GOALI: An Exploration of the Use of OpenCL for Numerical Modeling and Data Analysis

University Of Massachusetts, Dartmouth, North Dartmouth MA

Investigators

Abstract

This award supports research on using many-core compute architectures, such as modern GPUs and the Cell BE, for accelerating modeling and data analysis computations in the area of gravitational wave physics. A significant challenge in making effective use of such accelerator hardware is that they typically require specialized and vendor-specific software development. For example, Nvidia GPUs use the CUDA SDK, while the Cell BE uses the IBM Cell SDK, and these SDKs differ in many important ways. Thus, programming multiple accelerator hardware types can often become a difficult task for a computational scientist involving considerable redundant software development. The Open Computing Language (OpenCL), led by Nvidia (the industry partner on this GOALI award) and Apple (with a collaborating researcher on this project) is designed to be cross-platform and vendor neutral, and is supported on all the major computer hardware currently available (CPUs, Nvidia & ATI GPUs, IBM Cell BE, etc). OpenCL may give computational scientists a chance to explore several accelerator technologies with significantly less code development. The main goal of this project is to perform a careful evaluation of OpenCL (and closely related frameworks) for scientific computing (numerical modeling and data analysis) on a variety of different hardware architectures. A detailed investigation of these software development frameworks will be performed, starting with the ease of porting current scientific code, to the final performance levels that can be obtained on various different hardware architectures. This approach could yield significantly higher performance and cost-effectiveness when compared with traditional multi-core processors. More specifically, the project will focus on investigating this OpenCL-based optimization approach for projects in numerical relativity and gravitational wave data analysis. Preliminary work (based on CUDA and Cell SDKs) performed by the PI has yielded promising results. Many-core architectures such as GPU and Cell BE, promise significant performance gains for numerical modeling and data analysis tasks at a relatively low cost; however, programming such hardware accelerators can involve considerable redundant software development. Since human resources are often the most expensive and constrained aspects of code development, it would be highly desirable to make use of a framework in which such redundant work is unnecessary. It would give computational scientists a chance to explore multiple computer architectures with less effort. A specific outcome will be OpenCL-based optimization gravitational physics codes to make an immediate impact on the gravitational wave data analysis and source modeling communities. This GOALI award involves graduate students in the research and therefore contributes directly to student training, especially with a strong computer industry focus. In addition, the proposed research establishes collaborations between academia and industry. The work involves direct transfer of knowledge, skill and technology from industry to academia. The industry advisors will help to train the graduate students in cutting-edge software development frameworks like OpenCL, thus generating highly employable engineers and scientists. The outcome of the project and any resulting general purpose code or software libraries that are developed, will be freely distributed via the project's dedicated website.