Efficient 2-Body Statistics Computation on GPUs: Parallelization & Beyond

Various types of two-body statistics (2-BS) are regarded as essential components of data analysis in many scientific and computing domains. Due to the quadratic time complexity, use of modern parallel hardware has become an obvious direction for research and practice in 2-BS computation. This paper presents our recent work in designing and optimizing parallel algorithms for 2-BS computation on Graphics Processing Units (GPUs). First, we classify 2-body applications into three groups based on their data output pattern. Then, we introduce a straightforward parallel algorithm under the CUDA framework. The unique architecture of modern GPUs, however, provides abundant opportunities for optimizing the algorithm. To that end, we split the algorithm into two stages: pairwise distance function computation and writing output. Then, we present modifications to the basic algorithm by integrating various techniques at each stage. Since the architecture of modern GPUs is much more complex than that of multi-core CPUs, traditional wisdom on decomposing problems in a parallel platform is often insufficient in developing GPU-based algorithms. Therefore, our algorithms design focuses on effective use of hardware/software features that are unique in GPU platforms. In addition to the various programming cache and atomic operations, we also introduce novel load balancing and register content sharing techniques. We develop models to analyze such techniques and identify the best ones for each type of 2-BS. Experiments run on modern GPU hardware show that our GPU algorithms outperform the best known CPU program by at least an order of magnitude in various applications. Furthermore, our implementation achieves very high level of GPU resource utilization, indicating nearoptimal performance. This work builds a solid foundation towards realizing our vision of a framework that can automatically generate optimized code for any new 2-BS problems.