Modernizing the Meta Ads Service With an Open-Source Kernel Scheduler

This post details the implementation and impact of a custom BPF-based scheduling policy using sched_ext to address latency regressions introduced by a Linux kernel upgrade (v6.6 EEVDF scheduler). The custom policy soft-partitions CPUs into latency-critical and less sensitive pools, improving L3 cache locality and reducing DRAM access. This resulted in a 28% reduction in ads retrieval tail latency, 3.28 MW power savings, and a 1.1% increase in weighted-ads-ranked. Subsequent user-space policy updates further reduced p99 latency by 60% and timeout errors by 18%. The post highlights sched_ext as a strategic asset for independent, rapid scheduler optimization.