Investigation of a prefetch model for low bandwidth networks

Abstract

We investigate speculative prefetching under a model in which prefetching is neither aborted nor preempted by demand fetch but instead gets equal priority in network bandwidth utilisation. We argue that the non-abortive assumption is appropriate for wireless networks where bandwidth is low and latency is high, and the non-preemptive assumption is appropriate for Internet where prioritization is not always possible. This paper assumes the existence of an access model to provide some knowledge about future accesses and investigates analytically the performance of a prefetcher that utilises this knowledge. In mobile computing, because resources are severely constrained, performance prediction is as important as access prediction. For uniform retrieval time, we derive a theoretical limit of improvement in access time due to prefetching. This leads to the formulation of an optimal algorithrn for prefetching one access ahead. For non-uniform retrieval time, two different types of prefetching of multiple documents, namely mainline and branch prefetch, are evaluated against prefetch of single document. In mainline prefetch, the most probable sequence of future accesses is prefetched. In branch prefetch, a set of different alternatives for future accesses is prefetched. Under some conditions, mainline prefetch may give slight improvement in user-perceived access time over single prefetch with nominal extra retrieval cost, where retrieval cost is defined as the expected network time wasted in non-useful prefetch. Branch prefetch performs better than mainline prefetch but incurs more retrieval cost.