I have been runnning MD jobs with GULP (General Utility Lattice
 Program) on a 16 node Athlon based Beowulf cluster but have been
 unable to run consecutive MPI jobs without a remake of the source
 (just deleting the final source file and remaking - taking a second or
 two) each time. Although it works, it is a rather unsatisfactory way
 of working and limits me to one MPI job at a time.
 It seems that every time GULP finishes it somehow manages to
 corrupt the LAM MPI libraries, and when a second run of GULP is
 attempted an error occurs just as the first step is to proceed:
   Output for configuration   1
 MPI_Recv: process in local group is dead (rank 0,
 Rank (0, MPI_COMM_WORLD): Call stack within LAM:
 Rank (0, MPI_COMM_WORLD):  - MPI_Recv()
 Rank (0, MPI_COMM_WORLD):  - MPI_Reduce()
 Rank (0, MPI_COMM_WORLD):  - MPI_Allreduce()
 Rank (0, MPI_COMM_WORLD):  - main()
 Looking at the source code this seems to be the point in the code
 where the memory allocation takes place. I have therefore tried all
 combinations of -DMALLOC and -DF90 in the getmachine setup
 but none seems to fix the problem.
 I am using LAM-MPI compiled with the GNU compilers and
 configured to use the pgf90 compiler under Linux.
 Has anyone managed to compile a more robust LAM-
 MPI/pgf90/GULP setup (or similar), or does anyone have any
 suggestions of how to do so. An mpirun of a consecutive GULP job
 with np 1 gives no problem only with 2 or more processors. I have
 also tried using MPICH but I couldn't even run/compile/link GULP
 sucessfully then.
 Many thanks for your help in advance.
 Dr Stefan T. Bromley
 Laboratory of Applied Organic Chemistry and Catalysis
 DelftChemTech, Delft University of Technology
 Julianalaan 136, 2628 BL Delft
 The Netherlands
 Phone  : + 31 1527 89418
 e-mail  : S.T.Bromley - at -