Supporting I/O-intensive applications in a high performance storage system
Sangyup Shim · 1998
I/O intensive applications require a high performance and fault-tolerant storage system. In this thesis, we study how to support I/O intensive applications efficiently and integrate the latest storage technologies closely with applications. We consider two important I/O intensive applications which have quite different charteristics. They are Video-On-Demand (VOD) and databases. In VOD, popular movies may get large number of requests. One way to satisfy all the requests with limited resources is to accumulate requests over a predetermined interval and serve them together by a single stream. I investigate ways to place popular videos on disks and take advantage of zone bit recording of high performance disks. New high-speed interconnection standards are emerged to provide high performance as well as fault tolerance. Fibre Channel-Arbitrated Loop (FC-AL) and Serial Storage Architecture (SSA) are the two such standards for serial storage interfaces which can connect a large number of disks to servers. The serial interfaces offer special features such as a fairness algorithm, spatial reuse, and multiple host attachments. The fairness algorithm ensures that a fraction of data bandwidth is allocated to each device. Hence a storage system based on the serial interfaces is a better alternative in supporting continuous media. In this thesis, we discuss the effects of storage technology in terms of the performance of VOD by comparing the performance of SSA and SCSI in supporting continuous media. In databases, the performance of small writes are important because of the extensive read caching and a write-through policy. The performance of small writes can be improved using disk based writes. A disk based write is a RAID implementation based on an exclusive-or (XOR) unit in each disk. It improves the response times for small writes and reduces link traffic. An XOR unit on a disk adds relatively little cost to a high performance disk. The performance of disk based writes and implementation issues are addressed in this thesis. The implementation issues addressed include how to resolve a state of deadlock, how to enhance cache replacement policy, how to increase concurrency, and how to minimize interactions of concurrent commands.