ROCK-CNN: A Distributed RockPro64-based Convolutional Neural Network Cluster for IoT. Verification and Performance Analysis

Abstract

The paper is dedicated to optimization of machine learning and neural networks applications by replacing common servers with Single Board Computer (SBC) clusters to minimize mounting and service expenses, simplify node mounting process and organize parallel computing in IoT applications. Authors focus on former experience of using distributed computing, mainly, light-weight and cost-optimized SBCs to classify use-cases, then, choose an appropriate hardware platform enabling sufficient data processing and easy hot-replacing of nodes. This task requires organizing an efficient software architecture to make use of advantages of SBCs. A comparison for various SBSs is presented. Authors suggest their formerly-designed architecture with changes allowing using it for neural network applications. Authors pay attention to thorough parameter examination based on numerous tests. Parameter timelines are presented in the paper. The paper describes a number of test-cases to validate the efficiency of suggested architecture based on common use-cases. Performance analysis and cluster scalability potential estimation are conducted as well to estimate an efficient number of nodes required for future tasks.

Keywords

• single-board computer
• deep learning
• computer clusters
• iot
• convolutional neural network
• intelligent pseudo real-time architecture
• rockpro64
• performance testing