AMBER Archive (2009)
Subject: RE: [AMBER] Problems with running two separate jobs in the same time
From: Ross Walker (ross_at_rosswalker.co.uk)
It would be helpful to understand exactly how you are running these jobs, what your computer environment looks like etc.
A few things to check.
1) Are you running these jobs in separate directories? There can often be problems running in the same directory since the two codes will compete to overwrite certain files such as mdinfo and profile_mpi. Depending on your filesystem and whether it is local or network mounted this can cause all sorts of problems from minor to major performance degredation to complete lock up.
2) Do you have enough disk space free? Are you certain the two jobs running together are not filling the disk?
3) Do you have enough memory. It is possible that the two jobs running together are running out of memory since they have to share it. Often when something is killed because you ran out of memory you see no error message other than perhaps the word 'Killed'. Note if you are using MVAPICH there are numerous memory leaks in it, especially in v0.9.9. This means that the memory usage, for MPI buffers I believe, can increase linearly over time. With just one job running you may have enough memory + swap that you don't see any problems within the length of the run. With two jobs running you may be pushing things over the edge.
4) You don't mention the version of AMBER. You could try running the calculation in AMBER 10 PMEMD in place of sander. This, in parallel, will use less memory and so you may not encounter the same problems.
for example, if I try to run two separate molecular dynamics or