A study of divided disk cache performance using DiskSim

Abstract

The performance gap between the processor and the memory speeds has been increasing. Whereas processor speeds have improved by 60% annually, secondary storage speed has only improved by 10% annually. Although there have been many efficient techniques studied to minimize the speed gap, it still remains a bottleneck in various commercial implementations. Since secondary memory technologies are much slower than the main memory, it is challenging to match its speed to that of the processor. Usually, hard disk drives include semiconductor disk-caches to improve their performance. A hit in the disk-cache eliminates the mechanical seek time and rotational latency. To further improve performance a divided disk-cache, which is subdivided between metadata and data, was proposed earlier. In this paper, we evaluate the performance of such a disk-cache via extensive simulations on DiskSim. The simulations were run on standard benchmarks and synthetic benchmarks using normal, Gaussian and Poisson distributions. The divided point between data and metadata regions was varied to find an optimum divided point. The simulation runs for the standard and synthetic benchmarks show improvement up to 6% in hit ratio when the metadata region in the disk cache is between 10–30%. Since metadata is smaller in size, but accessed frequently, such a result seems reasonable. Although the performance improvement is small, it is important, because of the high access latency at the level of the disk. As such, our raw projections show that the overall response time improvement is expected to be between 15–20%. Such performance improvement could particularly be significant for long running processes.