Parallelization Scheme

What is the best Parallelization Scheme ?

• if you have lots of independant observations (scans) and few bolometers, use the PARA_FRAME parallelization
• if you have a few scans and a lot of bolometers, use the PARA_BOLO parallelization

Note that using the PARA_BOLO parallelization require that all your node/process access all the data, it is best to use a SMP architecture with a fast local disk.

What is the best CPU usage ?

If you want to obtain the best perfomance with SANEPIC with the PARA_BOLO option, it is best to bind the process to the cores when you launch your jobs:

• --bind-to-core --bysocket

  for Open MPI Project

• -binding cpu:cores

  for MPICH2 Hydra

Memory Usage

How much memory does SANEPIC require ?

SANEPIC keeps in memory 9 copies (double) of the filled pixels in the map (per process, 10 for the rank 0), 2 indexes (long) of the full map pixel, plus 2 to 4 times the lengths of a timeline (double).