Arkouda scalability runs

Benchmarks run on up to 512 nodes of a Cray-XC with Aries and 32 nodes of a Cray-CS With FDR InfiniBand. Per node hardware is similar. Using Arkouda from 12/12/19: https://github.com/mhmerrill/arkouda/commit/da0a124

Ran stream.py, argsort.py, gather.py, scatter.py, scan.py, and reduce.py. The data for scan/reduce isn't collated or graphed, but the raw data is available. The default problem size ~3/4 GB per node was run for both CS and XC. For XC there are also runs that use 16 GB per node (8TB at 512 nodes.)

• 36-core (72 HT), 128 GB RAM
  • dual 18-core (36 HT) "Broadwell" 2.1 GHz processors
• Aries network
• Software:
• CLE 7.0.UP02
• cray-python 3.7.3.2
• Chapel 1.20.0
• ZeroMQ 4.3.2
• cray-hdf5 1.10.5.2

• 36-core (72 HT), 128 GB RAM
  • dual 18-core (36 HT) "Broadwell" 2.1 GHz processors
• 56 Gb FDR InfiniBand network
• Software:
  • RHEL 7.6
  • Python 3.6.8
  • Chapel 1.20.0
  • ZeroMQ 4.3.2
  • HDF5 1.10.5

Dir structure:

• . -- contains combined .gpi and generated .pdfs
• cray-{cs,xc}
  • graphs/ -- contains .gpi files and generated .pdfs
    • data/ -- contains data collated into .dat files
    • raw-data/ -- contains output from runs
      • print.sh -- grep time/perf (modify for size and time/perf)
      • run.sh -- script I used to run problem sizes