Parking Lot for Experimental Results Ideas

[bushidocodes]

Current Target: now that you might be past the debugging phase. Is the focus on doing some aspect of the sensitivity analysis? A good first step (you'll recognize this from Nov) is comparing EDF vs. non. preemeptive vs. preemptive. I'd expect that the results will be different if we increase the concurrency and see latency/throughput

Concurrency Experiment run for the following permutations:

• EDF with preemption
• EDF without preemption
• FIFO with preemption
• FIFO without preemption

Additional Ideas:

• Jobs of varying distributions -- uniform (likely not interesting) to bimodal
• Jobs of varying execution time to emphasize system overheads (as they run more quickly)
• How does changing the deadline factor for this impact the system latency?
  • aside: We want the ability to set the deadline relative to the predicted execution time (e.g. a factor of it, 2x, 3x, etc...).
• Studying the above for the four scheduling policies.
• Determining how low we can make the quantum, and its relationship to throughput/latency.
• Understanding latency for tasks with real distributions (e.g. real applications with their runtime variations).
• Understanding runtime distributions for fib tasks so that we can understand how much variation is introduced by the system (vs. application tests where it could be introduced by the logic of the application).
• Memory stuff, that I'll leave for later discussion.

Needs clarification:

• We might want to send timer updates not to all cores, but to the specific one we want to enable preemption on. For now this is just a thought (that we will want to follow up on at some point), but I don't have an experiment in mind.

[emil916]

One additional idea:

• Building: multi-tenant infrastructure for tracking execution, and rate-limiting when necessary.