Reproduce “Affinity Tailor: Dynamic Locality-Aware Scheduling at Scale” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jin Xin Ng, Ori Livneh, Richard O'Grady, Josh Don, Peng Ding, Samuel Grossman, Luis Otero, Chris Kennelly and 2 more
Google's Affinity Tailor system introduces an OS scheduling architecture that improves application performance in multi-tenant datacenters by dynamically preserving microarchitectural locality. It achieves up to 11.7% per-CPU throughput gains by reducing cache misses and improving branch prediction, challenging the traditional work-conservation paradigm.
Key insights
- Empirical evidence demonstrates that prioritizing microarchitectural locality over strict work-conservation, even if it introduces transient localized queuing, yields substantial overall performance benefits on modern hardware.
- The coordination of userspace intelligence (demand prediction from the cluster agent) with kernel-enforced soft affinity policies effectively manages resource locality in dynamic, overcommitted datacenter environments.
- The negative impact of aggressive thread migration and cache pollution is particularly severe on chiplet-based architectures with split Last-Level Caches, highlighting the efficacy of topology-aware scheduling.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.