A Framework for Utilization-Based Absolute Delay Guarantees Using Adaptive Prefetching

University Of Virginia Main Campus, Charlottesville VA

Investigators

Abstract

We propose to develop an architecture and theoretical underpinnings for providing absolute delay guarantees for HTTP traffic. HTTP traffic constitutes an overwhelming majority of all Internet traffic today. Both network load and user-perceived end-to-end response time of web requests depend not only on network conditions but also on the performance of web proxy caches around major network backbones. For example, in an HTTP-dominated network, increasing the total amount of cached data may increase hit ratio and subsequently decrease both network load and network delay. Hence, an architecture for end-to-end web traffic delay guarantees should explicitly consider the effect of caching. It is the joint consideration of caching and network performance that separates our work from prior efforts on delay guarantees. Essentially, the joint problem considers data placement (replication) as a dimension to manipulate for affecting traffic delays. The approach is cost-effective since data storage is cheaper than network bandwidth. The first contribution of this project is to develop a scheme for network load control that relies on adaptive data prefetching. The architecture can be thought of as replacing admission control at the network boundary. In an HTTP context, while admission control would prevent a client's request from entering the network, data prefetching would bring the information to the client's side before it is requested, hence de ecting the request away from the backbone. While prefetching itself introduces traffic, the performance gain comes from the fact that, unlike serving live requests, prefetching can occur at a lower priority in the background without jeopardizing user-perceived network performance. Hence, prefetching removes time constraints from a big chunk of HTTP traffic which can now be served at a lower priority. Consequently, the remaining (live) HTTP traffic will receive better service from the network. The second main contribution of the project is a theoretical derivation of the relationship between network resource utilization and the satisfaction of end-to-end deadlines. Specifically, we prove that keeping network resource utilization due to live web traffic below a given threshold ensures that all ow deadlines are met. We call this threshold, the overcommitment threshold . This result allows us to associate deadlines with live web traffic and ensure their satisfaction simply by performing utilization control. The result obviates maintaining per- ow state in the network for the purposes of satisfying absolute delay guarantees. Merging the aforementioned contributions together, we propose to use our adaptive prefetching scheme to keep the amount of live web traffic below the overcommitment threshold. Hence, we ensure the satisfaction of absolute delay guarantees while requiring neither admission control on the network boundary nor per- ow state in routers. The architecture calls only for service differentiation in the network to separate live real-time web traffic from background prefetching traffic which may receive lower priority. The project will use the PI's existing evaluation testbed for implementing and evaluating architectural prototypes.