Saber: Sequential Access Based cachE Replacement to Reduce the Cache Miss Penalty

Ever since a long time ago, recency and frequency are not to be sneezed at design of cache management policy. Meanwhile, hit ratio is used to estimate whether a cache replacement policy is good or not. The influence of miss penalty on cache performance is neglected incorrectly. In fact, the cache miss penalty can vary by almost one or two orders of magnitude among different policies because of tremendous divergence between sequential access rate and random access rate of disk. We notice the phenomena and propose a novel cache management policy, namely, sequential access based cache replacement, or Saber for short, where the non-uniform weights of sequentiality are dynamically assigned to buffered pages depending on page locations in the disk to determine the candidate victim. In the case, isolated pages will stay longer than neighboring ones in the memory. Hence, hard disks will work under sequential access mode as much as possible, which can improve cache performance by reducing the cache miss penalty and the mean access latency.