Friday, February 04, 2011
2:00 PM
Computer Studies Bldg. Room 703
Stan Park
University of Rochester
Flash I/O Fairness for Concurrent Read/Write Workloads
Flash-based storage devices have the potential to eliminate the I/O bottlenecks in data-intensive applications. However, the large performance discrepancy between Flash reads and writes introduces challenges for fairness and even presents vulnerability to denial-of-service attacks. Evaluations using mixed read/write workloads on several Flash-based storage devices (including a CompactFlash card in a low-power wimpy node) show that read tasks may experience slowdowns at one or two orders of magnitude more than write-intensive tasks do. Evaluations also find that fair queueing-based I/O scheduling realizes resource utilization fairness at a substantial performance degradation (up to two orders of magnitude) due to the suppression of device read parallelism and improper management of I/O anticipation. This paper demonstrates that it is possible to attain high efficiency while achieving fairness in Flash I/O scheduling. Our scheduler employs a fair I/O timeslice management with mechanisms for read preference, read parallelism, and fairness-oriented I/O anticipation.