High-performance online spatial and temporal aggregations on multi-core CPUs and many-core GPUs

Abstract

Motivated by the practical needs for efficiently processing large-scale taxi trip data, we have developed techniques for high performance online spatial, temporal and spatiotemporal aggregations. These techniques include timestamp compression to reduce memory footprint, simple linear data structures for efficient in-memory scans and utilization of massively data parallel GPU accelerations for spatial joins. Our experiments have shown that the combined performance boosting techniques are able to perform various spatial, temporal and spatiotemporal aggregations on hundreds of millions of taxi trips in the order of a few seconds using commodity personal computers equipped with multi-core CPUs and many-core GPUs. The high throughputs in a personal computing environment are encouraging in the sense that high-performance OLAP queries on large-scale data is feasible when the parallel processing power of modern commodity hardware is fully utilized which is important for interactive OLAP applications.