Balancing Accuracy and Efficiency in TCP Flow Simulation

Efficient and accurate network simulation techniques are critical for evaluating protocols and systems at scale. For example, the magnitude of modern data center deployments requires the use of fast simulation techniques in order to evaluate newly proposed algorithms and system architectures. Unfortunately, efficient simulation can come at the cost of accuracy and realism. In this work-in-progress paper, we examine specific limitations of existing closed-form models of TCP throughput, which are commonly used to simulate aggregate performance of TCP flows. We evaluate two models in particular, comparing predictions of the models with actual performance measured using the Mininet platform. We find that the open-loop nature of these models and sensitivity to different parameters contributes to significant inaccuracies, which may in turn lead to incorrect conclusions when using such models as the basis for simulation. We describe our ongoing work on a new method for scalable TCP flow simulation that is based on ideas from XCP. Our proposed technique is highly efficient, incorporates network feedback in a closed-loop manner, and in our initial experiments shows appreciable improvement in accuracy over prior models.