From: Victor Ovchinnikov (ovchinnv_at_MIT.EDU)
Date: Thu Sep 27 2007 - 21:28:46 CDT
Ilya,
I'm attaching a table that I generated while benchmarking my cluster.
I'm comparing with a quad dual-core opteron cluster (also via Gbt
ethernet). The binary that I compiled from source is indicated by
`compiled' in the heading; `bin' means precompiled binary.
Do you get something similar?
On Thu, 2007-09-27 at 18:48 -0700, Ilya Chorny wrote:
> Victor,
>
> What kind of scaling do you get? Mine is not very good. I think 8
> processors per nic is a bottle neck.
>
> Ilya Chorny Ph.D.
>
>
> On Sep 27, 2007, at 5:11 PM, Victor Ovchinnikov <ovchinnv_at_MIT.EDU>
> wrote:
>
> > Shulin,
> > The existing 64 bit binaries compiled for AMD processors (both TCP/
> > UDP)
> > work fine (as also suggested by previous posts on this topic). I
> > have a
> > cluster of dual 4-core Xeons that I've tried this on.
> >
> > The execution command is unchanged (essentially as you describe)
> >
> > One thing that I don't know, and wonder if others have any
> > information,
> > is how NAMD scales with the number of these 8-cpu babies connected by
> > myrinet/infiniband. I have Gbit ethernet, and so the scaling is
> > decent
> > for big systems only (>100,000 atoms) and only up to around 4 nodes
> >
> > Victor
> >
> > On Wed, 2007-09-26 at 15:08 -0700, Shulin Zhuang wrote:
> >> Dear All,
> >>
> >> I want to ask a basic problem.
> >>
> >> Our lab plans to buy a 8 core (dual quad-core Xeons) to run parallel
> >> NAMD. Can NAMD be directly run on it or shoud I recompile it.
> >>
> >> For the parallel run, after set the nodelist file, should I also use
> >> the command: charmrun namd2 +p8 config file >outfile ?
> >>
> >> Great thanks
> >>
> >> Best regards
> >>
> >> Shulin
> >
9/24/07; NAMD test case; 110,000 atoms
t(s) per timestep
|---------------------------- dual 4-core Xeon--clock 2.33GhZ------------------|-quad dual-opteron-870 (2.0GhZ)--|
NCPU amd-64-net compiled amd-64-net bin amd-64-tcp bin amd-64-net bin amd-64-tcp bin
4 0.69
8 .51 0.44/0.44 0.40 0.52 0.51
16(2 nodes) 0.27 0.25 0.35 0.32
32(4 nodes) 0.191/0.195 0.18 0.32/0.33
48(6 nodes) 0.163 0.156
This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:45:19 CST