Modular supercomputing design supporting machine learning applications

Abstract

The DEEP-EST (DEEP — Extreme Scale Technologies) project designs and creates a Modular Supercomputer Architecture (MSA) whereby each module has different characteristics to serve as blueprint for future exascale systems. The design of these modules is driven by scientific applications from different domains that take advantage of a wide variety of different functionalities and technologies in High Performance Computing (HPC) systems today. In this context, this paper focuses on machine learning in the remote sensing application domain but uses methods like Support Vector Machines (SVMs) that are also used in life sciences and other scientific fields. One of the challenges in remote sensing is to classify land cover into distinct classes based on multi-spectral or hyper-spectral datasets obtained from airborne and satellite sensors. The paper therefore describes how several of the innovative DEEP-EST modules are co-designed by this particular application and subsequently used in order to not only improve the performance of the application but also the utilization of the next generation of HPC systems. The paper results show that the different phases of the classification technique (i.e. training, model generation and storing, testing, etc.) can be nicely distributed across the various cluster modules and thus leverage unique functionality such as the Network Attached Memory (NAM).