|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. To the best of our knowledge, we are the first to answer the following question on behalf of a user: what number of parallel flows will give the highest throughput with less than a 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.
The following paper illustartes our approach.
Our current implementation of the TameParallelTCP() call is a Perl script (that makes use of iperf). A binary version will soon be available in the next distribution.
Download version 1.0.
Effort sponsored by the National Science Foundation under Grants ANI-0093221, ACI-0112891, ANI-0301108, EIA-0130869, and EIA-0224449. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author and do not necessarily reflect the views of the National Science Foundation (NSF).