|The Leading Source for Global News and Information Covering the Ecosystem of High Productivity Computing / November 15, 2005|
HPC stalwart Jack Dongarra admits he still gets excited about going to SC -- even after all these years.
After all, it would be easy for him to be jaded. He's attended the annual industry gathering since the very first in 1988. Yet arguably, in 2005, Dongarra has even more to be excited about -- and especially today.
In the 17th year of SC, he is serving as the co-chair for the first year of the HPC Challenge award competition, the results of which are being revealed today. The goal of the competition, which is sponsored by DARPA's High Productivity Computing Systems Program and HPCwire, is to develop a set of HPC hardware and software capabilities that become de rigeur for the productive use of all HPC systems.
The awards will be presented in two designations. One is based on performance only; while the other spotlights productivity and elegant implementation. The winners will be announced at noon in a Birds of a Feather session.
Victory in the first class is based on the entry (or entries) that delivered the best performance on a base or optimized run as submitted to the HPC Challenge web site. From the suite of choices, the benchmarks judged were: Global HPL, Global RandomAccess, EP STREAM (Triad) per system and Global FFT. The prize is $500 plus a certificate, and up to four winners may be announced.
The results are not subjective; literally, a computer determines the order of finish. So far, more than 80 contributors have dared to measure up to the series of benchmarks, according to Dongarra. Unlike the well-worn Linpack benchmark, which is based solely on speed, the HPC Challenge takes in numerous other attributes to develop an all-around, better-rounded view.
"We're quite happy with the response," he added. "Like Linpack, it was slow at the start but continues to gain momentum." As the deadline goes right up to the last minute, Dongarra said he is maintaining a shroud over the results.
The second award, for most productivity, is based on the most "elegant" implementation of two or more of the HPC Challenge benchmarks with special emphasis being placed on Global HPL, Global RandomAccess, EP STREAM (Triad) per system and Global FFT.
This award, which will pay $1,500 and may be split, will be weighted equally between performance and code elegance, clarity and size. This portion of the HPC Challenge awards could be considered the closest the HPC community gets to hosting its own version of "American Idol."
"It's a beauty contest - almost," Dongarra said, with a chuckle. Or perhaps he meant it is a beauty of a contest. Either way, there is likely to be lots for the panel of 10 judges to review. As the deadline for this award was Oct. 15, the judges have been busy reviewing and assessing what turned out to be 10 entrants.
Co-chair Dongarra is one of the judges - and be assured he's no Paula Abdul, Simon Cowell or Randy Jackson. He's been huddling with his associates - co-chair Jeremy Kepner MIT Lincoln Lab; David Bailey, LBNL NERSC; David Koester, MITRE; Bob Lucas, ISI; Rusty Lusk, Argonne National Lab; Piotr Luszczek, University of Tennessee; John McCalpin, IBM Austin; Rolf Rabenseifner, HLRS, Stuttgart; and Daisuke Takahashi, University of Tsukuba. - to determine the three finalists. They will take to the spotlight at the awards session.
While Dongarra said he is "delighted" and "enthusiastic" about the response and the prospects, he also can't help but add that he believes next year's competition will be even better. "We're setting the right tone," he added. "We're expecting an ever-stronger collection of competitors."
HPC Challenge's test suite is designed to provide an overview of the major strengths and weaknesses of HPC systems. With this set of benchmarks, organizations looking to purchase HPC systems will have a better understanding of system capabilities.
"The HPC Challenge benchmark suite is a better indicator than any single test of how an HPC system will perform across a spectrum of real-world applications," said Dongarra. "HPC is much more complicated today then when Linpack was developed 30 years. The HPC Challenge awards are helping set new standards for benchmarking methodology and result-reporting with a control database/repository for both the benchmarks and the results."
The goal is for the suite to take no more than twice as long as Linpack to run, Dongarra said. Future goals are to reduce execution time while expanding the set to include elements such as sparse matrix operations and developing machine signatures.
SC05 represents a homecoming of sorts for Dongarra, one filled with many memories. For example, attendees need to know SC was not a success from the start, he said. While the early conferences were focused on the technical aspects of HPC, it wasn't until the early 1990s that participating vendors started to gain in numbers and prominence. It was then that SC as a whole started to attract broader attention.
"The vendors always put on a nice show," he said. "They're always trying to outdo each other. It's fun to look at the nice, shiny machines. After all, supercomputers are the Ferraris in any stable."
Yet he is quick to add that assuring SC's viable future means there must be continued emphasis on the "other side" of supercomputing, namely software and algorithms. "Hardware may be attraction, but software and algorithms makes things go and grow," Dongarra added.
For each SC gathering some theme emerges, and Dongarra said he's curious what this year will bring. While in the past Linux and clusters have enjoyed their time in the spotlight, the odds-on money this year is on Microsoft, especially with Bill Gates as the keynoter. "It will be interesting to see how his message resonates," Dongarra said.
Dongarra said he hopes Gates' address will clarify Microsoft's interest in HPC, but all that is certain is that it will set the tone for the Seattle gala. "In some sense, he'll be talking to a hostile audience," he added. "The HPC culture is very different than Microsoft's."
One place where perhaps it's not that different is with Craig Mundie, Microsoft's senior vice president and CTO, an HPC cronie of Dongarra's from SC's early days. "I knew Craig at Alliant [Computer Systems Corp., a company that developed massively parallel supercomputers]," Dongarra said. "It's amazing how far things have come since then. We'll see how he and Tony Hey [another high profile HPC veteran who's joined Microsoft] bring about change within the company."
Nevertheless, Dongarra added that some of his favorite parts of SC are seeing the displays and exhibits that the various labs and organizations put forth. He said he is especially encouraged that industry and commercial ventures are working as hard on HPC matters as the labs.
University of Tenneesee has 10 people at SC from Dongarra's group at the Innovative Computing Laboratory. They will be spread out over an number of exhibit booths, including Rice University, ORNL, Los Alamos and, even, Microsoft.
Even when this year's SC05 is a memory, it remains an exciting time in HPC," Dongarra said. "The advent of multi-core systems bares close watching as does the international competition to make petaflop computing a reality sooner than later. Both will put their fingerprint on the future of our industry."