Project Meeting 2021.03.16 - ActivitySim/activitysim GitHub Wiki

Technical Call

  • Discussed new issues and PRs
    • #390 - need to document and then close issue
    • #392 - Jeff to clean-up
    • PR #394 - Will merge draft telecommute model branch into develop
  • Merged develop2 to develop and deleted stale branches now that tests are passing
    • Newman can now merge develop into estimation_larch and issue #382
    • I need to work on the user guide updates: additional estimation functionality, tutorials, presampling, rep logsums, telecommute, test system improvements, etc.
  • Discussed representative logsums progress and TOD distribution before and after results match well
  • Discussed presampling and zone distribution before and after match well
    • The big zones are a bit under selected and the small zones a bit over compared to without presampling
  • Discussed non-mandatory tour scheduling expressions coding improvements since the expression file is inefficient
    • Add the ALTS PREPROCESSOR from mandatory tour scheduling that Clint added for ARC to pre-calculate the shift variables once
    • Segment by tour purpose
    • Put the ALTS PREPROCESSOR in a more efficient spot
    • The result will be a lot less calculations and much faster runtimes
    • We will take a look and maybe squeeze into the maintenance task
  • Discuss PSRC testing progress
    • PSRC updated runtimes here
    • Need to check on number of failed trips in trip_purpose_and_destination models since may have a data issue and/or funny results due to using mazs
    • Overall runtimes of 3.5 hrs in the reasonable ballpark
    • PSRC found 26 processors for a 32 processor machine is the sweet spot (or 40 for 48)
  • Discussed SANDAG testing progress
    • 100% full sample three zone example with SANDAG data running!
    • But runtimes slow, 30 hours
    • 10 hours for accessibility calculator
    • Could de-dup/cache to significantly speed it up
    • Accessibility calculator wasn't the focus of the 3 zone/TVPB work and so not surprised it needs performance tuning
    • We added a new timing-log.txt for submodel runtimes in develop that SANDAG can use
    • Trip destination running slow
    • SANDAG concerned with overall runtimes since current model (which is more complex) runs in 12 hours
    • SANDAG share their example so we can take a look and identify potential areas for speed up
    • SANDAG try 50 of 56 processors setup and MKL_NUM_THREADS=1
  • Another performance tuning task would be great
    • non-mandatory tour scheduling improvements, accessibility de-duping, trip purpose and destination speedups, etc
  • We may want to not improve the accessibility calculator much more and instead build the disaggregate accessibilities calculator that was discussed during the scoping phase
  • Discuss improved test system plan next time