INDUSTRY

GraphStream Incorporated announced the deployment of one of the world's most powerful graphics supercomputers, at Lawrence Livermore National Laboratory (LLNL), Livermore, CA, USA. This GraphStream terascale computer system, named "Gauss", features 256 NVIDIA Quadro FX 4500 Graphics Processing Units (GPUs) operating in parallel. Gauss is the eighth and largest in a series of increasingly powerful visual computing systems that GraphStream has delivered since 2003 to the U.S. Department of Energy's National Nuclear Security Administration, for the NNSA Advanced Simulation and Computing (ASC) program. Gauss is the visualization system for BlueGene/L ("BG/L"), which according to the traditional LINPACK performance metric is currently the world's most powerful computer system, at 280.6 trillion operations per second. BG/L has 131,072 CPUs operating in parallel, and is being used to simulate a variety of physical phenomena on a scale never before possible, enabling frontier work in theoretical and experimental science that supports the NNSA mission. The role of the Gauss system is to augment BG/L with visualization capabilities that BG/L does not natively have. Gauss will be used initially for visual exploration and interpretation of simulation results produced by BG/L. Future plans include work on interactive use of Gauss to follow and even guide certain in-progress computations on BG/L, to help optimize the use of BG/L's capabilities. According to Steve Louis, LLNL Subprogram Lead for ASC Computational Systems and Software Environment, "We realized that our current visualization solutions would be inadequate to handle the massive datasets that would be generated by BG/L codes. At early meetings, researchers commented that BG/L would produce output so massive that it could only be handled by BG/L itself. Gauss provides a solution". Gauss is designed to be able to completely fill its 3-terabyte main memory with simulation output data from BG/L in less than three minutes. Gauss is connected to BG/L via a shared 900-terabyte parallel storage system. Gauss has 256 separate Gigabit Ethernet connections to the storage system, supporting sustained storage throughput of over 17 gigabytes/sec in each direction. Gauss is currently one of the world's largest mainstream-architecture clustered computer systems to feature a dedicated, direct data link from every single node to shared storage. Highlights of the Gauss hardware configuration include: - 256 NVIDIA Quadro FX 4500 PCI Express GPUs - 42.3 teraFLOP/sec aggregate peak pixel-shader throughput (32-bit) - 8.6 terabytes/sec aggregate peak GPU memory bandwidth - 128 gigabytes aggregate GPU memory capacity - 512 AMD Opteron 2.4 GHz 64-bit CPUs - 3 terabytes aggregate main memory capacity - Voltaire InfiniBand cluster interconnect - 264-port network switch with 2+2 terabits/sec aggregate throughput - PCI Express host channel adapters - 256 separate Gigabit Ethernet links to external 900-terabyte parallel storage system - 18 rackmount enclosures - 225,000 watt three-phase power feed Highlights of the Gauss software configuration include: - Linux/GNU operating system based on Red Hat Enterprise Linux v4.0: LLNL "CHAOS" - OpenIB InfiniBand host software stack - Visualization applications and toolkits with support for cluster-based distributed graphics rendering - CEI EnSight - Kitware ParaView - LLNL VisIt - Chromium LLNL and GraphStream staff were able to fully install the Gauss system and complete initial LINPACK runs at over 1.8 trillion operations per second, all within two weeks after delivery of the Gauss hardware to LLNL.