Login   |    ITaP Home > ITaP Newsroom

 

Boilermaker Cluster Challenge team to use powerful, but lower-power supercomputer

Purdue’s 2008 Cluster Challenge team is looking to make its competition green with envy.

That goes not only for trying to win the undergraduate student competition at the SC08 conference in Austin, Texas, in November — the largest supercomputing conference in the world. It goes for the team’s choice of hardware as well.

Purdue is partnering with SiCortex, a Massachusetts company that makes innovative lower-power supercomputers whose processing chips — kind of the brain of a computer — draw as little electricity as a cell phone or small flashlight, which cuts down on heat and should reduce cooling costs in the bargain.

Jeffrey Evans, an assistant professor in the Department of Electrical and Computer Engineering Technology, said a SiCortex machine was an interesting prospect for the Cluster Challenge team for a number of reasons.

The system SiCortex employs to link its processors together for working in concert — the “interconnect fabric” — is unique, for example. SiCortex designed the speedy pipeline in part to work most efficiently with its not-so-speedy processors. The company describes it as an integral part of an overall system that delivers high performance using large numbers of slower, energy-efficient processors.

The Cluster Challenge’s rules also limit the amount of power the entries can draw from two 120-volt, 13-amp circuits, which limits the number of processors that can be employed. Purdue’s cluster had more than 40 processors in 2007, with most competitors falling between 30 and 60, Evans said. The 2008 Boilermaker team may be able to run nearly 1,000 processing cores concurrently in the SiCortex — orders of magnitude above the competition.

“I am confident that the talented members of the Purdue team will take full advantage of our innovative architecture,” said Mark Blessing, vice president of marketing for SiCortex. “More importantly, the Purdue team’s chosen approach to the SC08 Cluster Challenge rises above the event itself, addressing head-on the real-world issue facing data centers today: how to successfully respond to the exponential increase in demand for computing power in the face of severe data center power, cooling and space constraints.”

Evans said working with so many processors — likely the wave of the future in high-performance computing — also promises to be a good experience for the students on the Cluster Challenge team. The students work on their cluster as part of his high-performance computing class — normally graduate-level but open to undergraduates on the team.

The 2007 Purdue team had to assemble its cluster from standard rack-mounted machines. Since the SiCortex comes assembled out of the box — all the hardware in one package and basic software installed — that should give the ’08 team more time to install, configure and refine the benchmarking programs and working applications Cluster Challenge teams must run for the competition, Evans said.

“Hardware issues could come into play if something fails, or something begins misbehaving,” he said.

Preston Smith, a senior UNIX system administrator for Purdue's Rosen Center for Advanced Computing, said the SiCortex machine presents some extra software challenges, however. The system’s processors are atypical for one thing, so applications need to be adjusted to run on it, said Smith, who’s again serving as team leader.

The programs also need to be tuned to take advantage of the large number of processors and fast interconnect fabric SiCortex employs to offset the raw speed disadvantage of its lower-power chips, which run at a mere 500 megahertz.

Limited to six undergraduate students, teams have to apply to participate in the Cluster Challenge. In 2007, six teams participated. Evans said organizers of Purdue’s team, including himself, thought the unusual nature of the SiCortex computer might boost its case for participation in 2008.

According to the 2008 Cluster Challenge Web site, the winning team will be chosen based on workload accomplished, benchmark performance and overall architectural design.

Purdue's 2007 team received a lot of attention for an interactive comic book created to promote the team and its Cluster Challenge entry. In 2008, staff members from ITaP — Purdue's central information technology organization — and the Rosen Center, ITaP's research-computing arm, are creating a Web-based game — kind of a Sim cluster where players choose from different components, including SiCortex nodes, and try to balance their configurations to best handle a series of virtual jobs. Completed jobs earn grant money — the better to buy upgrades. Machines that don’t handle the load well go bust.

In real life, Purdue team members already have started working with likely competition applications. The Rosen Center — ITaP's research-computing arm — installed a top-of-the line SiCortex machine in June to test it from both performance and energy-consumption perspectives. The loaner that the Cluster Challenge team will use is an SC1458, the next step down from the Rosen Center’s SC5832. In both cases, the numbers refer to the maximum number of processors the machines can contain.

So far, the 2008 team includes: Andrew Howard, senior in electrical and computer engineering technology; Alex Younts, sophomore in computer science; David King, senior in electrical and computer engineering technology; and Matthew Sutherlin, sophomore in computer science. All of them work for the Rosen Center. Howard and Younts were on the 2007 team.

Recruitment of additional members for the 2008 team is in progress. Organizers are hoping to recruit students from earth and atmospheric and the life sciences — domains for which some of the scientific programs that must be run in the Cluster Challenge are geared. Evans and Smith said the Rosen Center machine also should give the team a chance to keep refining its work even after the competition SiCortex is shipped to Austin, which has to happen about a month before the Cluster Challenge.

“We’ll still have a machine we can do work on,” Smith said.

Writer: Greg Kline, (765) 494-8167, gkline@purdue.edu

Sources: Jeffrey Evans, (765) 494-7725, jje@purdue.edu
Preston Smith, (765) 494-9729, psmith@purdue.edu

Last updated: Sept. 12, 2008