By: Michael Feldman

In an industry as forward-leaning as high performance computing, the focus on exascale and buying machines with the maximum amount of FLOPS hardware can be a distraction. The average HPC user is just looking to find the best performance possible for their applications with the hardware at hand. And in more cases than we would like to think, sometimes that hardware is just a personal computer.

Three recent news items highlight how a set of users, armed only with PCs, were able to achieve supercomputing results through innovative software. Their success illustrates how innovative algorithms can overcome a lack of hardware, especially when those algorithms are carefully designed to exploit the hardware that is available. From another perspective, these stories also highlight how poorly many supercomputers are utilized under the burden of legacy software, or in some cases, just plain conventional wisdom. As Moore’s Law winds down, such innovation is going to become even more precious.

Quantum mechanics on the desktop

The first story along these lines involves a team of physicists at Moscow State University that used a PC to outrun a Blue Gene supercomputer. The Russian researchers made use of the GPU in the personal computer to solve a complex series of quantum mechanical equations in 15 mins, compared to a run-time of two to three days on the Blue Gene.

The supercomputer in question is a bit of a mystery, since the story implies it was JUGENE, the 220-teraflop Blue Gene/P system that was decommissioned four years ago at Forschungszentrum Jülich. The system that replaced it, JUQUEEN, is a 5.8 petaflop Blue Gene/Q. In either case, the PC the researchers used was far less powerful. Even with the on-board GPU, which they estimated cost $300 to $500, the PC couldn’t have delivered any more than a few peak teraflops.

The equations they were interested in involved the scattering of multiple quantum particles. To achieve realistic interactions between these particles required demanding calculations using large matrices (hundreds of millions of elements), which typically requires the computation power of a supercomputer. The researchers figured out a way to do the equivalent calculations in a fraction of the time on a single consumer-grade GPU.

In the Moscow State news story, professor of theoretical physics Vladimir Kukulin describes the implication of the PC effort (translated):

“This work, in our opinion, opens up completely new ways to analyze nuclear resonance chemical reactions,” said Kukulin. “It can also be very useful for solving a large number of computing tasks in plasma physics, electrodynamics, geophysics, medicine and many other areas of science. We want to organize training courses, where researchers from diverse scientific areas at peripheral universities that do not have access to supercomputers, could learn to do their work on their personal computers as we have done.”

Accelerated graph processing with GPUs and SSDs

The second story takes place at the Daegu Gyeongbuk Institute of Science & Technology in South Korea. On July 7, researchers there announced that have developed software for a personal computer that is able to process a graph with hundreds of billions of edges. The software, called GStream 2.0, uses a PC with two GPUs and two PCIe-connected SSD cards to achieve two giga-traversed edges per second (GTEPS). Graph processing is applicable to a large number of application domains, including bioinformatics, social networking, business intelligence, online retail, and logistics, to name a few.

According to an article published at Business Korea the software developed at the Daegu Gyeongbuk institute had far better performance than a well-known and highly scalable graph-processing framework (GraphLab) developed at in the US :

“The GraphLab of Carnegie Mellon University, which is the most advanced big data analyzer at this point in time, requires 1,400 seconds to process graph data having up to 32 billion edges with a supercomputer equipped with 480 CPU cores, a 2 TB memory and a 5 GB network,” the institute explained, adding, “However, GStream 2.0 can process data of 32 billion edges in 500 seconds and process data of up to 256 billion edges by storing large-scale graph data in the PCI-e SSDs, performing asynchronous streaming from the SSDs to the GPUs and using thousands of computing cores in the GPUs.”

Details of the GStream work were announced at the 2016 ACM SIGMOD/PODS Conference that took place in San Francisco on June 28.

AI outguns fighter pilots

The third story involves how “genetic fuzzy logic” turned a $500 PC into an ace fighter pilot. The application is an artificial intelligence program, known as ALPHA, a research tool than simulates air combat. As it turns out, ALPHA can consistently beat premier fighter pilots in simulated dogfights. The software was developed by Psibernetix, Inc., which was founded by Nick Ernest, a 2015 doctoral graduate from the University of Cincinatti (UC). Ernest is president and CEO of the company.

The software was recently assessed by retired United States Air Force Colonel Gene Lee, who could not score a kill against the application, despite repeated attempts. In each of his interactions with ALPHA, he was shot out of the air. Other expert fighter pilots fared no better in subsequent engagements. In fact, ALPHA bested its human competition even when it was artificially handicapped with degraded weaponry and sensors, as well as reduced hardware speed.

From the UC news story:

Lee, who has been flying in simulators against AI opponents since the early 1980s, said of that first encounter against ALPHA, “I was surprised at how aware and reactive it was. It seemed to be aware of my intentions and reacting instantly to my changes in flight and my missile deployment. It knew how to defeat the shot I was taking. It moved instantly between defensive and offensive actions as needed.”

Up until now, such programs could not consistently match real pilots, and the general feeling was that you would need a ground-based supercomputer to simulate the math-heavy scenarios necessary to engage live pilots. ALPHA though, works more like a real human, breaking down tactics, firing, and evasion maneuvers into a decision tree that is tested and refined as it runs through simulated dogfights. The genetic aspect of the software ensures the bad decisions are discarded and the good decisions are retained.

The ramifications of such a system are rather apparent, and the Air Force is continuing to test ALPHA. Although it remains a research tool, no doubt this software (or a descendent of it) will be inserted into a real fighter system, either to augment the pilot’s abilities, or perhaps one day, to replace him or her entirely.