Investigations into the robustness of computer-synthesized congestion control

Abstract

Recent work has shown that computer-synthesized TCP congestion control protocols can outperform the state of the art. However, these protocols are generally too complex to reason about. Human engineers therefore might not trust them enough to deploy them in real networks. This thesis presents two contributions toward the practical deployment of computer-synthesized congestion-control algorithms. First, we describe a simple, human-designed protocol that performs comparably to computer-optimized protocols using only 10 lines of code, suggesting that it may be feasible to optimize for interpretability in addition to performance. Second, we introduce techniques for reasoning about the behavior of black-box protocols via extensive simulation, which reveal regions of potentially undesirable behavior in both computer-optimized protocols and a NewReno-like TCP implementation, highlighting areas to focus further engineering effort.