Back in 1995, Thomas Sterling, along with academic comrades Paul Messina and Paul Smith collaborated on a forward-looking tome called, Enabling Technologies for Petaflops Computing, which explored a far-flung future that has finally arrived.
During a chat with Sterling this morning, the topic of the book cropped up, in part because the Indiana University professor (and notable luminary in Beowulf and thought leadership circles) has been biding his time until he could have a petaflopper to call his own—or at least one in cozy reaching distance at IU.
Later this month, Indiana University will formally introduce the successor to the Big Red system, the aptly-named, Big Red II. The Cray-crafted and tuned system is 25 times faster than its baby brother (the 4100-core original Big Red from 2006) and sports some notable improvements across its 1,020 nodes. With some Kepler spice and the snappy Gemini interconnect to push its peak one teraflop performance to an expected top 30 range for June’s list, the system will aim its big guns at true “big data” problems.
IU thinks some theoretical work on the “little” 210,000-core Big Red II can unleash some optimization dragons for systems like Titan and Blue Waters to ride, at least in theory. With a common, mixed-up architecture that is either homogeneous or heterogeneous, depending on how it’s feeling for particular applications, there are significant opportunities to fine-tune core operations to take best advantage of any configuration.
What’s needed for such systems is an execution model that can self-adapt on non-uniform systems. And since it’s the same big idea on a smaller canvas (than Titan and Blue Waters), Sterling said he has hope that tweaking the ParalleX execution model could yield some big returns.
Although Big Red II is far smaller than Titan or Blue Waters, it’s the same technology, architecture and software environment than its big hybrid peers—and this triad of features is likely to be at the top of the trend list for new systems in the coming years.
On that note, Sterling was in the midst of a trip this week to Sandia National Lab (both part of the XPRESS project) to talk about Big Red II and these experimental pieces of a potential programming model and runtime system that might play nicely with such unique, hybrid supers. There are obvious architectural similarities between big boy systems like Titan and Blue Waters, and the Hoosiers have hopes that Big Red II can help create a playbook for similar system operators to score maximum performance, scalability and of course, efficiency out of their supers.
There are a few things that Sterling and his many counterparts expect from Big Red, including counting on its iron hand to help shove some new ideas about using these trend-setting systems efficiently and at massive scale. “The trick is to address the challenges of asynchrony and compensate for that uncertainty. That’s what our runtime system will demonstrate, or so we hope—at least for some applications on Cray systems like the one we have now.”
The other proof pudding they hope to whip up at IU relates to tackling new classes of data-intensive problems that are memory-bound, exploit locality and move beyond traditional numerically-oriented approaches. We need to move back toward an older concept that never enjoyed its day in the sun, argues Sterling—we need to think back to the promise of symbolic computing and how systems like Big Red and others can turn the standard model on end. Overused buzzword or not, this is all about “big data,” a topic that can’t be shoved under the HPC rug as a trend when it’s already influencing the shift toward Titan-esque systems.
On the big data front, Sterling and his team at IU, under the university’s VP of IT and CIO, Brad Wheeler, set about driving stakes in new supercomputing ground, the emphasis was on pushing performance. But just as important as floating point was the need to make critical decisions about memory. He pointed to a number of people at IU that helped make core decisions, and also to Bill Blake from Cray who helped them refine and tweak to perfection.
Sterling notes that in terms of system design (full specs here), the choice to snap in AMD Interlagos and Abu Dhabi processors wasn’t an Intel versus AMD decision, it was “purely generational” for this pre-Intel Cray design. The Kepler cores were a key investment since, as Sterling described, there are “many science codes that, with sufficient refactoring, could take advantage of GPUS.” He said, “It doesn’t mean it’s easy, but under the right circumstances, we’re looking at a 5x to 10x speedup.” This is going to boost their production capabilities to new levels, he notes, and is aided by the fact that Geoffrey Fox and other critical folks at IU were pushing fresh envelopes on the GPU and parallel computing fronts before this Kepler-sporting system landed on their datacenter doorstep to begin with.
In terms of extracting ultra performance on a system designed with data-intensive problems in mind, Sterling said there is a balance between FLOPS and big data considerations, including laying down memory foundations and keeping data and compute at the end of the same stick. “The importance of FLOPS will continue to grow,” he notes, “but the importance of big data and knowledge analytics will grow faster.”
“It’s the symbolic graph structures and future architectures we need to make computers understand its data, not just manipulate…right now, that’s a big constraint. The work on Big Red II will let us move closer to knowledge, knowledge management and most importantly, machine understanding of knowledge and that will change how we pursue problems in climate change, drug design, very complex system design as in large aircraft that doesn’t take decades–but weeks or less.”
“Computers manipulate data and take action on it. Human beings manipulate knowledge and make that actionable and there’s a gap between data and knowledge. We don’t want big data—we want smart knowledge. And that’s where the research has to be and that’s how the usage patterns of big computers need to reflect.”
Whether or not “big data” is just a new term for wrapping HPC around a new architecture that’s optimized for certain applications or problems, it’s a term that has staying power, even for the self-confessed “hype hater” Sterling. But at the end of the day, for the guy who helped write the book on what petascale systems would be made of, Sterling says that “here so many years later, I’m finally getting my hands on [a system]. It’s closure in my career and truly an exciting time.”