EAGER: Collaborative Research: Using PDE Descriptions to Generate Code Precisely Tailored to Energy-Constrained Systems Including Large GPU Accelerated Clusters

点击翻译按钮获取中文摘要

Modern computer system architectures are forcing computational scientists to move scientific applicationsfrom traditional homogeneous cpu-based systems to heterogeneous multi-core/accelerator architectures. Obtaining performance in the presence of accelerators requires close attention tothe memory hierarchy and chip-level parallelism to reach even a modest fractionof the potential performance. As a result, coding tasks which were once the province oflone graduate students in a single discipline now require interdisciplinary teams of people. Project Chemora will explore the design of a new application framework for automaticallycreating highly optimized code for high-end computational machines. The systemwill use as input a set of partial differential equations (PDEs) that describe aproblem, it will then construct a machine-specific abstract performance model, and using theseit will generate well-tuned code and execution configurations for accelerated(e.g., hybrid CPU/GPU) computing clusters at various scales. Chemora willimprove programmability in this simplified domain by decoupling the science andcomputer science at a high level, thereby reducing the complexity and number of issues scientists need tocollectively understand and allowing individual scientists in the team to focus on their area ofspecialty. Chemora will improve performance (both wallclock time and energy) forsystems with both simple and complex sets of equations by making use of detailedinformation describing the problem and machine, and will provide improved loadbalancing through the AMPI framework.The Chemora project has chosen the Einstein equations as the primary science driver becausethese equations are one of the more complex PDE systems, one with manyhundreds of terms, and a problem scale that is challenging to optimize for mostcompilers. Achieving this vision for a general scientific problem would indeedbe a "Grand Challenge" in computational science, but in order to give ourresearch a sharper focus we have chosen as a science driver thesimulation of Intermediate mass ratio Binary Black Hole (IBBH) systems. Suchsystems, consisting of a black hole of mass 100 to 1,000 solar masses orbited bya smaller black hole of mass 5 to 20 solar masses are expected to be importantsources of gravitational waves for advanced Laser Interferometer GravitationalWave Observatory (LIGO) and the Einstein Telescope (ET). Accurate modeling ofthe waveforms from IBBH systems will be necessary in order to extractgravitational wave signals using template-matching data analysis techniques.