NPACI's Teraflops IBM SP Accepted by SDSC

San Diego 12 Jan 00 The 1,152-processor IBM RS/6000 SP system at the San Diego Supercomputer Center (SDSC) was officially accepted by SDSC management December 30 after successfully completing a battery of tests that demonstrated stable operation, good performance, and high throughput. The test results show that the new machine will provide the capability to solve problems in days that typically require weeks, months, or years on smaller machines.

The IBM SP computer, installed for NPACI at SDSC, has a peak speed of one teraflops--a trillion floating-point operations per second--and is the most powerful available to the U.S. academic community for unclassified research.

"This is a milestone achievement in IBM's rapidly evolving partnership with the team at San Diego Supercomputer Center," said Michael J. Henesey worldwide sales and marketing, IBM RS6000 Scientific and Technical Computing. "The significance here is the massive computational capability now available in an unclassified environment. This gives the research community the tool needed for breakthroughs in areas such climate modeling, mapping and modeling the human brain, and genomic research. At IBM, we're proud to be delivering on this commitment and also very focused on developing the next-generation systems that will attempt to satisfy the scientific community's insatiable desire for insight."

The highlight of the cooperative IBM and SDSC effort was the discovery and resolution of a problem in the mapping of memory to cache that led to excessive variation in program run-times. Working together, IBM and SDSC improved the cache management for large systems using Power3 SMP High Nodes. The operating system patch will be included in the next release of AIX.

As part of the acceptance tests, performance of four scientific applications on both the teraflops SP and NPACI's current production SP was compared for various numbers of processors, from one to 128. The applications included AMBER for molecular dynamics; GAMESS for quantum chemistry; and PARTREE and SCF for astrophysics. In all cases, these applications ran 1.11 to 1.92 times faster on the teraflops SP.

One application, SCF, was also run on all 1,152 processors and showed good scaling over a large processor range. Using a fixed problem size per processor, the run-time on 1,152 processors was only 1.5 times slower than the time on two processors.

One further test demonstrated high node throughput. Over four consecutive days in December, the system's 144 compute nodes were in use more than 86 percent of the time, peaking at 95.7 percent usage on Christmas Day.

Also during acceptance testing, the system was unofficially tested as a participant in distributed.net's RC5-64 code-breaking challenge . While participating in force, the machine placed among the top five daily participants for a week .

 

Ad Emmen

[News on Advanced IT]   [Calendar]   [Analysis]   [IT in Medicine]