CIRCA

A research prototype for strict placement-level OOD co-location scheduling with counterfactual scoring and audited decision evaluation.

CIRCA

CIRCA is a research prototype for strict placement-level OOD co-location scheduling in multi-tenant GPU systems. It studies how to score candidate placements under genuinely unseen deployment contexts, and how to evaluate whether a scheduler can still rank actions correctly when the placement structure was never observed during training.

What It Does

  • Scores candidate placements with a counterfactual decision-oriented view under strict OOD settings
  • Provides a reproducible pipeline for benchmark collection, training, evaluation, and audit
  • Includes audited main-result artifacts and representative failure-case extraction
  • Focuses on decision quality for co-location scheduling rather than only fitting historical interference patterns

Why It Is Interesting

In practical GPU scheduling, the key question is not only whether a predictor can output plausible interference scores, but whether it can make the right placement decision when the runtime context is genuinely new. CIRCA is interesting because it emphasizes decision evaluation under unseen placements, which is closer to the real deployment challenge than standard in-distribution scoring.

Repository Highlights

  • Core package in circa/ for data handling, benchmark utilities, and predictors
  • Runnable scripts in scripts/ for collection, training, evaluation, audit, and analysis
  • Locked benchmark and audited result artifacts for reproducible evaluation
  • Public research release under the MIT License