Web Caching Architectures: Hierarchical and Distributed Caching

Web caches are installed to reduce the receivers’ latency and save network bandwidth. These caches need to cooperate in order to improve their performance. A popular technique to make caches cooperate is by setting up a caching hierarchy. However, there are several problems associated with a caching hierarchy: i) every hierarchy level introduces additional delay, ii) redundant document copies are stored at every hierarchy level, and iii) higher level caches tend to become bottlenecks. Recently, a number of researchers have proposed the setup of a totally distributed caching scheme where only caches at the bottom level of the network cooperate and there are no intermediate caches. In this paper we analyze the performance of both hierarchical and distributed caching. Our main performance measure is the expected latency to retrieve a Web document. We find that hierarchical caching gives shorter connection times than distributed caching. We also find that distributed caching gives shorter transmission times than hierarchical caching since the traffic flows through less congested traffic links. Distributed caching has higher bandwidth usage than hierarchical caching. However, the network traffic generated by a distributed scheme is better distributed, using more bandwidth in the lower network levels, which are less congested. We also discuss administrative issues concerning the deployment of distributed caching. Additionally, we study a hybrid scheme where there is a caching hierarchy and a certain number of caches that cooperate at every caching level using a distributed caching scheme. We find that a “well configured” hybrid scheme can combine the advantages of both hierarchical and distributed caching, reducing the connection time as well as the transmission time. Depending on the hybrid caching architecture, the current parent caches load, and the document size, there is certain number of caches that should cooperate at each network level to minimize the overall retrieval latency. We propose small modifications of the existing cache-sharing protocols to dynamically determine the degree of cooperation between caches at every level of a hybrid caching scheme.