A mathematical hard disk timing model for full system simulation

Abstract

This paper introduces and validates a mathematical hard disk timing model designed for use in an execution-driven full-system simulator. While very accurate disk simulators and emulators exist, their complexity is often not warranted when disk operations are not the sole focus of an experiment. This model depends far less on the details of the disk structure or the physical layout of the disk, making it less complex and easier to configure. By combining traditional hard disk modeling methods with novel ways of mathematically and probabilistically modeling the platter layout and reorder queues, the need for a physical model of the disk platter is eliminated. Current full-system simulators do not include a realistic disk timing model, giving them large and variable errors when modeling disk intensive workloads. With this model, we show that disk intensive full-system simulation can be accurate, with benchmarks averaging 12% error for individual disks and 18% error for disk arrays. This model will benefit full-system performance simulations by bridging the gap between complex and highly specific disk simulators and the actual performance modeling requirements needed for effective use of full-system simulators.