From: Niklaus Johner (nij2003_at_med.cornell.edu)
Date: Fri Apr 26 2013 - 10:17:00 CDT
Hi,
My understanding is that PLUMED does indeed use only one processor, but this is just for the calculations needed for your collective variable, which is usually very small compared to the processing that has to be made for all the forces in the system. The latter are still spread on multiple processors and the GPUs. Therefore this doesn't slow down your namd run (except in some extreme cases, such as using as collective variable the contacts between two large sets of atoms for example).
You should see that right away form the performance of your runs anyway, but I have used PLUMED-NAMD with GPUs and it works fine.
N.
Niklaus Johner
Weill Cornell Medical College
Harel Weinstein Lab
Department of Physiology and Biophysics
1300 York Avenue, Room D-501
New York, NY 10065
On Apr 26, 2013, at 10:46 AM, Francesco Pietra wrote:
Hello:
This is to ask whether there is any plan to extend NAMD-CUDA to bias-exchange metadynamics, temperature-exchange metadynamics, and reconnaissance metadynamics. This is just to plan with my work.
I was not careful enough to appreaciate that PLUMED-NAMD is at a single processor. I compiled NAMD2.9-CUDA4.0 with PLUMED1.3, including recon, and it seemed to run on multiple GPUs, by inserting commands for PLUMED into the NAMD conf file (just a "distance" job on a large protein I was engaged with). According to nvidia-smi, all GPUs were involved, and the log file showed PLUMED involved. Now, having been informed that PLUMED is at a single CPU with NAMD, it must have been fake success.
Thanks
francesco pietra
This archive was generated by hypermail 2.1.6 : Tue Dec 31 2013 - 23:23:10 CST