 |
|
|
|
|
|
|
|
|
|
|
2004 NEWS AND HIGHLIGHTS
10.26.04 - NASA's Columbia Supercomputer Tops Charts
NASA's 10,240-processor Columbia supercomputer has produced a sustained performance of 42.7 teraflops on 16 of the 20 installed nodes, achieving an 88 percent efficiency rating - ranking it as the fastest known supercomputer in the world today. This large supercomputer of twenty SGI Altix 512-processor systems is located at NASA Ames Research Center, Moffett Field, Calif.
While planning the installation of the Columbia system, engineers designed their approach to measuring its performance using the Linpack benchmarks - the accepted performance index in the high-performance computing industry. On 16 nodes, Columbia has a peak (theoretical) speed of 49.15 teraflops, but because no system can reach peak speed - due to inherent interprocessor communication inefficiencies - the trick is to get the most efficient performance possible in order to produce top results. Typically, as the number of processors increases, the efficiency (as a percentage of peak) decreases further. So, with NASA installing the largest number of processors of any system on the current Top 500 list of supercomputers, efficiency is very important for Columbia's performance.
The top spot on the TOP500 List for several years running has gone to Japan's Earth Simulator, with a currently posted speed of 35.86 teraflops. Initially, the Columbia team was confident it could achieve enough efficiency to surpass that number - possibly reaching 40 or more teraflops. The team crafted several approaches to adapting the Linpack benchmarks, and worked to analyze and remove performance bottlenecks in the code and system interface routines.
Their approach was to first maximize efficiency on a few processors, then run the benchmark on more processors and analyze any drop-off. At one point, the efficiency on four 512-processor systems was 76 percent, but a subsequent test on eight systems dropped to 63 percent - too big a drop to be acceptable to the team. Another test on four systems, run in mid-October, produced an 82 percent efficiency rating. Even more encouraging, the subsequent eight-system test showed a decrease of only two percentage points (to 80 percent). That test achieved 19.6 teraflops, just missing the mark set by the second-fastest system in the world in June 2004, the 19.94 teraflop mark posted by a Lawrence Livermore National Laboratory system known as "Thunder."
In the days before submission of Columbia's final official Linpack results to the TOP500 organization, the team aggressively tackled the efficiency challenge and improved the number yet again. The system count was expanded to 16 in a later run that yielded 42.7 teraflops and is reported here. With additional adjustments, the final number on all 20 systems (totaling 10,240 processors) was submitted on October 24, 2004. Final results will be published on Monday, November 8.
|
|
|
|
NEWS ARCHIVE
|
|
|
