Write Caching in Distributed File Systems

Disk caches are employed in distributed le systems to avoid network accesses at clients and to compensate for the speed diierential between main memory and disk at le servers. Because of concerns about volatility, however, write requests have typically not beneetted from the presence of caches. Instead, they have been processed with some sort of write-through or periodic write-back approach to ensure the integrity of the stored data. The introduction of reasonably priced non-volatile (NV) memories has prompted interest in the use of such memory for write caching, at the server and/or at the client. This paper describes an investigation of several approaches to write caching in distributed systems with both volatile and non-volatile caches through trace-driven simulation experiments. The results support the ndings of earlier work that suggests important diierences between caching in the traditional single-level caching environment and caching in a two-level caching environment. While policies focusing on temporal locality perform well for a single-level caching system, or at the client of a two-level caching system, they may not be suitable for use at the server in a two-level caching system. This is because locality characteristics in the reference stream seen at the server in a two-level caching system may be destroyed by caching at the client with a NV write cache large enough to hold the client's working set of dirty blocks. Policies focusing on amortizing the cost of a disk seek operation over multiple write-back operations perform better at the server of a two-level caching system.