We have developed mechanisms for multiplexing disk bandwidth, storage space, and file server cache among heterogeneous applications. The most noteworthy of these mechanisms is the one for managing disk bandwidth.

Conventional disk scheduling algorithms are optimized either for best-effort requests or for requests with real-time deadlines. A disk scheduling algorithm that simultaneously supports applications with heterogeneous performance requirements should: (1) align the service it provides with the application needs, (2) protect application classes from one another, (3) be work-conserving and adapt to changes in work-load, (4) minimize the seek time and rotational latency overhead incurred during access, and (5) be computationally efficient. We have developed the Cello disk scheduling algorithm that addresses these requirements. Cello achieves this by allocating disk bandwidth to application classes at two time-scales. At a coarse time-scale, it determines the number of requests from each application class to be serviced, and at a fine time-scale, it determines the order for servicing the set of selected requests. Whereas the former enables Cello to protect application classes from one another and to adapt disk bandwidth allocation with changing work-load, the latter enables it to align the service provided with the application requirements while minimizing the seek time and rotational latency overhead. These two tasks naturally map to an extensible two-level disk scheduling architecture consisting of a application-independent and a set of application-specific schedulers.

Representative Publications:

1. P. Shenoy and H.M. Vin, Cello: A Disk Scheduling Framework for Next-generation Operating Systems, In Proceedings of ACM SIGMETRICS'98, the International Conference on Measurement and Modeling of Computer Systems, June 1998 [ Abstract | Paper ]