Data Mining On The Cell Broadband Engine

The STI Cell Broadband Engine architecture represents an interesting design point along the spectrum of chipsets with multiple processing elements. In this article we investigate key mining tasks such as clustering, classification, anomaly detection and PageRank on the Cell along the axes of performance, programming complexity and algorithm design. As part of our comparative analysis we juxtapose these algorithms with similar ones implemented on state-of-the-art uniprocessor and multicore architectures. For the workloads that are more floating point intensive, and where data is accessed in a streaming fashion the Cell processor is up to seven times faster than competing technologies, when the underlying algorithm uses the hardware efficiently. However, when required to write in a non-streaming fashion, as with PageRank, the processor is many times slower than competing processors. An outcome of our benchmark study, beyond the results on these particular algorithms is that we answer several higher level questions, which are designed to provide a fast and reliable estimate to application designers for how well other workloads will scale on the Cell.