Iím trying to compile charmm31a1 to run in parallel on an IBM cluster with shared memory and several 4- and 8-processor nodes. Jobs are submitted with Loadleveler and nodes are managed by POE.

I tried installing with these options:

./include.com ibmaixmp large FULL M

It compiles, but I run into two different errors. When running on one processor, I get the message

VEHEAP> Error: MALLOC returned ZERO.

as soon as it begins a minimization. When I use two processors charmm runs properly for the first several lines of the script but then stops without returning an error message, and LoadLeveler returns the following:

ATTENTION: 0031-408 2 tasks allocated by LoadLeveler, continuing...
ERROR: 0032-117 User pack or receive buffer is too small (4000) in MPI_Bcast, task 1
ERROR: 0031-250 task 1: Terminated
ERROR: 0031-250 task 0: Terminated

Iíve tried several combinations of options in the pref.dat file, and in one instance (sorry I donít remember the settings) the second error became something very similar but without reference to MPI, simply

ATTENTION: 0031-408 2 tasks allocated by LoadLeveler, continuing...
ERROR: 0032-017 Receive buffer too small, task 1
ERROR: 0031-250 task 1: Terminated
ERROR: 0031-250 task 0: Terminated

I tried borrowing the pref.dat settings from a successful installation of charmm 29b1 on the same cluster, but the run still crashed when I ran it on multiple processors. (pref.dat contained IBMSP UNIX SCALAR LARGE PUTFCM FCMDIR=fcm OLDDYN PBOUND PARALLEL PARAFULL SYNCHRON RCFFT NOGRAPHICS END). I need the new version to use GBSW.

I also played with environmental variables that I thought might matter (MP_EAGER_LIMIT, MP_BUFFER_MEM), but the execution crashed in exactly the same place regardless of these settings.

Now Iím stuck and donít know what else to try. I could be overlooking the obviousóI havenít done this before. Does anybody have suggestions?

By the way, is there any substantial difference between developmental versions of charmm and standard releases?

Thanks much.


Nathan Hammond Massachusetts Institute of Technology