NetCDF for Parallel I/O

• Each PE writes its subdomain as hyperslab
  • NetCDF output array specified with global array dims.

  • At least one disk seek per PE

• CM5 performance (Rick Light's optimized library)
  • Dump 3D 64-bit array 32x32x16 (potential from PIC code)
    • 3D domain decomposition

    • Measured 1-2 MB/s using 1-128 PEs
      • Adequately small impact on run time

• T3D performance (Steve Luzmoor's optimizations)
  • In progress; early indications are O(10) MB/s

Generated with CERN WebMaker