Computing Bottleneck Structures at Scale for High-Precision Network Performance Analysis
Data Analytics, Compression, and Management
Performance/Productivity Measurement and Evaluation
TimeThursday, 12 November 20205:40pm - 6:05pm EDT
DescriptionThe Theory of Bottleneck Structures is a recently-introduced framework for studying the performance of data networks that describes how local perturbations in one part of the network propagate and interact with one another. This frame- work provides a powerful analytical tool for network operators to make accurate predictions about network behavior and optimize performance. Previous work implemented a software package that leveraged the Theory of Bottleneck Structures to address several network optimization problems, but applied it only to simple examples. In this work, we introduce the first software package capable of scaling the bottleneck structure analysis to production-sized networks. We benchmark our system using logs from ESnet, the Department of Energy high-performance data network used to connect research institutions in the US. Using the previously published tool as a baseline, we demonstrate that our system achieves vastly improved performance, allowing the bottleneck structure analysis to be applied to rapidly-changing network conditions in real time. We also study the asymptotic performance of our core algorithms, demonstrating strong agreement with theoretical bounds and improvements over algorithms of the baseline. These results indicate that the proposed software package is capable of scaling to very large data networks. Overall, we demonstrate the viability of using bottleneck structures to perform high-precision bottleneck and flow analysis.