Performance analysis of the disk subsystem

An important performance measure in client/server systems is the response time. Response time is generally limited by how fast data is retrieved from the underlying storage system. The storage system performance is especially limiting in I/O intensive applications, where the response time is of critical concern. Therefore, there is a need to model the performance of disk subsystem in the early stages of its design as well as during its execution, so that a better configuration is achieved. This research analyzes the performance of disk storage subsystems by obtaining metrics such as response time and throughput. The mechanics of a single disk is studied in detail. Analytical service time moment equations are derived, and the results are compared against the Disksim*. As far as we know, the analytical approach presented is the only one close enough to represent Disksim results. The analytical model used for the single disk is further extended to be used in RAID I system time approximations. In a typical client/server environment, arrivals to the disk subsystem are departures from the server CPU. Therefore, the Markovian arrival assumption to the disk subsystem, as was made in the early work of [Xiong06] needs to be improved by a more general arrival process. Here, this is done by expressing the interarrival times observed between two disk arrivals mathematically, which then enables us to compute its first and second moments. A phase-type distribution is used to model this general arrival process and to carry out a queueing analysis to compute a set of performance measures. GI/G/1 queueing analysis of the system is done, and performance metrics such as the average waiting time at the disk queue, as well as the average number of requests waiting are found. The results are compared against the results our simulator, which is fine-tuned and statistically proven to represent the Disksim behavior. It is shown that the accuracy of the proposed approximation is highly acceptable. Finally, different request scheduling algorithms are analyzed and compared with the one we have proposed. *Disksim is a simulator developed at the University of Michigan and enhanced at the CMU.