Management policies for non-volatile write caches

Abstract

Many computer hardware and software architectures buffer data in memory to improve system performance. Volatile disk or file caches are sometimes used to delay the propagation of writes to disk (called delayed writes). While delayed writes improve system performance, volatile caches can cause the loss of vital data during sudden failure. In this study, we investigate managing non-volatile RAM (NVRAM) caches with different simple strategies to delay writes to disk. We evaluate the performance of NVRAM caches using three measures of merit: the number of stalled writes which wait while the cache is cleaned before being serviced the mean service time far I/O requests, and the number of writes generated by cleaning the cache. Our results show that even small non-volatile write caches using simple management policies can reduce the number of writes to disk by at least 70% and as much as 80% in some cases. Our results also show that the number of stalled writes is high: 30% at best and nearly 100% at worst. Adding pro-active purging effectively decreases both stalled writes and disk write activity.