Task scheduling for GPU accelerated OLAP systems

  • Authors:

  • Lubomir Riha;Colin Shea;Maria Malik;Tarek El-Ghazawi

  • Affiliations:

  • The George Washington University, Washington, DC;The George Washington University, Washington, DC;The George Washington University, Washington, DC;The George Washington University, Washington, DC

  • Venue:
  • Proceedings of the 2011 Conference of the Center for Advanced Studies on Collaborative Research
  • Year:
  • 2011

Quantified Score

Hi-index 0.01

Visualization

Abstract

Processing data related to business intelligence is an ever important and complex task set. One approach to answering multi-faceted analytical queries quickly is online analytical processing, or OLAP. OLAP allows for quick query response times thanks to its use of n-dimensional models referred to as OLAP cubes. As with all data laden systems, OLAP systems are dealing with ever-increasing dimensionality of the data cube while expecting system responsiveness to be maintained. As queries become more complex and the dimensionality of the cube grows ever larger, runtime required to aggregate queries increases. To coalesce these requirements, without impacting the apparent dimensionality of the cube, an agile method for reporting must be found. In this paper we propose a task-scheduling algorithm for GPU accelerated OLAP systems. This scheduling algorithm looks to balance the GPU and CPU load to meet a minimally acceptable completion time for OLAP queries. A partial in memory cube is formed using highest-level general queries. To ensure fast response time of aggregations, the cube is restricted in dimensionality. If a query requires data outside of the dimensional cube, or the time to search the cube is greater than the time to execute a raw aggregation, the task is scheduled on the GPU for processing. Our evaluation of the heterogeneous task scheduler shows a performance increase of 8.5x over a CPU only OLAP system.