Cyberinfrastructure for Accelerating Physics & Astronomy Applications With Many-core and Accelerator-Based Systems

University Of California-Berkeley, Berkeley CA

Investigators

Abstract

The rapid increase of data rates and volumes (peta-operations/second, 15 petabytes/year) from physics and astronomy simulations, observations, experiments and analyses are reaching critical computational impasse. To meet the demands of petascale to exascale computing challenges, fundamentally new energy efficient supercomputing architectures and solutions will have to be researched and developed. The uniqueness of this work is in adopting science/application-driven approach where the key application-drivers are identified up-front to assess the efficacy of this new approach. With wide appeal, the new period of energy efficient science with new performance metrics such as operations-per-watt presents a great opportunity to lead the progress of scientific research. We will build a cyberinfrastructure of comprehensive software libraries, tools, frameworks, easy-assembly common hardware modules and complete turnkey solutions by leveraging emerging many-core architectures with emphasis on Graphics Processing Units (GPUs). The three-year research plan will develop algorithms, and hardware infrastructures for efficient scalable solutions directly applicable to a broad range of compute-intensive scientific problems. The set of applications are categorized into three separate domains - simulation, instrumentation and data processing - covering specific real-case challenges in cosmology, astronomy, optics, and image/data processing with potential of interdisciplinary relevance. The developed cyberinfrastructure will be released to the broader scientific community with methodologies for easy implementation. Successfully harnessing the power of the parallel architectures such as GPUs for compute-intensive scientific problems via the planned cyberinfrastructure will open doors for new discovery and revolutionize the growth of science. The infrastructure will actively identify interdisciplinary acceleration overlaps and will alleviate adoption. Extremely high-speed massive simulations will cut the overall execution times by several orders of magnitudes, thereby reducing monthly time cycles, prone to malfunctions and delays, to hours and minutes. Remote on-site handling of high data rates will make real-time imaging in radio astronomy possible for the first time. Partnerships have been established with international groups in National Astronomical Observatories of China, and University of Heidelberg, Germany. The proposed research will engage and enable students. The combination of low cost devices and cyberinfrastructure will supply affordable high performance computing for young researchers and students. The infrastructure will be released to the broader community in yearly cycles with open source license punctuated with workshops to widen the scope of the research.