In-memory OLAP aggregation on GPUs using CUDA Dynamic Parallelism

Most queries involved with Online Analytical Processing (OLAP) depend on the functionality of aggregating data along the multidimensional hierarchies of an OLAP cube. In real-time OLAP, aggregated data for interactive operations e.g. roll-up and drill-down is computed on-the-fly. Fast response times are essential and can be accelerated significantly through data-parallel computation on graphics processing units (GPUs). In this thesis, an existing parallel algorithm is modified to use a technology called CUDA Dynamic Parallelism (CDP). Using this technology, GPU programs can be launched directly from within other GPU programs to extract more parallelism. Furthermore, we present a preaggregation method using the CUDA shuffle command to optimize both GPU implementations. For evaluation purposes, we additionally implement a sequential aggregation algorithm. Our experiments show that the single-threaded CPU implementation is outperformed by the GPU implementations by 16 to 218 times. The experiments further show that the CDP implementation reaches a speedup of 3.72 times over the non-CDP implementation when processing queries for an artificial OLAP cube. However, using CDP causes an average of 1.42x slowdown to the processing of queries in a typical OLAP scenario.