Selecting which ready source to service next is the core scheduling task in low-latency systems. The conventional approach routes that decision through queues; however, fan-in queues concentrate contention and throughput degrades under load, and priority usually compounds it - adding complexity. Polling avoids some contention but injects jitter and wasted cycles. This talk deep-dives into Signal Trees: an alternative lock-free, low-latency readiness scheduler that sidesteps queue-centered degradation at scale and eliminates the need for polling entirely.
We start with measurements - scaling, tail percentiles, and service distribution - showing up to 40× higher throughput than queue-based schedulers under contention, then explain how that performance improvement emerges from the design. We will examine the internal design, invariants, and the deliberate choices behind them. We cover how Signal Trees support priority and fairness without sacrificing performance or introducing the contention inherent in queue-based schedulers.
We close by delineating the boundary between the Signal Tree scheduling primitive and an execution layer above it: Signal Trees provide readiness and selection, deferring responsibility for work and data to the execution layer.
View Slides
Michael been building C++ systems professionally for over 25 years. He started in data compression, developing algorithms such as M99, M03, and MSufSort. For the past decade he has worked in finance as a principal engineer and software architect, building low-latency infrastructure for high-frequency trading. His interests include performance engineering, lock-free concurrency, and algorithm design.