Charm++ parallel programming system

[Chris Harrison <charris5 AT gmail.com>]

Marcel,

I am forwarding you question to the charm++ list for further help.

Best,

Chris

---------- Forwarded message ----------
From: Marcel UJI (IMAP) <arzo AT uji.es>
Date: Fri, Mar 9, 2012 at 6:49 AM
Subject: namd-l: Compiling NAMD
To: namd-l AT ks.uiuc.edu

Dear all,

I have been compiling NAMD for my exact machine, an Intel Xeon E5345 processor, to get the most of it. However I have found that using -march=native (or -march=core2, which seems also correct) gets SLOWER namd code!!

I'm using gcc compiler version 4.4.5 and introducing the option via a configuration like this:

./config Linux-x86_64-g++ --charm-arch net-linux-x86_64 --with-cuda --cuda-prefix /usr/local/cuda/ --cxx-opts "-march=native -mtune=native" --cc-opts "-march=native -mtune=native"

The difference goes from about ~2.72 days/ns in a ~450.000 atoms MD-NVT system with external field applied to 3.33 days/ns using the above optimizations. charm++ was also compiled using gcc but without any optimization.

Do someone have an explanation for this??

Thank you in advance

Marcel

--

Chris Harrison, Ph.D.

NIH Center for Macromolecular Modeling and Bioinformatics

Theoretical and Computational Biophysics Group
Beckman Institute for Advanced Science and Technology
University of Illinois, 405 N. Mathews Ave., Urbana, IL 61801

http://www.ks.uiuc.edu/Research/namd       Voice: 773-570-6078

http://www.ks.uiuc.edu/~char                          Fax: 217-244-6078