GPU-Accelerated VoltDB: A case for indexed nested loop join

Graphics Processing Units (GPUs) are traditionally designed for gaming purposes. The new GPU hardware and new programming platforms for GPU applications have enabled GPUs to work as co-processors alongside Central Processing Units (CPUs) in order to speed up general purpose applications. In this paper, we focus on the design and implementation of the GPU-Accelerated indexed nested loop join (INLJ) for in-memory relational database management system (RDBMS). Previous studies have proposed novel approaches for using GPU to improve the performance of the relational INLJ, but they are only implemented on simulation systems. Their performance in current industry RDBMS still needs to be clarified. To this end, we implement the GPU-Accelerated INLJ algorithm and perform various experiments on that join in VoltDB, an inmemory commercial RDBMS. We also propose a method for handling skewed input data, which is a critical problem in the GPU INLJ. Our evaluations indicated that though the GPU-Accelerated INLJ is 2-14X faster than the default INLJ of VoltDB, the memory copy between the host and the GPU memory is the major factor that holds back the join's speedup rate.