ALPS: An Adaptive Learning, Priority OS Scheduler for Serverless Functions

Explore a groundbreaking conference talk on ALPS (Adaptive Learning, Priority Scheduler), a novel application-aware kernel scheduler designed for serverless functions. Delve into the unique challenges posed by Function-as-a-Service (FaaS) workloads and their impact on kernel scheduling. Examine the limitations of Linux CFS (Completely Fair Scheduler) and the trade-offs of preemptive shortest job first scheduling. Discover ALPS' innovative decoupled scheduler frontend and backend architecture, which combines approximate SRPT and proportional-share scheduling. Learn how ALPS leverages adaptive learning and eBPF functions to enhance workload intelligence in OS scheduling. Gain insights into the extensive evaluation of ALPS using production FaaS workloads from Huawei and Azure, showcasing its significant performance improvements over CFS.

USENIX ATC '24 - ALPS: An Adaptive Learning, Priority OS Scheduler for Serverless Functions