The University of Maine uses high-performance computing to model the impact of global warming on ice sheets and...
glaciers, so it seemed counterintuitive to use a resource-hungry system to perform those calculations.
When it comes to high-performance computing (HPC) systems, real estate is a key concern. One question "when you talk about getting a supercomputer system, is where do I put the thing?" said Phil Dickens, an assistant professor of computer science at the school. Both physical space and power capacity at the supercomputing center were filling up.
And, as in enterprise data centers, power and cooling infrastructure is an issue. The High Performance Computing Challenge, a group of HPC-specific benchmark applications, wants to integrate energy efficiency into its metrics. One HPC vendor, SiCortex, recently released an online calculator comparing the energy efficiency of various HPC machines.
Jack Dongarra, a University of Tennessee professor who launched the HPC Challenge, said that while discussions are under way, there's little that is concrete to deliver yet.
"We have to approach it very carefully," he said. "We're trying to determine what metric is the right one."
Power consumption benchmarking evolving
The group would have to determine, for example, what to measure to get the power readings and what would be used to measure it. Dongarra and others have started to collect power numbers for Linpack, which measures a system's floating-point computing power. For that, they're measuring the power consumption of the CPU, memory and switches that connect parallel processors.
But there are gaps. Linpack doesn't, for example, factor in the power consumed by disks that are connected to the machine or the power for lights in the room, he said.
Dongarra isn't entering unchartered territory here. The Standard Performance Evaluation Corp. developed a benchmark for measuring server power compared to performance, but it currently measures only the performance of server-side Java.
SiCortex uses the HPC Challenge benchmarks and matches them up with the power consumption of about a dozen HPC machines, including one from SiCortex. Under most scenarios, according to the SiCortex calculator, SiCortex comes out as more efficient. Mark Blessing, a key developer of the calculator, said he wouldn't mind turning it over to the HPC Challenge.
"What I'd like to do is pass this off to an independent body," he said. "One could argue that this is very convenient that SiCortex has created the index and they come out as No. 1. It would be better for everyone if there was a third party that owned this index."
SiCortex and Dongarra haven't discussed this possibility, but Dongarra said the company would have a seat at the table.
SiCortex "can participate along with other interested vendors and parties," he said.
Dickens said the University of Maine picked a SiCortex server -- which has 648 processor cores -- in part because of its efficiency. He recalled during his time at the Illinois Institute of Technology, which bought a Sun Microsystems HPC server. After the purchase, the organization realized it had no facility that could accommodate the system's power needs and had to build a special room for it.
Dickens admitted that the primary motivation for getting a more energy-efficient HPC machine was so he could maintain it in his own lab.
"I wanted to be able to keep supercomputing local to my lab so my students and I could work on it directly if we needed to," he said. "If we put it in the data center, we would have lost administrative control. I would rather have the opportunity to keep it in my lab, because that's where it's primarily being used."
Let us know what you think about the story; email Mark Fontecchio, News Writer.