HASTE: A Heterogeneously Accelerated SQL Transaction Engine
MetadataShow full metadata
Databases are the backbone of the digital age and empower the storage and processing of massive amounts of data. As private and public data grows at an astonishing rate, the technologies that drive data processing must adapt or be discarded. Conventional database engines struggle to provide responsiveness at the level required of them when faced with ever-expanding datasets and more demanding use cases. With the recent surge in public adoption of hardware parallelism and co-processor offloading, we have explored the concept of employing new parallel processing techniques and technologies to a database management system (DBMS) to achieve higher query processing performance. In this paper, we demonstrate a custom-designed, modular DBMS targeted at parallel platforms including the Intel Many Integrated Core architecture and an Nvidia CUDA platform. Our Heterogeneously Accelerated SQL Transaction Engine (HASTE) uses a novel query parsing methodology to create a hardware agnostic query definition which can be processed by adaptable modules written for new and existing hardware and software platforms and executed on one more such modules simultaneously. This paper demonstrates these modules designed for a modern CPU, Xeon Phi 5110 co-processor from Intel, and Tesla K20 GPGPU from Nvidia, but can also be extended to run on virtually any technology that interfaces with the HASTE host kernel. Through experimenting on both synthetic and real-world data, we achieve a speedup of up to 2000 percent with the Xeon Phi and 6700 percent with the Tesla hardware.
CitationRomoser, B. M. (2014). HASTE: A heterogeneously accelerated SQL transaction engine (Unpublished thesis). Texas State University, San Marcos, Texas.