Switching from GPU to CPU: failing with too many CPUs?

From: Smith, Harper E. (smith.12510_at_buckeyemail.osu.edu)
Date: Thu Sep 23 2021 - 10:41:47 CDT

Hi mailing list,

I was running some equilibrations of my ~500k atom system on GPU, then decided to benchmark to potentially speed things up. Strangely, everything works under 480 CPUs, but the simulation slows dramatically above that: at 560 CPUs, I get 99 steps in 15 minutes (compared to 5,000 steps in 66 seconds for 480).

For the slow (560 CPU) simulations, all the output values are comparable to the faster (480 CPU) ones. I have run the faster simulation for some time, and the trajectories are fine.

In the TIMING lines from the slow simulation, it predicts that 5,000 steps will be completed in ~0.04 hours (~150 seconds), despite never completing in 15 minutes.

Any idea what may be causing this? My initial equilibrations were GPU via NAMD 2.12, and I am continuing on CPU for NAMD 2.13.

Harper Smith

This archive was generated by hypermail 2.1.6 : Fri Dec 31 2021 - 23:17:11 CST