Flash-Conscious Cache Population for Enterprise Database Workloads.

Abstract

Host-side flash caching has lately emerged as a suitable and e↵ective means of accelerating enterprise workloads. How-ever, cache management for flash-based caching is di↵erent from traditional DRAM-based caching. A flash cache sits underneath the DRAM cache. Its position in the hierarchy combined with the unique characteristics of flash, calls for a di↵erent cache management solution. Specifically, cache population, an aspect of cache management which is not at-tributed much importance in DRAM caches, becomes cru-cial in flash-based caches. In this paper, we first present a performance evaluation of three popular open-source flash cache implementations: flashcache, bcache, and EnhanceIO. We evaluate them un-der an industry-standard database benchmark and identify their limitations. We demonstrate that several shortcom-ings are due to sub-optimal cache population. We propose a novel set of techniques for cache population, and present the design of the Scalable Cache Engine (SCE) – a new flash cache solution that incorporates our cache population tech-niques. We demonstrate that SCE remarkably outperforms the existing open-source solutions: 45 % higher throughput, 55 % lower latency, 12 ⇥ faster cache warm-up than flash-cache, and 95 % less memory usage than EnhanceIO. 1.