Agenda

• AMPO Conference
• Update on Phase 9a Optimization task

Action Items

• Everyone to review the AMPO conference proposed panel discussion submittal. Joe to review and edit, particularly in relation to the criteria that Caitlin provided.

Meeting Notes

AMPO Conference

• Michelle shared a draft proposal for a panel discussion, recycling what was proposed and sent for the 2022 conference.
• Folks agreed that the format/structure last time can be used again this time.
• Caitlin sent out an email with a link that goes over how to put together a good proposal and requested that we be sure to specifically address the 3Ps outlined as that is what the committee will be scoring you against. Joe will review the abstract in this context.

Update on Phase 9a Optimization task

• Several new PRs and issues were added, some key ones include:
  • School escorting optimization - this takes out unnecessary calculations (no need to run households without kids through the school escorting model)
  • Crash during sharrow compiling (https://github.com/ActivitySim/activitysim/issues/811)
  • Jeff opened a PR to update some packaging infrastructure.
  • Jeff found a bug and opened a PR related to the categorical variables that are attached to school escorting participants.
• Sijia reported on the latest results in profiling memory usage and run time
  • Good news
    • Vehicle type choice model improvements worked and reduced run time for that model in sharrow mode from 17 hours to 16 minutes.
    • In non-sharrow mode, the crashed because of a memory issue but the implemented fixes dropped peak memory usage from 500GB RAM to 200GB RAM.
  • Bad news
    • Both runs (with and without sharrow) are still crashing in trip mode choice for school purpose (high school), trying to choose the mode for a trip that is 146 miles long. The team is just starting to look into why. There could potentially be two bugs – school location choice and trip mode choice (relating to escortees). Engineering team to dig into these things more offline. Sijia to open issues: https://github.com/ActivitySim/activitysim/issues/814.
    • Noticed that sharrow is not releasing the memory – Sijia created an issue
  • David Hensle reported on school escorting.
    • Previous, the school escorting model was running for all households. A new PR removes households without children from the choosers table, which resulted in a drastic improvements in memory usage.
    • There are still has some school escorting performance questions (run-time issues) to address. There’s a slow, low memory process towards the end (the “tails” after the choices have been made, there’s some processing to bring results into the pipeline). The question is whatever is taking so long, could that be optimized?
• Sijia is comparing individual steps between legacy (no sharrow) and with sharrow to see where sharrow is taking longer. If any step is taking longer with sharrow, that should be an issue to look into. For example, non-mandatory tour scheduling and trip purpose/destination are taking longer in sharrow and needs to be reviewed. Small increases are happening in vehicle allocation, at work subtour mode choice, stop frequency, etc.
• Things still to do:
  • Still need to do BayDAG PRs, but want to do more work on optimization and fixing these things before merging.