Scheduling I/O Requests with Deadlines: A Performance Evaluation

Abstract:

A real-time computing system allocates resources to tasks with the goal of meeting individual task deadlines. The CPU, main memory, I/O devices and access to shared data all should be managed with the goal of meeting task deadlines. This paper examines the I/O scheduling problem in detail. We present a detailed, realistic model for studying this problem in the context of a system which executes real-time transactions. The model takes advantage of the fact that reading from the disk occurs before a transaction commits while writing to the disk usually occurs after the transaction commits. We develop new algorithms that exploit this fact in order to meet the deadlines of individual requests. The algorithms are evaluated via detailed simulation and their performance is compared with traditional disk scheduling algorithms.