S2HAT - performance results:



Description of the tests

The performance test have been performed on two computational platforms at NERSC Supercomputing Center in Berkeley: 1. IBM SP RS/6000 computer called Seaborg, running AIX UNIX
2. Opteron cluster called Jacquard and running LINUX
The test runs were all assuming full sky polarized maps and optimal load-balancing in both pixel and harmonic domains as performed by the auxiliary routines (distribute_map, distribute_alm, etc) provided in the package. For each of the transforms and platforms two cases were investigated: 1. with fixed total work load, i.e., nside = const, and thus the one in which the amount of the data and work load per processors is decreasing with a number of allocated processors.
2. with fixed theoretical work load per processors, i.e., in which the map size grows together with a number of processors. i.e. nside^3/nproc = constant.
In all runs lmax = mmax = 2 nside.

The results (not yet complete)

  • map2alm transforms:
  • i. fixed total work load (nside = const)
    ii. fixed per-processor work load (nside^3/nproc = const)

  • alm2map transforms:
  • i. fixed total work load (nside = const)
    ii. fixed per-processor work load (nside^3/nproc = const)


    radek stompor 2007-02-13