GPU-Based Aggregation of On-Line Analytical Processing

Abstract

OLAP (On-Line Analytical Processing) is data and compute intensive application, how to improve the performance of OLAP are researchers always pursued goal. Aggregation is one of high frequently used operations which have a great impact on OLAP performance. Modern GPU (Graphic Process Units) have more raw computing power and higher memory bandwidth, so utilizing GPU accelerating aggregation computation is straight forward. But now GPU equipment does not supports float atomic operation and incremental memory allocation, so GPU algorithm need to be well-designed. In this paper, we discuss real-time aggregation in OLAP based on dense and sparse dataset, which fully utilize the high parallelism and high memory bandwidth and achieve performance improvements approximately 20X over CPU-based algorithms. On dense dataset, source data are chunked based on shared memory size, each thread block processes one chunk, each thread in block computes one cell in chunk cuboid. Algorithms adapts to GPU architecture and high parallelism which ensure high performance of algorithms. But on sparse dataset, there is a complex relationship between the compression dataset and the unknown size of result cuboid, it is impossible to define a straightforward parallelization. So we utilize sort, map and prefix sum primitive finishing source data partition, and reduction primitive aggregation data. At last, we introduce prototype system GPUOLAP (GPU-based OLAP) architecture which is under development now. Our work is a good attempt to real-time OLAP using new hardware.