RNN-based Approach to TCP throughput prediction

Luyao Bai, Hirotake Abe, Chunghan Lee

Abstract: Conventional TCP congestion control algorithms, such as newReno and CUBIC, were designed to avoid bandwidth collapse at times of network congestion. However, the different network scenarios and various network applications associated with congestion would make it difficult to maintain TCP throughput during this time. Thus, we focus on how a recurrent neural network (RNN)-based TCP throughput prediction method can improve throughput under the conditions of congestion to solve this problem. Our proposed method involves training an RNN on the metrics of a TCP connection to predict the TCP throughput. We prepared eight groups of metrics for training, and applied three types of RNNs (i.e., simple-RNN, gated recurrent unit (GRU), and long short-term memory (LSTM)) in our method design. Under our congestion scenarios, a combination of metrics can contribute to improve the throughput. The throughput predicted via our method indicates that it performs better than the TCP CUBIC algorithm. Furthermore, the GRU-based predictive model demonstrated the best performance, achieving 35% higher throughput than the TCP CUBIC algorithm.