RETAIL

Two years after Quadrics helped build the fastest Linux cluster of the day, called MCR (www.llnl.gov/linux/mcr/), Quadrics and Lawrence Livermore National Laboratory (LLNL) have done it again. This time the system, called Thunder, is not only the fastest Linux cluster in the world, but also the fastest computer system in the US and the #2 system in the world (see www.top500.org), surpassing the previous #2 system (ASCI-Q), also based on Quadrics QsNet. Thunder uses the new Quadrics QsNetII 64-bit technology to connect 1024 Intel Tiger Quad Itanium® II Processor servers to deliver 19.94 TeraFLOPs/s on parallel Linpack. Peak performance of the system is 22.9 TeraFLOPs/s. The level of efficiency, at 87%, is unrivalled by other clusters and exceeds that of many vector supercomputers. Key factors in achieving these levels of performance and efficiency are the floating-point performance of Itanium II processor and the ultra low latency of the QsNetII interconnect. "Thunder represents the next generation of Linux cluster for scientific simulation," remarked Mark Seager, Livermore's Assistant Department Head for Advanced Technology. "Our applications are seeing a 50% to 400% speed up over our Intel® Xeon® based clusters." The new cluster enhances the existing LLNL simulation environment and integrates with three other existing Xeon® - QsNet based commodity clusters. This combined 32-bit and 64-bit simulation environment now provides a total of 36 TeraFLOPs/s of Linpack Performance. Data sharing between the compute and visualization clusters is enabled by a single shared Lustre filesystem that runs both QsNet and Ethernet networks. The LLNL simulation environment is a major achievement in supercomputing and an approach now being followed by similar organizations. 'Thunder is a milestone in the use of commodity components in supercomputing' declared Duncan Roweth, head of software R&D at Quadrics, 'providing not only massive computational power but also the I/O resources and operational environment necessary visualize and interrogate numerical simulations in situ'. Quadrics has collaborated with LLNL over the past 4 years to integrate its technologies in large-scale production Linux clusters that satisfy the Laboratory's demanding requirements in terms of performance, scalability and reliability. This collaboration has included system design, software development, integration and commissioning, code optimization, 24-hour mission critical support and further collaboration with LLNL third parties to optimize the whole simulation environment. "This confirms our focus on high performance computing to develop our technology and motivate us to improve it on an ongoing basis; we are already working on the next generation of QsNet" concluded Cristoforo Romanelli, CEO of Quadrics. http://www.quadrics.com About Quadrics Quadrics (www.quadrics.com) is a leading supplier and developer of high performance networking products and software for massively parallel systems. Quadrics has its HQ in UK and local presence in US and in Italy. Quadrics Ltd. is part of the Finmeccanica group (www.finmeccanica.com). The world's leading vendors and system integrators as well as major Supercomputing Centres such as Pittsburgh Supercomputer Center, Lawrence Livermore National Laboratory, Los Alamos National Laboratory, Pacific Northwest National Laboratory and Sandia National Laboratory have selected Quadrics. Quadrics Ltd is part of the Finmeccanica Group (www.finmeccanica.com) QsNetII is the latest interconnect technology released by Quadrics. With high bandwidth and ultra-low latency QsNet, is a crucial component for building scalable systems. Quadrics software ensures that productions clusters spanning thousands of processors can be programmed, used and administered effectively. The QsNetII PCI-X 64-bit host adaptor delivers MPI latencies of 1.5-2.5 microseconds (depending upon host type) and bandwidths of up to 910 Mbytes/sec/adapter. Multiple adaptors can be used to provide more bandwidth and high availability where required. (www.quadrics.com/whitepaper) Thunder (www.llnl.gov/linux/thunder/) is a highly integrated, well-balanced capability compute resource with 1,024 nodes. The cluster is capable of supporting a complex workload consisting of medium (1,024-2,048) and large (2,048-4,004) MPI task parallel jobs for unclassified Multi-programmatic & Institutional Computing (M&IC) simulations.