Analyzing Storage System Workloads

Disk storage subsystems have not kept up the speed with processors. Processor performance has been increasing at a much higher rate than that of disk drives. Therefore, I/O subsystem has become a bottleneck in current computer systems. With this in mind the research community is looking into ways of improving the I/O subsystem. IBM and HP are among the organizations doing research and development of high performance storage systems also referred to as Enterprise Storage Systems (ESSs). Much of this effort goes into the evaluation of these systems for correctness and performance. For these evaluations, using simulations or otherwise, to be reliable, there is need to correctly understand and model the disk I/O workloads also known as I/O traffic or physical I/O workloads. Therefore in our work, we analyzed I/O workload traces to derive statistics which can be used as a guide in the (a) modelling of I/O workload and subsequent production of I/O workload representative of the actual I/O workload for evaluation and (b) optimization of the storage systems. Our results show that the distribution of inter-arrival times of I/O requests are heavy-tailed, and that the I/O request sizes are a function of the operating system