Toward Fair and Efficient Congestion Control: Machine Learning Aided Congestion Control (MLACC)

Abstract

Emerging inter-datacenter applications require massive loads of data transfer which makes them sensitive to packet drops, high latency, and fair resource sharing. However, current congestion control (CC) protocols do not guarantee the optimal outcome of these metrics. In this paper, we introduce a new CC technique, Machine Learning Aided Congestion Control (MLACC), that combines heuristics and machine learning (ML) to improve these three network metrics. The proposed technique achieves a high level of fairness, minimum latency, and minimum drop rate. ML is utilized to estimate the ratio of the available bandwidth of the bottleneck link while the heuristic uses this ratio to enable end-points to cooperatively limit the shared bottleneck link utilization under a predefined threshold in order to minimize latency and drop rate. The key to achieving the desired fairness is using the gradient of the link utilization to control the sending rate. We compared MLACC to BBR (which is at least on par with the state-of-the-art ML-based techniques) as a base case in different network settings. The results show that MLACC can achieve lower and more stable end-to-end latency (25% to 52% latency saving). It also significantly reduces packet drop rates while attaining a higher fairness level. The only cost for these advantages is a small throughput reduction of less than 3.5%.