Optimizing a Collaborative Filtering Recommender for Many-Core Processors

The web is moving from an era of "search" to that of "discovery". Collaborative filtering (CF) recommender systems are now commonly used to predict user\´s preference towards an unknown item from past ratings. To be scalable or effective, they are typically deployed in distributed clusters and operate on extremely large apriori datasets. Improvement of the efficiency of these systems is increasingly recognized important and challenging. Meanwhile, emerging many-core processors present an opportunity to optimize these systems on a per-node basis. We identify and address challenges in fast computation of correlations by maximizing data locality and minimizing communication cost between individual cores. We experiment run-time, power and energy consumed on: (1) Intel\´s experimental single chip cloud computer (SCC), (2) NVIDIA\´s CUDA-enabled GPGPU co-processor and (3) traditional server class x86 processor. We achieve super linear speedups (~30x), reduction in energy consumption (~90%) for benchmark workloads. Introduction of this design in CF systems can significantly reduce the number of servers required in a data center, energy consumption, operation costs and floor area bringing in significant savings.