SCION: Lightweight Cache Policy Orchestration for Nonstationary Workloads

SCION is an efficient framework for orchestrating policies in object caches, tailored for diverse, dynamic, and throughput-limited production environments. It identifies the optimal cache policies from a limited selection by utilizing a minimal workload fingerprint derived from the critical path. The prototype, AUTO, leverages brief prefix statistics regarding object size, cacheability, reuse, and cache dimensions, employing an offline-trained linear selector to differentiate among GDSF, S3-FIFO, SIEVE, LHD, W-TinyLFU-AV, and DynamicAdaptiveClimb. A more straightforward version, SCION-P90, relies solely on a p90 threshold. In a CPU-only, trace-driven assessment involving 30 public object-cache traces and a distinct HR-Cache simulator subset, AUTO enhances miss ratios for cacheable objects. This research confronts the challenge posed by straightforward non-ML policies like SIEVE and S3-FIFO, which establish a formidable baseline, necessitating that any learned approach be aware of overhead, resilient to drift, and competitive with proficient experts.