Mean delay analysis of cache assisted file transfers over the Internet

We address the problem of analyzing the mean delay experienced by end-users in a Content Distribution Network (CDN) consisting of a set of surrogate (cache) servers connected via persistent TCP connections over the Internet cloud to a set of remotely located origin servers. We first consider the simplest scenario of a single cache server and a single origin server. Taking into account several factors such as cache hit probability, pending interest table, number of concurrent sessions, background load, bottleneck link speed etc., we derive an exact expression for the mean file transfer delay for a wide range of service disciplines at the origin server, under the restriction of equal rates of request arrivals for all the files. The derivation involves establishing the ergodicity of the underlying stochastic process, and is of interest on its own. The expression for the mean delay involves the second moment of the file sojourn time in the origin server. This expression enables us to compare, using an adaptation of certain standard sample path arguments, several service disciplines at the origin server in terms of their mean delay performance. Among a wide range of commonly used service disciplines, FCFS achieves the best mean delay performance. We discuss an implication of this result for system design. Finally, we provide an approximate mean delay analysis for the general scenario of multiple caches and origin servers, which is accurate when the number of files is large compared to the number of requests in service at any point in time.