high performance computing on graphics processing units: hgpu.org

In the past few years, performance improvements in CPUs and memory technologies have outpaced those of storage systems. When extrapolated to the exascale, this trend places strict limits on the amount of data that can be written to disk for full analysis, resulting in an increased reliance on characterizing in-memory data. Many of these characterizations are simple, but require sorted data. This paper explores variations on two classic algorithms for distributed sorting-radix and sample sort – under two novel constraints imposed by the projected requirements of an exascale machine, heterogeneity and limited external storage. The two approaches are evaluated on the GPU-based NSF Keeneland system, including an analysis of data movement and the effects of GPUs on performance and scalability. Results from Keeneland indicate a substantial performance advantage for sample-based approaches on some data distributions, but this advantage comes at the cost of randomized behavior and load imbalance.