Elijah's Notes

Scheduling Benchmarks and Simulators

1 min read

Purpose

These are not microbenchmarks against kernel schedulers. They are small simulators for checking the math and building intuition:

  • queue growth versus utilization
  • mean and tail latency under different scheduling policies
  • per-flow fairness under packet schedulers
  • load balance and steal counts under task schedulers
  • TTFT/throughput tradeoffs under simple LLM-serving schedulers

The point is to preserve the pathways for grokking the models, not to pretend these scripts reproduce production systems exactly.

Files

  • queue_sim.py: M/M/1-style queue simulator with average and percentile latency
  • cpu_policies.py: FIFO, SJF, SRPT, and RR on the same job set
  • packet_fairness.py: FIFO, RR, and DRR over multiple packet flows
  • work_stealing.py: toy task scheduler with local queues and steals
  • llm_batching.py: simple request scheduler with KV-budget and batch-token constraints

Reproduction

python3 content/systems/scheduling/benchmarks/queue_sim.py --rho 0.8 --jobs 20000
python3 content/systems/scheduling/benchmarks/cpu_policies.py --jobs 40 --seed 0
python3 content/systems/scheduling/benchmarks/packet_fairness.py --rounds 200
python3 content/systems/scheduling/benchmarks/work_stealing.py --workers 8 --tasks 2000 --seed 0
python3 content/systems/scheduling/benchmarks/llm_batching.py --requests 256 --seed 0

Graph View