High Performance Machine Learning (HPML) Framework to Support DDDAS Decision Support Systems: Design Overview

This paper presents a design for a High Performance Machine Learning (HPML) framework to support DDDAS decision processes. The HPML framework can provide a high performance computing environment to implement large scale machine learning algorithms that leverages Big Data tools (e.g., SPARK, Hadoop), parallel algorithms, and MapReduce programming paradigm. The framework provides the following capabilities: • High Performance Parallel Algorithms: For a suite of important ML, we will develop three parallel implementations of each algorithm that are based on Message Passing Interface (MPI), Shared Memory (SM) and MapReduce programming model. • High Performance and Scalable Platforms: This will enable us to identify the best high performance platform that maximizes performance and scalability of the parallel ML methods. We will experiment with and evaluate the performance and scalability of different parallel architectures (shared memory and message passing), Clusters of GPUs, and cloud computing systems. By leveraging the emerging Big Data tools and high performance computing algorithms (traditional and emerging paradigm such as MapReduce), we will be able to achieve the following: 1) reduce significantly the ML processing time, 2) enable StreamlinedML users to leverage Big Data tools to perform large scale ML tasks over structured and non-structured data sets; and 3) enable users to identify the best parallel platform and storage allocation and distribution that maximize performance and scalability of the selected ML algorithms.'