Re: parallel execution speedup

From: bo liu (liubo.njuer_at_gmail.com)
Date: Thu Aug 28 2008 - 01:51:29 CDT

Hi Roman and all,

In addition,for my experiences,

the benchmark you get is just the speed at the time NAMD anticipates, maybe
not the real average speed of the simulation.
many factors account for this benchmark, for example, if your system is far
from equilibration, the benchmark reported in the beginning of the
simulation would be very different from the real one.

That's my understanding, if i made any mistakes please correct me :)
Cheers!
2008/8/22 Eric H. Lee <ericlee_at_ks.uiuc.edu>

> Yes, the beginning of the log file will show the benchmark speed for how
> many CPUs you used for that particular run.
>
> - Eric
>
> On Aug 21, 2008, at 4:28 PM, Roman Petrenko wrote:
>
> Thank you for you responses,
> so the time i see in namd.log file is the actual speed of calculations
> using multiple CPUs, right?
>
> or is it the calculation speed of just one CPU?
>
> How can i check our network connection speed limit? What other info
> should i supply?
>
>
> On Thu, Aug 21, 2008 at 5:07 PM, Eric H. Lee <ericlee_at_ks.uiuc.edu> wrote:
>
> As a rule, the speedup one obtains with scaling with greater number of CPUs
>
> is based on a variety of factors. One of these are the number of patches
>
> NAMD divides your system into for independent calculation on individual
>
> nodes. Depending on the shape of your system, the number of patches
>
> generate may roughly correspond to the number of atoms in the system you
> are
>
> simulating as well. Historically NAMD achieved excellent linearity in
> speed
>
> up down to a limit of 1000 atoms per processor. However, this performance
>
> has been improved somewhat in the latest version that you can grab from the
>
> NAMD CVS, which scales quite well to 500 atoms per processor from my
>
> experience.
>
> In practice, one should benchmark simulation speed before a full production
>
> run, using different numbers of processors, and choose the one with the
> best
>
> tradeoff between computation speed and cost in processor hours.
>
> -Eric
>
> On Aug 21, 2008, at 3:51 PM, Roman Petrenko wrote:
>
>
> Dear all,
>
> when i run a job on one cpu i have this speed
>
> Info: Initial time: 1 CPUs 0.196443 s/step 1.13682 days/ns 17304 kB memory
>
>
> but when i run it on 6 computers with ppn=4, i get this speed
>
> Info: Benchmark time: 24 CPUs 0.0710909 s/step 0.411406 days/ns 8328 kB
>
> memory
>
>
> the question is: do i have mere 3 times speedup using 24 CPUs???
>
>
> --
>
> Roman Petrenko.
>
> Physics Department
>
> University of Cincinnati
>
>
> Eric H. Lee
>
>
> Medical Scholars Program
>
>
> Theoretical and Computational Biophysics Group, UIUC
>
>
> ericlee_at_ks.uiuc.edu
>
>
>
>
>
> --
> Roman Petrenko.
> Physics Department
> University of Cincinnati
>
>
> Eric H. Lee
>
> Medical Scholars Program
>
> Theoretical and Computational Biophysics Group, UIUC
>
> ericlee_at_ks.uiuc.edu
>
>

-- 
-Liu bo
-----------------------------------------------------------
Computational Biochemistry&Biophysics, Nano-bio systems: MD method;
College of Chemistry and Chemical Engineering, Graduate University of
Chinese Academy of Sciences, Beijing P.R. China
Office: (86)-010-88233187
Home: (86)-010-88259765
Cell: 13426057875

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:48:13 CST