Scheduling Performance Working Group - April 28th, 2023

[mattdailis]

• Participants: @mattdailis @adrienmaillard @bashbash @JoelCourtney @Mythicaeda @camargo
• Scribe: @mattdailis
• Agenda:
  • Overview of current scheduler flow to get on the same page
  • Highlight all places where simulation occurs in the above flow
  • Discuss current and future work in context of which simulation points it reduces

Discussion notes

All bold text: this is where the scheduler triggers simulation
Highlighted in gray: this is what #708 allows users to toggle on and off per goal
Highlighted in yellow: the "hot spot" where a potentially very large number of simulations occur

• Bulk insert?
  • Any correctness risks?
  • If one of the activities “doesn’t work”, can we pull that activity without reprocessing the other ones?
    • Partial satisfiability
  • Possibility: allow scheduler to produce some invalid placements (but maybe call them out explicitly)
• Measure performance benefit of simulate after
• Consider option to simulate once between

Adrien: What is the role of the scheduler? What is the role of the user?
Either way user should provide guidance

• Option 1: annotations from the mission model
• Option 2: let the user add more constraints to help the scheduler find the right placement
  • May help disable simulation, if the user has confidence in the constraints they provided
  • Narrow search space as much as possible

The above would support disabling simulation in return for placing constraints that help narrow the scheduler's search space

• Max Duration or Estimated Duration from mission model
  • Results in idle time - can be problematic for Clipper, but not for landers

Summary

We discovered a few misconceptions we had about how the scheduler triggers simulation

• When calling simulateActivity, the scheduler will simulate until all activities in the plan complete
• createAndInsertActs is called on one activity at a time, inside the "for each conflict" loop

We discussed the potential to move createAndInsertActs to the outer loop. We're not sure about the correctness implications of doing so. This seems related to the "partial satisfiability" of the goal, and hinges on whether future activities depend on past activities. One option would be to have the scheduler still place an activity, even though it knows it causes a conflict, and highlight the problem to the user.

Next steps

• Schedule follow-up working group (@mattdailis)
• Continue work on removing extra simulation when uncontrollable search guesses right (@adrienmaillard)
• Investigate utility of NonDecreasing annotation (@JoelCourtney)

Unassigned possible future tasks:

• Measure speedup from toggling simulation (unassigned)
• Measure speedup from FixedDuration feature (unassigned)
• Investigate correctness implications of moving createAndInsertActs into the outer loop (unassigned)