Casa Parallelization meeting minutes

Monday June 18th Room 317, 3:30pm MDT

Attendees:

Targets/tasks for June 18th - 25th

• Complete installation of test hardware James
• Complete or nearly complete test script James
• Begin completion of pclean vs clean parametrization Kumar
• Identify pcube ALMA tester Brian
• Identify large number channel data set Butler

Discussion/Agenda

• Benchmarking outline for final cluster orders
  • Use 400GB 3c147 data set, large A array image with 1K channels
    • Sufficiently large to measure multinode performance with pcube and pclean
    • Re-demonstrate non-imaging vs imaging task times and I/O performance
    • Effects of per core and per node parallelization per task
  • Performance against variety of hardware
    • 5500, 5600 Westmere Xeon 6core processorss, E5-2400 and E5-2600 Sandy Bridge Xeon 6 and 8 core processors
      • Generate various imaging/cost measures (e.g. visibilities imaged per dollar) (assumption is large number of cheapest processors is best)
      • Measure parallelization performance and impacts due to cache starvation
    • Dell, SGI and Supermicro motherboards with similar processors and memory
      • Measure Infinband and over all performance and cost
  • Demonstrate memory impacts for various sized images, particularly swapping impact on performance
    • Need better description of what imaging tasks and sizes must be supported
  • Ultimately generate optimal price configuration of processor vs memory tradeoffs vs specific imaging cases to create hetergenous cluster config
• Discussion of priorities for 3.5 imaging
  • Full parameter implementation of clean to pclean (e.g. multifield, utilitarian options) versus memory consumption issues
    • Devote 2 weeks to pclean implementation
    • Debate about examine memory vs scratchless/cal library interface

• Memory requirements in imaging
  • Total memory scaling is generally understood
  • Kumar has identified expected vs realized memory usage discrepency.
  • Need to understand better how temp image selection works
• Describe and implement engine vs threaded gridder determination as a function of memory
  • Need dynamic image partitioning scheme for 2, 3, 4 or more threads.
  • (Total memory - some delta ) / memory per engine = number of engines
  • int (Total cores / number engines) = number of threads per engine
• Update on Partitioning status (null selections, non-conforming tasks) Jeff
  • First tasks being implemented (flagger, calibration), some read in parallel, all can read serially, writers are still problematic
  • Re-examination of Parallel-go for cluster specification (cores vs engines, memory)
• Model data tests James
  • Pending completion of test script
  • James to get with Kumar in late June
• Memory resident subtable tests James late June
  • Pending completion of test script
  • Check that it affects number of FDs as well
• Spectral imaging tests James
  • Understand channel chunk choice effect on performance James
  • Pending completion of test script

CASA HPC Initiatives for the 3.5 Cycle

• Tentative list from last meeting
  • Reduce in memory requirements during imaging (maybe removed)
  • Parallelize multi-field imaging
  • Parallelize MFS terms > 1 in the gridding step
  • Parallelize Cube imaging scatter gather
  • Expand threaded gridder case to >4 threads
  • Implement core affinity for threaded gridder case (first out)
  • Parallel framework (partitioning, parallel go) work in ESO
  • Filler to write a MMS in parallel

Deferred Topics

ARDG Targets