ACADEMIA

Silicon Graphics today announced its SGI Altix 3000 system has beaten comparable servers from IBM, Cray and Sun on performance tests that comprise the new HPC Challenge Benchmark. The results reveal that SGI Altix delivers as much as eight times the performance of IBM servers and four times the performance of Cray supercomputers in key categories. Results submitted March 2 by SGI show that a 64-processor SGI Altix system bested similarly configured servers from IBM, Cray and Sun on five of the HPC Challenge (HPCC) Benchmark’s list of eight tests. The HPCC benchmark extends the Linpack benchmark used to determine the well-known Top 500 list of the world’s fastest supercomputers. “The most relevant measure of a system’s real-world capability is application performance, which tends to rely on several aspects of the system architecture,” said Jeff Greenwald, senior director of product management and marketing, SGI. “The HPC Challenge Benchmark cuts a wide swath across many metrics, offering a more balanced indication of how a system will deliver for real-world users in real-world conditions. This additional information will provide customers with a well-rounded profile of a system’s capabilities.” The 64-processor SGI Altix not only swept the competing servers on different combinations of five performance metrics, but Altix also beat all of these systems on two HPC Challenge tests: G-PTRANS, which measures the global communication rate achieved during a matrix transpose operation; and G-FFTE, which measures the floating-point performance of one-dimensional Fourier transforms, a common component of scientific computing codes. On floating-point-intensive calculations as measured by G-FFTE, for instance, Altix outperforms the IBM pSeries 655 by more than eight times. And in HPCC’s Random Ring bandwidth test measuring data transfer bandwidth, Altix performs nearly four times faster than the Cray XD1 at 64 processors. Altix also outperformed IBM, Cray and Sun with the fastest minimum system latency, underscoring the Altix system’s superior communication architecture, in tests of non-synchronous “Ping-Pong” data transmissions between nodes. The Altix is the first supercomputer-class system to break the one microsecond barrier with a score of 0.993 on the MPI Ping-Pong latency test as measured by the HPC Challenge benchmark. The Altix result is nearly twice as fast as the nearest competitor’s. SGI Altix demonstrated superior scalability by surpassing all other competitors, no matter the configuration, in three key HPC Challenge tests: G-HPL (High-Performance Linpack), G-PTRANS, and G-Random Access, a key measurement of how well a system can update its memory. The results, submitted on March 29, were achieved using two, 504-processor Altix systems integrated via the high-speed SGI NUMAlink™ interconnect and located at the National Center for Supercomputing Applications. Complete results are available at: http://icl.cs.utk.edu/hpcc/hpcc_results.cgi. The HPC Challenge benchmark suite measures a range of HPC computing functions, including floating-point performance, memory bandwidth, memory access, and communication latency and bandwidth. Unveiled in 2003, the suite was developed under sponsorship of the U.S. Department of Energy, the National Science Foundation, and the DARPA High Productivity Computing Systems program. With its powerful global shared-memory architecture, SGI Altix systems handle large data sets with ease, helping to enable customers to achieve groundbreaking improvements in life sciences, manufacturing, oil and gas exploration, homeland security, earth and environmental sciences research. Leveraging a proven 64-bit Linux environment and powerful Intel Itanium 2 processors, Altix has consistently set numerous records for sheer performance, and for its ability to efficiently run manufacturing, engineering and scientific applications across hundreds of processors in a Linux operating environment. SGI Altix holds the highest result on the STREAM shared memory bandwidth test, and is one of the fastest and most efficient supercomputers in the world as shown by the TOP500 list.