Project Meeting 2021.05.04

Technical meeting

• Feel free to request an agenda item in the future as this is a forum for technical sharing
• Issues review
  • No new issues since last time but let's keep talking about performance needs by commenting on issue #377
  • Based on the recent chunking improvements work, Jeff believes even more so that migrating from strings to factors, especially for the tour and trip tables, should help with performance
  • Did we speed up sampling for one zone systems by skipping zones with size term == 0? We think so, but can check
• Thursday we'll discuss any questions related to the bench IRFP
• Update from Jeff on moving from chunking to available RAM settings, #406
  • Goal is to specify amount of RAM available and then stay within it
  • It is difficult though because of paging and virtual memory
  • If you watch memory usage in the Windows Resource Monitor, you see you often don't use all that you expect
  • So users may want to set higher than installed RAM
  • Three ways to check RAM usage - bytes allocated, RSS, and USS
  • bytes allocated reported by python but unreliable
  • RSS always too big since using disk resident memory
  • USS is amount if you killed the process now
  • And remember we need across processes as well for mp
  • Currently testing a hybrid bytes allocated + USS approach
  • Need to test 1 zone mtc and 3 zone sandag examples a number of times and each run takes 1 to 3 hours
  • The exploratory (dynamic) chunking takes 10 to 20% overhead so saving and re-using improves runtime 10 to 20%
  • User will want to specify machine RAM, run with exploratory chunking, see usage, and adjust settings, and then re-run with cached settings
  • It is working well and should have numbers to share next week
  • There's lots of new settings to help instrument it and we'll need to write some documentation on how to use it
  • Run times should improve as well
  • What if there just isn't enough RAM for the number of processors? Can we provide some info to the user about this?
  • What's a typical server config? Agencies to share machine specs
  • Typically the machine is dedicated to just running the model
  • As the model progresses through the submodels, more stuff is in RAM and therefore the headroom is reduced
  • The tour and trip tables for example include strings which are inefficient memory hogs and slow
  • Jeff plans to wrap up this week and then I'll test and document