DAG-aware harmonizing job scheduling and data caching for disaggregated analytics frameworks

Abstract

Modern data analytics frameworks often integrate with external storage services, which can lead to storage bottlenecks. Existing caching and prefetching solutions utilize high-level information from data analytics frameworks to forecast future data accesses. They employ these predictions to prefetch data into the cache and manage the cache contents. However, this approach overlooks a fundamental opportunity: rather than caching data given a prediction of job execution, influencing the job execution order can enable more effective caching and prefetching. With this key insight, we introduce a novel system called Tripod, designed to synchronize job scheduling and data caching for analytics frameworks.Leveraging the higher-level information provided by analytics frameworks, Tripod explores the best-suited job execution order, guided by developed heuristics, to facilitate prefetching and caching. To fully exploit the potential of Tripod, we also introduce a novel caching strategy named CAP. This strategy not only acknowledges the job scheduling order but also offers fine-grained control over object prefetching and eviction. Our evaluation, conducted using standard analytic benchmarks (TPC-H and TPC-DS), demonstrates that Tripod achieves a speedup of up to 1.7x on state-of-the-art approaches. Moreover, when employing CAP to make caching decisions, the performance can further be improved (as much as 1.5x).