Pushing the Petaflop Envelope

Todays announcement of the Top 500 supercomputers in the world is notable, not just for the current popularity of quad-core chips and Gigabit Ethernet, but for the sheer speed of the world's fastest system.

The U.S. Department of Energy’s new Roadrunner machine has basically blown the doors off the Top 500 ratings, touting speeds of more than a Petaflop (over 1,000 teraflops, or one thousand trillion calculations per second).

Contrast this with the previous No. 1 system, the IBM BlueGene/L at Lawrence Livermore National Lab, which is capable of 478.2 Tflops.

BlueGene/L had topped the list since November 2004, but Los Alamos’s Roadrunner now offers more than double the performance of its predecessor.

“The Roadrunner system is a bit of an anomaly,” says Jim Kasdorf, director of special projects at the Pittsburgh Supercomputing Center, explaining that Roadrunner’s phenomenal speed comes at a price. “It’s a very difficult system to program - you have to be a hero to make all these pieces work together.”These “pieces” are an unusual mix of different processors, according to Kasdorf, including AMD’s Opterons, IBM’s PowerPC’s, and even a hybrid version of the "cell engine" chips developed for Sony’s Playstation 3.

”On Roadrunner you have to program for three computers; you have to program for the Opterons, the PowerPC cores, and then the cell cores,” says Kasdorf. “There’s three different architectures with three different characteristics.”

Despite taking the roof off the Top 500 list, Kasdorf feels that other sites will soon catch up with Roadrunner, namely the fifth-placed "Jaguar" system at the Oak Ridge National Lab, and the IBM offering at Lawrence Livermore.

”The Oak Ridge system will probably exceed a Petaflop by next year, and the Blue Gene people plan to double the [system’s] size,” he says, adding that both of these systems also use less complex processor architectures.

For enterprises with limited programming resources, it is these systems, as opposed to Roadrunner, which may provide the real benchmark for supercomputing.Have a comment on this story? Please click "Discuss" below. If you'd like to contact Byte and Switch's editors directly, send us a message.