Modeling and Taming Parallel TCP on the Wide Area Network

Parallel TCP flows are broadly used in the high performance distributed computing community to enhance network throughput, particularly for large data transfers. Previous research has studied the mechanism by which parallel TCP improves aggregate throughput and proposed a model to determine its upper bound when the network is not congested. In this work, we address how to predict parallel TCP throughput as a function of the number of flows without such constraints, as well as how to predict the corresponding impact on cross traffic. This combination allows us to answer the following question on behalf of a user: what number of parallel flows will give the highest throughput with less than a p% impact on cross traffic? We term this the maximum nondisruptive throughput. We begin by studying the behavior of parallel TCP in simulation to help derive a model for predicting parallel TCP throughput and its impact on cross traffic. Combining this model with some previous findings we derive a simple, yet effective, online advisor. We evaluate our advisor through simulation-based and wide-area experimentation