Massively Parallel Analysis of Similarity Matrices on Heterogeneous Hardware.

We conduct a study that investigates the performance characteristics of a set of parallel implementations of the recurrence quantication analysis (RQA) using OpenCL. Being an important tool in climate impact and medical research, a central aspect of RQA is the construction of a binary matrix that captures the similarities of multi-dimensional vectors. Based on this matrix, quantitative measures are derived. Starting with a baseline implementation, we diversify its properties along four dimensions: the representation of input data, the materialisation of the similarity matrix, the representation of similarity values and the recycling of intermediate results. We evaluate the performance of ve implementations by varying the input parameter assignments, the hardware platform employed for execution and the default OpenCL compiler optimisations status. We come to the conclusion that the performance of conducting RQA highly depends on the selected implementation as well as the combination of these variables under investigation. Dierences in runtime of up to one order of magnitude are observed, emphasising the importance of performance studies as presented here.