Dear all,
I'm trying to run a protein simulation using PMEMD on CPUs, and I'm getting some strange (and possibly worrying) behaviour. When analyzing the trajectory with cpptraj, I get recurrent spikes in bond
energy and angle energy (see attached file). To figure out what is happening, I checked the positions of the atoms residue by residue, and I found that each spike corresponds to several bonds just
about a bit overstretching, but doing it all together. To me, this seemed like a synchronization artifact, so I restarted the run for a while longer on a different number of processors (and on a
different machine, for that matters). What I found is that, passing from 128 to 8 CPUs, the density of spikes decreases. In the attached plot, the last 50 ns were run on 8 processors, whereas the
previous 200 were run on 128. The density of spikes in the last 50 ns is exactly 4 times smaller than in the previous 150 ns, which seems to be quite good proof that what I'm observing is, indeed, a
synchronization artifact between parallel tasks (the system also contains 32238 OPC waters). Note that, in my input file, I do explicitly use ig=-1.
So, am I correct in my thinking, or are there possible alternative explanations?
Cheers,
Charo
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Mar 20 2024 - 05:00:02 PDT
