[gmx-users] MVAPICH2 mdrun problem for SD and MD, GROMACS 4.0.5

[Mark Abraham]

Daniel Adriano Silva M wrote:
> Dear Gromacs users,
> 
> I am experimenting the next problem on an infiniband-cluster (8
> intel-cores per node, GROMACS compiled with icc 11.1, all run through
> mvapich2):
> 
> I have a molecule (protein 498aa, solvated or in vacuum I get the same
> problem at any box shape), when I try to SD minimize it with
> mvapich2-mdrun, it minimizes well with 1, 2 or 3 cores and reaches
> convergence in around 1000steps, however any further combination (4,5,
> ...n cores) makes it to immediately stop (less than 20 steps) with:
> 
> "Steepest Descents converged to machine precision...".
> 
> Further if I take "the 1 core minimized structure" and try to make a
> solvated-pr dynamics(2fs, MD, NTP, etc.) it also works with 1
> processor, but with more cores it begins immediately to bring LINCS
> warnings: and dies:
> 
> "Too many LINCS warnings" or "Water molecule starting at atom 16221
> can not be settled"
> 
> For a "long time" I had made another md simulations on this cluster
> with the same mdps and other proteic systems, and I only see this
> behavior with this particular protein, of course  before send this
> mail I re-tested previuos-working tprs.
> Finally, the most suspicious is that I have another very similar
> 8-core box (with the same processors) but with gromacs gcc compiled,
> and it actually runs very well the same problematic molecule (even the
> same tpr) with mpi and 8-cores.
> What do you think??? Please, if you have some tpr to test something send it.

I'd guess you're having some problem with (dynamic) linking of the MPI 
library. Perhaps the version of some library has changed since recently, 
etc. I'd suggest compiling two fresh copies of GROMACS with either icc 
and gcc on the troublesome machine and seeing what happens with them.

Mark