GPU Testbed and Evaluation

GPU Testbed and Evaluation

GPUs are becoming more widely used for computational science applications, and GPU testbeds have been requested at the requirements workshops NERSC has been conducting with its users. In April 2010 NERSC fielded a 48-node Fermi GPU cluster called "dirac" that consists of InfiniBand-connected Nehalem host-nodes. The GPU accelerated cluster consist of 44 nodes containing NVidia Fermi C2050 GPU's, containing 3 Gigabytes of memory each, and 8 NVidia C1060 GPUs. The system includes a complete set of development tools for hybrid computing, including CUDA, OpenCL, and PGI GPGPU-targeted compilers. The Dirac testbed provides us the opportunity to engage with the NERSC user community to answer the following questions:

• What parts of the NERSC workload will benefit from GPU acceleration?
• What portions of the workload see no benefit, or insufficient benefit to justify the investment?
• Do GPUs represent the future of HPC platforms, or a feature that will be beneficial to a subset of the community?

We have collected a broad range of preliminary performance data that support the idea that GPUs in their current configuration will be an important way to augment the effectiveness of NERSC systems for a subset of the user base.