Proc. SAICSIT ’98 (poster session), Gordon’s Bay, South Africa, November 1998, pp 142-144

Technical Report TR-Wits-CS-1998-5

Latency goals often relate to response times seen by users, which are slow by computer standards, but scaling up to large numbers of users presents a problem. Examples include transaction-based systems and web sites. While a transaction-based system presents performance challenges other than disk latency, it is interesting to develop a model of disk architecture in which disk latency no longer presents a challenge, which allows sytem designers to focus on other areas in which performance goals may be hard to meet. The disk delay line concept relies on the fact that a disk can stream data quickly. A single-disk delay line is a disk which constantly streams its entire contents, and a request for data or for a write waits until the required portion of the data stream appears (a strategy similar to ultrasonic delay lines used in early computer systems). A given latency goal can be achieved by replicating disks, with copies of streams evenly spaced apart in time, and a given number of transactions per second can be supported by sufficient memory to buffer requests.

(PDF 9.5KB)
A shorter version of this paper has been submitted elsewhere (not accepted for presentation but published on the conference session web site).