different results for parallel vs serial runs for NAMD

From: Yousung Jung (yousung_at_caltech.edu)
Date: Mon Apr 16 2007 - 23:04:56 CDT

Dear NAMD users,

I've been playing with NAMD lately (as an MD beginner), and while
doing a quick benchmarking on the parallel performance of NAMD, I
realized that the parallel version yields a slightly different result
than the serial calculation. 2 cpus were used within a single node
for a parallel job.

In the minimization, both parallel and serial runs start with the
same energy, but then slowly deviate from each other as iterations
accumulate as below:

Parallel vs Serial

0 = 39049347.3758 vs 39049347.3758
100 = -184673.8475 vs -184673.8514
200 = -196574.6772 vs -196574.6758
500 = -207848.1239 vs -207848.0719
750 = -211889.4635 vs -212039.551
1000 = -214041.1238 vs 214074.6360

Evolution of the deviation doesn't look so much systematic either.
Has anyone come across this erratic behavior before? I'm using 64-bit
Opteron computers, and a precompiled binary from the NAMD website.

Thanks very much!


This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:44:34 CST