Analytical and Experimental Methods for High-Performance Network Testing

Abstract

There has been an increasing number of large-scale science and commercial applications that produce large amounts of data, in the range of petabytes to exabytes, which has to be transported over wide area networks. Such data transport capability requires high performance protocols together with complex end systems and network connections. A systematic analysis and comparison of such data transport methods involves the generation of the throughput profiles from measurements collected over connections of different lengths. For such testing, the connections provided by production networks and testbeds are limited by the infrastructures, which are typically quite expensive. On the other hand, network emulators provide connections of arbitrary lengths at much lower costs, but their measurements only approximate those on physical connections. We present a differential regression method to estimate the differences between the performance profiles of physical and emulated connections, and then to estimate “physical” profiles from emulated measurements. This method is more general and enables: (i) an objective comparison of profiles of different connection modalities, including emulated and physical connections, and (ii) estimation of a profile of one modality from measurements of a different modality by applying a differential regression function. This method is based on statistical finite sample theory and exploits the monotonicity of parameters to provide distribution-free probabilistic guarantees on error bounds. We present an efficient polynomial-time dynamic programming algorithm to compute the underlying differential regression function. We provide a systematic analysis of long-haul InfiniBand and TCP throughput measurements over dedicated 10Gbps connections of several thousands of miles. These results establish the closeness of throughput profiles generated over plain, encrypted, physical and emulated connections. In particular, our results show that robust physical throughput profiles can be derived using much less expensive emulations, thereby leading to significant savings in cost and effort.