High Performance Information Filtering on Many-core Processors

Abstract

The increasing amount of information accessible to a user digitally makes search difficult, time consuming and unsatisfactory. This has led to the development of active information filtering (recommendation) systems that learn a user’s preference and filter out the most relevant information using sophisticated machine learning techniques. To be scalable and effective, such systems are currently deployed in cloud infrastructures consisting of general-purpose computers. The emergence of many-core processors as compute nodes in cloud infrastructures necessitates a revisit of the computational model, run-time, memory hierarchy and I/O pipelines to fully exploit available concurrency within these processors.

This research proposes algorithms & architectures to enhance the performance of content-based (CB) and collaborative information filtering (CF) on many-core processors. To validate these methods, we use Nvidia’s Tesla, Fermi and Kepler GPUs and Intel’s experimental single chip cloud computer (SCC) as the target platforms. We observe that ~290x speedup and up to 97% energy savings over conventional sequential approaches. Finally, we propose and validate a novel reconfigurable SoC architecture which combines the best features of GPUs & SCC. This has been validated to show ~98K speedup over SCC and ~15K speedup over GPU.