[gmx-users] Biomolecule shuffling in parallel runs crash MD

[Marcin Wojciechowski]

Hi fellows,

Problem is as follows:

I have a system with one protein neutralized with ions and solvated in
water box. I runned it for a total 1ns of MD on 16 cpu cluster. Recently i
tried to start it on 32 cpus ('cos i have scali mpi and scaling on 32 is
acceptable, approx. 60%). Unfortunatelly i always had an error regarding
LINCS warnings and terminating the simulation. The facts I observed:

1. Division over CPUs is like 1 protein on 1 cpu and some waters, plus
rest of the water on other cpus (shuffle and sort used): no problem,
simulation goes fine.

2. Division over CPUSs is like 1 protein and 0 waters (when you use more
cpus), plus rest of the water on other cpus (shuffle and sort used): MD
crash after first step (LINCS warnangle warnings).

3. Turning off shuffle and sort does't work. MD also crash.

So to choose maximum allowable number of CPUs i have to make trial&error
runs of grompp and select MD where at least some waters "accompany"
protein on the same CPU.

Have anybody encountered somethig similar?

I runned "homogenous systems" like dppc benchmark up to 64 CPUs and it was
fine.
My protein system was after 1 ns so i assume it's equilibriated enough.

Best regards, Martin.