Wrong energies in CUDA version

From: Jeff Wereszczynski (jmweresz_at_mccammon.ucsd.edu)
Date: Thu Jun 03 2010 - 11:45:13 CDT

Hello NAMD List,

I am trying to get a CUDA version of NAMD running. We have access to
a GPU cluster with each node having 8 CPUs and either a Tesla C 1060
GPU or a Tesla C 2050, with the cuda 3.0 libraries installed. I was
able to compile a namd executable using the with-cuda flag which runs
and detects the gpu, but the simulation quickly crashes with RATTLE
errors. I tried to do a minimization and I observed that the energies
at step 0 were all correct except for the pressure, gpressure,
pressavg, and gpressvg columns. As I continue the minimization these
terms fluctuate wildly and the potential energy of the system explodes
as the vdw terms increases drastically (atoms begin to overlap with
one another).

I know pressure controls do not work in CUDA NAMD, but I am running
NVT so I would think that NAMD would ignore these terms. I've tried
this with both 2.7b2 and the CVS version and got the same results.
Additionally, I tried this simulation (same input files) on a local
GPU with the pre-compiled binary and it worked fine (the simulation
progressed and the pressure values were normal). One major difference
was on our local computer we had CUDA 2.3, whereas on this cluster
only 3.0 is available, could this be the cause of the problem? Does
anyone have any ideas of other things I could try to get this to work?

Thanks,
Jeff

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:55:50 CST