Scalable Multi-core Architectures for Data Mining Applications

Abstract

Over the past few decades we have witnessed an exponential growth in the amount of data being used in virtually all domains. Extracting useful information from massive amounts of data is crucial and hence lot of data mining applications employing complex algorithms have been developed to aid in this task. As these data mining applications are gaining prominence it is very important to understand and address these application needs and translate them into design choices to have efficient and scalable processor architectures. This thesis is a step in this direction and involves investigating architecture alternatives that offer maximum scalability for data mining applications. For investigation we choose the Minebench benchmark suite from the data mining application domain and specifically worked with clustering and classiffication benchmarks. For evaluating the different architectural alternatives, we employ the SESC simulator. We identify the different architectural alternatives to be used for exploration by carrying out performance analysis experiments and identifying characteristics common to all the applications in this domain. Based on these observations, we select Asymmetric Chip Multi-Processor and Heterogeneous Multi-core Processor as candidate architectures and evaluate them by comparing them to a baseline Homogeneous Chip Multi-Processor. The evaluation shows that Asymmetric Chip Multi-Processor architectures provide better performance scalability than Homogeneous Chip Multi-Processor architectures for all the benchmarks considered. For Heterogeneous Multi-core Processor architectures we use two different scheduling strategies and observe large differences in performance results. Based on the results we conclude that Asymmetric Chip Multi-Processors consistently perform better than Heterogeneous Multi-core Processors.