The nature of high performance computing is undergoing rapid change with the advent of general purpose GPU-accelerated architectures. These changes open new opportunities for scientific discovery at a much reduced capital and energy cost, but to exploit them, new algorithms and methods are required.
This project supports the collaborative development and implementation of GPU (Graphical Processing Unit) codes and innovative algorithms for the numerical simulation of the strong interactions of quarks and gluons, namely quantum chromodynamics (QCD), on a lattice. Specifically, this project supports the development of CUDA (parallel coding model from NVIDIA)-based code modules and OpenMP threading that can be used with the MILC [Multiple Instruction (and Multiple Data) Lattice Calculation] code. This 200,000-line code suite is used by several research groups around the world and supports some 100 million core-hours of computing at NSF and DOE national centers and laboratories mostly on conventional supercomputers. Our code development will allow improved performance on GPU clusters and on computers such as Blue Waters at NCSA that contain both GPUs and conventional multicore processors. Performance models developed will guide the performance tuning and help to plan future calculations and should also be of value in studing the effectiveness of exascale architectures.
Interpretation of many of the experimental results in particle physics is currently limited by lack of theoretical understanding. This reduces our ability to determine parameters of the Standard Model and to find evidence for physics beyond the Standard Model through precision experiments. The calculations that are enabled are thus important for the interpretation of many experiments in elementary particle physics. Some of these experiments at Fermilab, Cornell, and at SLAC have been completed. Others such as in Beijing, Geneva, and Japan are still running.
We have publicized previous work on code development and performance modeling by presentations at appropriate conference and publication. We will continue to do so for the current work in order that others can benefit from our methodology and innovations. The MILC code is widely used in benchmarking by several high-performance computing centers. This in turn has led to several vendors contacting us regarding performance of our code on their current and future chips. Thus, having GPU-accelerated code available in MILC could help in the evaluation of or co-design of future high-performance computing architectures.
Finally this project will contribute to the professional education and training of future scientists via the participation of postdoctoral reseachers.
