Explore a cutting-edge conference talk from FAST '24 that delves into Baleen, an innovative approach to machine learning admission and prefetching for flash caches in data center services. Learn how this system addresses the challenges of flash write endurance while optimizing backend load reduction. Discover the novel cache residency model called "episodes" and its role in guiding model training. Understand how Baleen focuses on the end-to-end system metric of Disk-head Time to more accurately measure backend load. Examine the impressive results from evaluations using Meta traces across seven storage clusters, showcasing a 12% reduction in Peak Disk-head Time compared to state-of-the-art policies. Gain insights into Baleen-TCO's ability to optimize flash write rates and reduce total cost of ownership by 17%. Access the provided code and traces to further explore this groundbreaking research in flash cache optimization for bulk storage systems.
Overview
Syllabus
FAST '24 - Baleen: ML Admission & Prefetching for Flash Caches
Taught by
USENIX