[gmx-developers] BUG appears with intel-compilers

[Berk Hess]

H.J.Risselada wrote:

> Dear GMX-users/devolopers,
>
> I noticed a fatal bug in gromacs versions 3.3 (>3.3)
> When this version is compiled by intel-compilers and mpiecex
> as a message passing interface (like the standard setup on the SARA, 
> grid clusters) it remains in a "segfault" IN ANY CASE as soon as the 
> non-bonded c.q bonded force routines are touched. When GNU compilers 
> and mpirun are used however this problem disappears. All older 
> versions work smoothly on the same tpr files in any case.
>
> Although this bug can be avoided, using GNU compilers on intel 
> processors will lose some computational efficiency.
> Didn't feel brave enough yet myself, to track down an architecture 
> dependend bug in the dark woods of the non-bonded force routine ;)
>
I can vaguely remember having heard of a problem like this before.

But what do you mean with mpirun?
Many mpi implementations use a binary called mpirun.

Do you know if the error is related to the use of the intel compiler or 
the mpi implementation?

Could you compile a version with CFLAGS+=-g as an option to configure
and run with a debugger to find out where is crashes?

PS Intel compilers will not make Gromacs significantly faster,
as most time is spent in the innerloops which are assembly code.

Berk.