Performance analysis of machine learning applications on rapid: a highly parallel computer architecture

Over the past few years, the interest and application of machine learning algorithms has risen exponentially. Machine learning has found extensive use in diverse fields like self-driving cars, speech recognition, image processing, computer vision, molecular biology, security etc. A lot of recent research involves evaluation of machine learning applications on different architectures. In this thesis, we evaluate the performance of six common machine learning algorithms: K-Means, K-Nearest Neighbors, Linear Regression, Latent Dirichlet Allocation, Deep Neural Network, and Radix Sort on RAPID. RAPID is a highly parallel computer architecture developed at Oracle Labs for accelerating and improving the performance of database analytic workloads. We find that the RAPID platform performs well on the performance-per-watt metric i.e. it is a power-efficient architecture. Moreover, the machine learning applications can be easily scaled to hundreds of nodes of the RAPID architecture, thereby making it suitable for distributed machine learning applications. However, we find certain bottlenecks in the micro-architecture, memory system and network of the RAPID architecture and propose optimizations to make it a more performance efficient architecture for machine learning applications