Re: NAMD 2.7b1 running & performance issues

From: Marius Micluta (marius_at_biochim.ro)
Date: Mon Jun 01 2009 - 07:15:56 CDT

Dear Axel and Nicholas,

Thank you for your replies. I tested your suggestions and, although they did
not solve my problem, I found interesting things.

Building Charm++/NAMD with the latest (11.0.083) Intel compiler runs smoother
and succeeds irrespective to the optimization flags I use. However, I noticed
no significant gain in speed between the versions compiled with various
optimization flags, from -O0 to -O3, but I did not run extensive tests. As
compared to the code compiled with the 10.1.011 compiler, the code generated by
the latest Intel compiler runs 3% faster, but meanwhile I also enabled Jumbo
frames on the Ethernet switch, for which I previously lacked the administrator
password. Unfortunately, changing the compiler and the optimization flags does
not solve my sporadic freeze at startup problem.

I also tested the precompiled UDP version of NAMD-2.7b1 (which I had never used
before, trusting the NAMD documentation stating that the TCP version might be
faster on Gigabit clusters) and it is indeed much faster than the TCP version
on my cluster (about 3.5 times faster than the TCP version and 2.5 times faster
than the MPI version). However, it is 25-30% slower than the SMP version. On
the other hand, I found no significant gain in performance running the NAMD-2.6
precompiled UDP version on an old 32 bit/2.4.x Linux kernel cluster.

I had discovered the Norma cluster before posting my first message and I was
amazed by its performance. After reading your message, I examined more
carefully the startup file published on the Web site and noticed that you used
rigid bonds and a timestep of 2.0 fs. Using similar parameters, the UDP
benchmarks on our cluster give results close to your figures.

My guess is that both the performance and lock at startup issues are network
related, but I lack the means of getting a deeper insight.

Kindest regards,

Marius Micluta

--
Marius Micluta
Institute of Biochemistry of the Romanian Academy
296 Splaiul Independentei
060031 Bucharest
Romania
Tel: +40 21 223 90 69
Fax: +40 21 223 90 68

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