HPC Users Mull Licensing Issues

Software licensing is a major hurdle in the development of ultra-high end supercomputers, warned execs at this weeks High Performance Computing User Forum at the Oak Ridge National Lab.

A number of organizations are currently shifting to what are described as "petascale" machines -- supercomputers with the capacity to perform a thousand trillion operations a second (a petaflop) (see Oak Ridge Plans Petaflop Supercomputer). But, in order to build these monsters, users are boosting their processor capacities, potentially into the tens of thousands, creating a software licensing headache.

Many software packages are priced on a per-processor basis, which poses a real dilemma, according to one supercomputing site manager from the U.S. government sector, who asked not to be named. “It’s a problem,” he said. “We’re reaching a point where software becomes much more costly than hardware.”

The exec is concerned about the possible impact on his internal users, particularly when he shifts applications to a larger number of processors. “What we don’t want [is for] users to crash because there’s no license (in place).”

Andrew Jones, high-end computing development manager at Manchester University, echoed these fears. “The current assumption is that petaflop [supercomputers] will be thousands of processors,” he says. “Your software licensing costs might be a few thousand dollars now, but they will be millions of dollars then.”The software licensing issue is even approaching NASA’s radar. “I can see that in some of the large software analysis tools it could be a problem down the line,” said Tsengdar Lee, the agency’s high performance computing program manager. However, Lee added that licensing is not such a “big deal” for NASA at the moment because most of its software is homegrown.

Alternatives to processor-based pricing cited by attendees included "per site" and even "per end user" pricing. However, Jones warned that organizations should be careful not to push vendors too far, particularly given their reliance on specialist software for the likes of debugging. “Something towards per user licensing could be an answer, but you do have to take into account the economics of the software vendor,” he said. "If you cut their margins too much they won’t exist."

To make matters even more complicated, processor manufacturers Advanced Micro Devices (NYSE: AMD) and Intel Corp. (Nasdaq: INTC) are moving toward chips offering two or more cores. Intel this week announced its swansong single core Xeon processor (see Intel Pushes Xeon ).

The idea is that, by combining two processors on a single chip, users gain both a performance boost and a reduction in power consumption. (See Sun Spawns Galaxy, Penguin Offers Opteron and Dell Adds Multi-Core .) But this makes it even more confusing for the folks running supercomputers. “Some software vendors are creating their cost models based on the chip and some are creating their software models based on the core,” said the U.S government supercomputing site manager.

Ultimately, some of these issues will be resolved in the course of business. “In a few years, on your standard PC or laptop, you may have a chip with eight cores, and obviously you’re not going to want to pay for an eight processor license,” said Jones. “Personally, I believe that the mass market will solve the problem for us.”Some users are taking the licensing issue into their own hands. “We have gone back to a few vendors and asked for site licenses or per node licenses, which could be six, ten, or however many processors,” said the U.S supercomputing site manager. “We created our own terms.”

Supercomputing is not the only blot on the software landscape. Enterprises, for example, have already voiced their concerns to NDCF about the impact of grid computing on current software contracts. (See Software Licensing Gridlock.)

— James Rogers, Site Editor, Next-Gen Data Center Forum