Parallel Error Calculation using MPI and SLmpi

Parallel Error Calculation with MPI and SLmpi

• For interactive error calculation in isis you can use conf or conf_loop (using Isis_Slaves.num_slaves+1 cores on the host machine).
• Fast possibility: mpi_fit_pars. It's a routine in the isisscripts, which uses MPI to calculate the confidence intervals in parallel on different machines.

• Requirements:
  1. SLmpi installed: http://www.sternwarte.uni-erlangen.de/git.public/?p=slmpi.git
  2. A Portable Batch System (PBS), e.g., Torque (see Torque for an Introduction)
  3. the Remeis-isisscripts http://www.sternwarte.uni-erlangen.de/git.public/?p=isisscripts

• Create an isisscript to be submitted via PBS:
  1. load and group data, load best fit model
  2. call result = mpi_fit_pars([pars]);
  3. if you have the isisscripts you can store everything in a fits-file via fits_save_fit(filename,result)

• How to start the Job on the Remeis Cluster?
  • Create a job-file with mpi-Option (e.g. by qsub_array -p %%--%%mpi file.com)
  • the Command File file.com contains the list of scripts to be executed.
  • Set nodes to the number of free parameters. (See below for an example).

#!/bin/bash 
#PBS -S /bin/bash -V 
#PBS -t 0-0\%10000 
#PBS -l nodes=11 
#PBS -l walltime=03:00:00 
#PBS -N file.com 
#PBS -o example_script.out  
#PBS -e example_script.err 
cd /home/user/datadir 
COMMAND[0]="mpiexec isis-script example_script.sl" 
/usr/bin/nice -n +15 ${COMMAND[$PBS_ARRAYID]}