Spots: Exploring expiration time-based distributed caching System on Flash

Abstract

As the demand for high bandwidth applications and the amount of data stored increases, new approaches to storage devices and cache algorithms are necessary. Development in flash-based solid-state drives (SSDs) has lowered their cost per capacity significantly and improved their read/write speeds. Unfortunately, flash-based SSDs have a limited lifetime, as erase/write cycles wear down the cells of the device, eventually rendering it useless. This can be countered by designing cache systems that are aware of these physical limitations, such as being conscious of how objects are grouped on the device. By labeling objects based on an expiration time, we can group objects in a manner that reduces the number of erase/write cycles for a better flash-based cache. This paper discusses transitioning Leopard, a cache system that is based on expi- ration time to improve cache performance on flash, from a single SSD to a distributed caching system. It lays down the initial considerations and framework for Spots - a version of Leopard controlling multiple SSDs to form a distributed cache. We ex- plore in-depth how the grouping of objects changes when there are more zones to place them in, and reconsider core parts of Leopard that may not transition well to a multi-device cache. The findings are that there is generally little improvement from having increased flexibility in object placement. Modifying aspects of the grouping policy has shown some preliminary results that improve the byte miss ratio/write volume Pareto frontier.