Optimising the Go runtime scheduler

Abstract

The current Go scheduling runtime is based upon work-stealing, a widely-used approach in scheduling. This work looks into the basic components that constitute a scheduler, reviews recent progress made in scheduler development, and analyses an alternate form of work scheduling, namely in the form of batch-threading, by re-engineering the Go runtime scheduler. The focus is set on observing runtime performance, primarily using synthetic benchmarks so as to contrast with previous work that has been done on batching schedulers. This is complemented with a comparison between the work-stealing implementation and our batching implementation using the standard Go language benchmark suite. The results are aimed to show how a thread batching scheduling is a viable solution in achieving better performance, even achieving up to 20% improvement through simple modifications of the current Go scheduler.