Project Meeting 2021.06.08

Technical call

• Wu share SANDAG's 1, 2, and 3 zone system testing progress
  • Goal is better runtimes than existing model
  • Would be good to make sure specifications are similar enough for comparison purposes
  • Would be good to hold num processors settings constant (and to use the new chunking functionality next time)
  • Location choice models appear to degrade in performance as we move from 1 zone to 2 zone to 3 zone
  • Performance of 2 zone versus 1 zone and 3 zone should be closer to 1 zone?
  • Runs done with MKL setting
  • Tour mode choice is a lot slower in 2 zone, which is unexpected so why?
  • Would be good to review log files side by side to see performance of steps within submodels
  • Look at presampling, location models, logsums, etc.
  • Can also do testing/investigation with smaller sample sizes since patterns should be similar
  • We'll use 6b support tasks to help SANDAG with the task
  • Phase 7 will have funds to keep working on performance as well
  • SANDAG setup shared google doc for tracking results
  • Would also be good to setup a dedicated team discussion on the topic so we can more quickly make progress
  • Newman's benchmarking system improvements could help here as well
• Jeff Doyle share usability feature improvements
  • Auto-chunking, configurable checkpointing, and logging of zero prob expressions todos complete
  • No checkpointing could help a little with runtimes
  • Logging of zero prob expressions is a global setting (for all submodels) for now. If the need arises, then we can make it configurable by submodel
  • I'll test with and without checkpointing to see runtime differences
• v1.0 release just about ready
  • Wrapping up doc updates, all code merged to develop and passing tests
  • Plan to release later this week