Author Topic: TURBOMOLE 5.9.1 on IBM Power5+  (Read 6183 times)

I.Swart

  • Newbie
  • *
  • Posts: 4
  • Karma: +0/-0
TURBOMOLE 5.9.1 on IBM Power5+
« on: September 10, 2007, 12:29:28 PM »
Dear all,

I'm trying to run TURBOMOLE on an IBM Power5+/Linux system using the loadleveler batch system. A description of the machine can be found here:
https://subtrac.sara.nl/userdoc/wiki/huygens/description
The serial binaries work fine. However, when trying to run the parallel binaries problems arrise.

The 'sysname' script prints out the correct '*_mpi' directory. A test calculation (ridft calculation of a co2 dimer) requests 16 processors on a single node. The job seems to start okay. The ridft output file gives the usual information when a parallel calculation starts. Also the 15 'slave*.output' files are generated. However these are all empty and the calculation exits.

Does anyone else experienced simiar errors before? Also did anybody manage to run TURBOMOLE v5.9.1 in parallel on Power5+/linux? Any help is highly appreciated!

Best regards,
Ingmar Swart

Dept. Chemistry, Utrecht University, the Netherlands



I.Swart

  • Newbie
  • *
  • Posts: 4
  • Karma: +0/-0
Re: TURBOMOLE 5.9.1 on IBM Power5+
« Reply #1 on: September 14, 2007, 08:16:03 AM »
Just a note to inform you that the problem has been solved! The code is running properly now! The problem was a too small job assigned to too many processors. It turns out TURBOMOLE crashes when one of the cpus isn't assigned a task.......

It's good to know that 16 cpus are (way) too many for a 54 electron system! 

regards,
Ingma