Performance study of parallel I/O systems
Varman, Peter J.
Master of Science
The use of parallelism in I/O systems is becoming increasingly important as the performance gap between processors and disks continues to widen. This thesis studies the performance of multiple external merge sorts in an I/O system with multiple disks. Specifically, we investigate the impact of data placement on I/O performance. For one intuitively good placement policy, a race develops among concurrent merge sorts, resulting in the serialization of job executions and significant performance degradation. We present a model of a system with two jobs performing concurrent I/O and analyze the model. Our analysis accurately predicts the development of the race condition. We also present methods to control the race based on data placement and disk scheduling policies, which are shown to be effective through simulations.
Electronics; Electrical engineering