• Question re meeting and workload

Advancing subdomains in a 1D wave simulation

Walk through a tuned 1D wave code for running the time stepper (getopt-based argument parsing added after the meeting).

Breakout groups

  1. Look at the “telephone” example in the demos subdirectory. Add the requested MPI calls to actually pass the messages. Change the job.pbs script (PBS) to an appropriate job submission script for the Graphite cluster, and test the code there.
  2. Take the batched version of the wave stepper code and think through how you might implement an MPI version. What is the effective flop rate for this on Graphite on a million points? Given what you know from the MPI ping-pong tests, can you come up with a crude estimate of the speedup?