Time Series Data Mining using Matrix Profiles on FPGAs

Time series data mining has become essential for a large variety of disciplines ranging from seismology to medicine. With the use of matrix profiles, most target analyses, especially motif or discord discovery, have become trivial to a large extend. As the original formulations for calculating the matrix profile were relatively ineffective, substantial effort has been invested in computing them efficiently. In particular, the state-of-the-art algorithm SCAMP has convinced with its simplicity, parallelizability, and efficiency. As FPGA acceleration promises a combination of high performance and competitive energy efficiency, we present a - to the best of our knowledge first - systolic array-based design, based on the SCAMP algorithm, to compute matrix profiles efficiently on FPGAs. Our design is mapped to a concrete architecture using a High-Level Synthesis tool. This approach allows us to maintain a high level of abstraction, use streaming abstractions, and enable maintainability and portability across FPGA devices. Kernels synthesized from our design are shown to offer competitive performance in practice, scaling with both compute and memory resources. Finally, we outline possible optimizations as a basis for future work.