https://joulie-k8s.github.io/Joulie/versions/v0.0.2/docs/experiments/Recent content in Experiments on JoulieHugoen-ushttps://joulie-k8s.github.io/Joulie/versions/v0.0.2/docs/experiments/kwok-benchmark/Mon, 01 Jan 0001 00:00:00 +0000https://joulie-k8s.github.io/Joulie/versions/v0.0.2/docs/experiments/kwok-benchmark/<p>This document describes the first benchmark harness implementation under:</p> <ul> <li><code>experiments/01-kwok-benchmark/</code></li> </ul> <h2 id="motivation">Motivation</h2> <p>The benchmark focuses on repeatable comparisons of scheduler+control behavior across baselines:</p> <ul> <li>baseline A: no Joulie control path (simulator only),</li> <li>baseline B: static partition style,</li> <li>baseline C: queue-aware style.</li> </ul> <p>The setup keeps real scheduling semantics while using fake KWOK nodes and simulator-driven telemetry/control effects.</p> <h2 id="assumptions">Assumptions</h2> <ul> <li>Kubernetes scheduler and API are real.</li> <li>Fake nodes are tainted and selected by workload pods.</li> <li>Simulator injects and advances batch work from trace input.</li> <li>Experiment scripts orchestrate install/run/collect/plot.</li> <li>A/B/C workload fairness is preserved by generating a canonical per-seed trace and deriving baseline A by stripping only power-profile affinity.</li> </ul> <h2 id="what-is-measured-now">What is measured now</h2> <p>Current harness outputs:</p>