Dimension Reduction of Network Bottleneck Bandwidth Data Space

Abstract

The network proximity metrics, such as bottleneck bandwidth and round-trip time, are very useful in different network applications. The round-trip-time prediction has been studied extensively. However, the prediction of bottleneck bandwidth has received much less attention. Therefore, we attempt to design a new bottleneck bandwidth prediction system by matrix factorization. As a first step, we focus on the dimension reduction of network bottleneck bandwidth data space in this paper. Evaluation is carried out based on real-world bottleneck bandwidth datasets, which are collected in the past three months. The results show that a 250D data space can be compressed to 10D and the average median-relative-error is only 8.65%. Although preliminary, our work provides some insights into the design direction towards matrix factorization based distributed system to predict the bottleneck bandwidth.