I-Corps: Graphics Processing Unit-Based Data Management System Software

University Of South Florida, Tampa FL

Investigators

Abstract

The broader impact/commercial potential of this I-Corps project lies in the improved database productivity that the technology can bring to many industries. Existing software systems based on traditional pull-based database engine design and CPU hardware often do not provide the desired high throughput and/or short response time in the management and analysis of large-scale data. The technology developed here offers a more efficient solution to meet such challenges and its commercialization will potentially impact the practice of data management and analysis in a large number of industries including healthcare, retailing, and online advertising. Successful deployment of these the systems developed here will enable a range of industry sectors to efficiently harness big data analytics. This I-Corps project explores the commercial potential of a data management system with a novel software architecture and advanced data analytics functionalities built on massively parallel hardware such as Graphics Processing Units (GPUs). The technology features a data processing engine under a data streaming (push-based) design and achieves very high utilization of a multitude of computing resources on GPUs. Current experiments show that the data processing throughput and latency are about one order of magnitude better than best known systems built for the same purposes. The research behind this technology complements current work in database systems and parallel computing fields. The system-oriented approach is in sharp contrast to existing GPU work that typically focuses on solving individual problems and thus highly creative. Innovations are also seen in the formal methods used for meeting system design challenges such as dynamic resource allocation in GPUs.