

> charm++ handle a heterogeneous system setup with core-core and
> node-node interfaces? Thanks a lot for the input.

yes, you have to adjust the compilation to add SMP or multi-core
support and provide some topology information through arguments
(e.g. +ppn). i have not experimented with it very much, but i am
certain, it is explained somewhere in the charm++ or namd2
documentation.

cheers,
   axel.
>
> Regards,
> Guanglei
>
> On Tue, May 19, 2009 at 11:00 AM, Axel Kohlmeyer
> <akohlmey_at_cmm.chem.upenn.edu> wrote:
> > On Tue, 2009-05-19 at 09:56 -0400, Guanglei Cui wrote:
> >> Hi Lev,
> >
> > dear guanglei,
> >
> >
> >> Thanks for the reply. I have not tried to compile Charm++ with any MPI
> >> implementations, and been using the net- option. Is there any
> >> advantage of using MPI (just any MPI)vs. using the net- option?
> >
> > yes there is an advantage. the net version always uses tcp/ip,
> > which has a high communication overhead. now for a small number
> > of nodes or a large system this does not matter much, as NAMD
> > can hide communication latencies through the ingenuity of charm++
> > over plain MPI. however, if you have an infiniband network, the
> > low latency that the infiniband communication offers, is only
> > available if you use a charm++ installation that uses MPI for
> > lowlevel communication or even the ibverbs layer directly.
> >
> > the difference between the two approaches will be larger the
> > more nodes you can use in your tests. the net version will scale
> > out _much_ faster than the custom compile.
> >
> > cheers,
> > axel.
> >
> >>
> >> Best regards,
> >> Guanglei
> >>
> >> On Tue, May 19, 2009 at 2:31 AM, Lev Lafayette <lev_at_vpac.org> wrote:
> >> > Hi Guanglei,
> >> >
> >> > Although I am no NAMD expert, I would suggest compiling from source and creating an application. As a binary you probably won't be picking up the performance benefits, which should be determined by whatever you're using as message parsing interface (OpenMPI?). Then you should seem some real gains.
> >> >
> >> > All the best,
> >> >
> >> > --
> >> > Lev Lafayette (03) 9925 4909, Systems Administrator, Victorian Partnership for Advanced Computing
> >> > "If you are sending a reply to a message or a posting be sure you summarize the original at the top of the message, or include just enough text of the original to give a context. This will make sure readers understand when they start to read your response."
> >> > RFC 1855
> >> >
> >> >
> >
> > --
> > =======================================================================
> > Axel Kohlmeyer akohlmey_at_cmm.chem.upenn.edu http://www.cmm.upenn.edu
> > 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.
> >
> >

-- 
=======================================================================
Axel Kohlmeyer   akohlmey_at_cmm.chem.upenn.edu   http://www.cmm.upenn.edu
   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 - 15:50:59 CST