TCP-PARIS: a parallel download protocol for replicas

Abstract

Parallel download protocols have the potential to reduce file download time and to achieve a server-side load balancing in replica systems, such as peer-to-peer networks, content distribution networks and mirrored servers, by simultaneously establishing connections to multiple replicas and downloading disjoint file parts in parallel. This paper presents TCP-PARIS, a novel parallel download protocol from multiple replicas to one receiver. Because the ideal partitioning of the transfer volume from each server is a dynamic and a difficult-to-predict function of network conditions, server load and data size, TCP-PARIS uses the stream segmentation of TCP and congestion window information to continuously adapt the assigned volume to each server in proportion to the bandwidth-delay product to best approximate the optimal data partitioning. Analytical results, simulation and Internet experiments with a transport-layer implementation characterize the performance and the resource requirements of TCP-PARIS and allow a comparison with related protocols. Extensive simulations with varying network and application parameters show download time reductions of up to 52% compared to single-flow downloads and up to 52% compared to related protocols.