bby Brookhaven National Laboratory
GPU Research Centers are institutions that embrace and use graphics processing unit (GPU)-accelerated computing across multiple research fields, and are at the forefront of some of the world's most innovative scientific research. GPU-accelerated computing leverages the parallel processing capabilities of GPU accelerators and enabling software to deliver dramatic increases in performance for scientific, artificial intelligence, machine learning, graphics, engineering and other demanding applications.
Brookhaven Lab was recognized for its use of GPU-accelerated computing to conduct research in fields including materials science, physics and climate science, and for its vision to further the application of GPU-accelerated computing in those and other research fields with a high computational demand. Current GPU computing resources at the Lab include GPU-accelerated Linux clusters. A new 200-node GPU-accelerated institutional cluster is expected to go into operation in June 2016.
Working under this initiative, computational scientists at Brookhaven Lab are seeking to establish partnerships with key players in academia and industry. Barbara Chapman, head of Brookhaven's Computer Science and Mathematics Group, and Nicholas D'Imperio, chair of Brookhaven's Computational Science Laboratory, led the proposal to bring an
"The center will enable Brookhaven Lab to collaborate with NVIDIA on the development of widely deployed codes that will benefit from more effective GPU use, and in the delivery of on-site GPU training to increase staff and guest researchers' proficiency," said Kerstin Kleese van Dam, director of CSI and chair of the Lab's Center for Data-Driven Discovery.
These codes will include streaming data-analysis codes for the near-real-time analysis of experimental results at the National Synchrotron Light Source II, the Center for Functional Nanomaterials, and the Lab's other DOE Office of Science User Facilities, and extreme-scale materials science codes, such as the QuantumESPRESSO open-source computer codes for electronic-structure calculations and materials modeling at the nanoscale. According to D'Imperio, efforts in this area are already underway.
"In the Computational Science Laboratory, we are actively converting a materials science code to run on GPU accelerators. Early results show a 31 percent increase in performance over the CPU [central processing unit]-only version," said D'Imperio.
Brookhaven Lab scientists are also trying to make it easier to develop GPU-accelerated applications.
"At Brookhaven Lab, we are exploring productive, portable approaches to developing application codes that will run on GPUs," said Chapman. "Brookhaven is ideally positioned to demonstrate the benefits of using compiler directives — or source code instructions for the compiler to influence its translation of a program — for GPU application development. Because we have experts who develop programming models and experts who can use those models to develop applications, we can take user feedback into account when experimenting with and evaluating our ideas."
Brookhaven's GPU training program will be expanded to include courses on GPU architecture and programming tools, such as the NVIDIA CUDA parallel computing platform and OpenACC (a parallel programming standard for compiler directives).