University of Texas at Austin
About the talk:
|Study after study shows that prefetching can significantly
improve network services' response times and availability. Yet few network
services prefetch. Why?
In this talk, we argue that threshold prefetching -- the dominant strategy for balancing prefetching's latency and availability benefits against its bandwidth and server-load costs -- is fundamentally flawed. In particular, prefetching thresholds represent magic numbers that are difficult to tune due to their tenuous relationship to the actual goals of the system and due to the changing benefit/cost trade-offs over different time scales as technology trends alter resource costs over months and years and as workload variations affect the amount of spare capacity over seconds, minutes, and hours. Furthermore, the consequences of incorrectly setting this threshold parameter can be dire because systems often exhibit a highly non-linear relationship between prefetching load and the resulting interference with demand requests that is not exposed by threshold-based schemes.
We propose a new prefetching architecture that separates the task of ranking objects for prefetching from the task of scheduling prefetch requests. We have applied this approach to two prototype replication systems: one with client-initiated prefetching and the other with server-driven speculative pushes of content. Our experience suggests that the separation of concerns at the heart of this architecture has three advantages -- (i) it simplifies the design and deployment of prefetching systems by eliminating the need to choose appropriate thresholds for an environment and update these thresholds with changing conditions, (ii) it reduces the risk of prefetching by avoiding interference even during unanticipated periods of high load, and (iii) it increases the benefits of prefetching by prefetching more aggressively than would otherwise be safe during periods of low or moderate load.
About the speaker:
|Mike Dahlin is an Associate Professor in the Department of Computer Sciences at the University of Texas at Austin. His work focuses on large scale distributed systems. Dr. Dahlin received his PhD from the University of California at Berkeley in 1996, the NSF CAREER award in 1998, and the Sloan Research Fellowship in 2000.|