2C: A New Approach to Design the Second Level Buffer Cache

Multi-level cache hierarchies are very common in today's networked storage environments. Dedicated data servers or file servers are simultaneously accessed by multi-applications of multi-clients. Workload shift results in cache pollution and degrade of hit ratio. Widely used cache policies work well on the case when a single level cache separates the data provider and data consumer, but poorly on the second level buffer cache scenarios. To avoid the threshing induced by improper cache policies, Lustre file system even abandons the idea of employing server-side buffer caches. However, it suffers from poor performance on random access. We propose a novel cache replacement policy 2C, namely the 2nd level Cache replacement policy, which provides fast access to random pages without extra degrade of hit ratio. 2C detects sequential pages and reclaims memory for random access, so hit ratio of random access is increased significantly. Because 2C reduces the overhead of seeking and rotating time mainly induced by disk head movements of random access, it improves the effective access time of overall storage systems. Our simulation results show that 2C performs better than LRU for a wide range of cache sizes and random/sequential workloads.