A Cache Management Algorithm for the Heterogeneous Storage Systems

The scale of storage system is becoming larger with the rapid increase of the amount of produced data. Along with the development of computer technologies, such as cloud computing, cloud storage and big data, higher requirements are put forward to storage systems: higher capacity, higher performance and higher reliability. In order to satisfy the increasing requirement of capacity and performance, modern data center widely adopts multi technologies to implement the dynamic increasing of storage and performance, such as virtualization, hybrid storage and so on, which makes the storage systems trend more and more heterogeneous. The heterogeneous storage system introduces multiple new problems, of which one key problem is the degradation of performance as load unbalance. Thats because the difference of capacity and performance between heterogeneous storage devices make the parallelism technologies hardly to obtain high performance, such as RAID, Erasure code. For this problem, we propose a caching algorithm based on performance prediction and identification of workload characteristic, named Caper (access-pattern aware and performance prediction-based cache allocation algorithm). The main idea of Caper algorithm is to allocate the load according to the capacity of the storage devices, which aims to alleviate the load unbalance or eliminate the performance bottleneck in the heterogeneous storage systems. The Caper algorithm is composed of three parts: prediction of performance for I/O request, analysis of caching requirement for storage device, and caching replacement policy. The algorithm also classifies the application workload into three types: random access, sequential access, and looping access. In order to ensure high caching utility, the algorithm adjusts the size of logic cache partition based on the analysis of the caching requirement. Besides, in order to adapt to the heterogeneous storage system, the Caper algorithm improves the Clock cache replacement algorithm. The experimental results also show that the Caper algorithm can significantly improve the performance about 26.1% compared with Charkraborty algorithm under mixed workloads, 28.1% compared with Forney algorithm, and 30.3% compared with Clock algorithm. Even adding prefetching operation, Caper algorithm also improves the performance about 7.7% and 17.4% compared with Charkraborty algorithm and Forney algorithm respectively.