Measuring and evaluating TCP splitting for cloud services

Abstract

In this paper, we examine the benefits of split-TCP proxies, deployed in an operational world-wide network, for accelerating cloud services. We consider a fraction of a network consisting of a large number of satellite datacenters, which host split-TCP proxies, and a smaller number of mega datacenters, which ultimately perform computation or provide storage. Using web search as an exemplary case study, our detailed measurements reveal that a vanilla TCP splitting solution deployed at the satellite DCs reduces the 95th percentile of latency by as much as 43% when compared to serving queries directly from the mega DCs. Through careful dissection of the measurement results, we characterize how individual components, including proxy stacks, network protocols, packet losses and network load, can impact the latency. Finally, we shed light on further optimizations that can fully realize the potential of the TCP splitting solution.