Highly concurrent shared storage - Carnegie Mellon University.pdfVIP

Appears in the Proceedings of the International Conference On Distributed Computing Systems, Taipei, April 2000. Highly concurrent shared storage Khalil Amiri, Garth A. Gibson Richard Golding Carnegie Mellon University, Pittsburgh, PA Hewlett-Packard Laboratories, Palo Alto, CA {amiri+,garth+}@ golding@ Abstract1 One of the major limitations of today’s I/O subsystems is their limited scalability caused by shared controllers that Switched system-area networks enable thousands of storage data must pass through, typically from server to RAID devices to be shared and directly accessed by end hosts, controller, and from RAID controller to device. Emerging promising databases and filesystems highly scalable, reli- shared, network-attached storage arrays, like the one shown able storage. In such systems, hosts perform access tasks in Figure 1(a), enhance scalability by eliminating the shared (read and write) and management tasks (storage migration controllers and enable direct host access to potentially and reconstruction of data on failed devices.) Each task thousands of storage devices [9, 18] over cost-effective translates into multiple phases of low-level device I/Os, so switched networks [4, 15]. In these systems, each host acts that concurrent host tasks accessing shared devices can as the storage controller on behalf of the applications corrupt redundancy codes and cause hosts to read inconsis- running on it, achieving scalable storage access bandwidths tent data. Concurre