Using Machine Learning Method for Probing Congestion in High-Performance Datacenter Network

Abstract

Even if, as we all known, High Performance Computing (HPC) systems can greatly reduce application performance, there is little quantification of congestion on credit-based interconnected networks. This paper proposed a method for detecting, extracting and characterizing congested regions in the network. The authors have implemented this methodology in a deployable tool, Monet, which can provide such analysis and feedback at runtime. Using Monet, we can characterize and diagnose the congestion in Blue Waters, the world’s largest 3D torus network. Blue Waters is a 13.3- petaflop supercomputer at the National Center for Supercomputing Applications.