One of the prominent themes of this week’s High Performance Computer and Communications Council (HPCC) Conference revolved around the question of why many users with a need for HPC are still resistant to adopting the technology. John West, the Director of the DoD’s High Performance Computing Modernization Program, and the organizer of this years HPCC program, talked at length about this particular phenomenon in his conference kickoff presentation on Monday morning, titled “What’s Missing From HPC?”
There are plenty of drivers for bringing more users into the HPC fold, from the practical motivations of hardware and software vendors, who would like to move more product, to the more altruistic interests of the HPC’ers, who want to expand the community, and the government, who sees the technology as a way to improve industry competitiveness and create jobs.
The problem has been coined with the term “Missing Middle,” referring to the absence of HPC users between the topmost supercomputing practitioners at the national labs and those doing technical computing via MATLAB and CAE/CAD tools on personal computers and workstations. Many of these missing users are in the manufacturing sector, but they also inhabit more established HPC enclaves such as defense, life sciences and finance.
All things being equal, one would expect there to be a continuum of HPC practitioners from the bottom to the top, with a pyramidal distribution that reflected application level and complexity. But that’s not the case. While there are millions of people doing technical computing on the desktop and perhaps tens of thousands of supercomputing users at the top, the middle ground has a lot more in common with supercomputing group population-wise.
For these types of users, system size is in the “closet cluster” realm, on up to maybe a few racks of servers. In fact, this represents the average size of HPC systems for people who are not doing “big science”-type supercomputing. In that sense, the middle is not so much missing, as grossly underpopulated.
According to West, most people using supercomputing today came to the technology because they didn’t have of choice. Astrophysicists couldn’t create two galaxies in a lab and watch them collide; they had to simulate the whole thing digitally. Since supercomputing practitioners are more or less a captive audience, in many cases the tools that are available are not all that great. They often rely on specialized compilers and development environments, legacy programming languages, command line interfaces, and obscure Linux commands. Meanwhile, the larger computing community has moved on to pretty GUIs and a rich ecosystem of more intuitive tools.
That by itself has made the jump from desktop computing to clusters a painful one. But as West mentioned later, there are a number of new interfaces being developed (usually specialized for individual applications or application domains) that are much more user friendly.
Another barrier to moving up the computing food chain is expensive hardware and software. “We’re mostly over this one,” West noted. “It’s not so expensive anymore, although if you’re talking about small manufacturers or small businesses, $50,000 is still real money.”
Then there’s the management of the cluster. If you don’t have an IT admin in your organization or if you do have one, but they are used to managing only Windows PCs, then the decision to add an HPC system is a lot more difficult. The choices (ignoring the cloud option) are to either hire a cluster administration or convince IT that they have to come up to speed on the technology.
Compounding that problem is the lack of a complete tool chain — the various codes, libraries and development tools that are needed to create the models and other user applications. Since these are often missing even at the high-end of HPC, their absence for entry-level users should come as no particular surprise. The solution here, said West, is non-trivial, and comes down to filling in those software gaps on a case-by-case basis.
One barrier that is not discussed as much is the lack of expertise and social support for HPC systems. For a workplace with no previous experience using the technology, the initial user is often the loneliest guy or gal in the building, with no one to ask questions of when something goes wrong. “This is a skills problem, at its heart,” West said, adding that what is needed is a lot more people in industry who are at least computational literate and then a smaller number of computational professionals.
Related to the cultural and technical unfamiliarity with high performance computing is the fact that most non-HPC users already have something that works today. It might not be the fastest or slickest solution, but it serves its purpose. A typical desktop workflow might mean starting up a job on a PC before going home for the evening, and then getting the results back the following morning. If that doesn’t sound like an optimal workflow, at least it’s comfortable one.
The opportunity for HPC arises when the pace of desktop computation isn’t fast enough, either because it’s limiting product innovation, it’s causing deadlines to be missed, or both. It’s been estimated that maybe half the 280,000 or so US manufacturers fall into that category. And given that only 4 to 8 percent of those manufacturers currently employ HPC, the opportunity does indeed appear to loom large.
Of course, the underlying assumption here is that Moore’s Law is not sufficient for technical computing at any level. In other words, desktop systems that are regularly replaced with ones based on faster chips would not be powerful enough to keep up with an escalating demand for better application fidelity or more complex computations. While it’s true that desktop machines of today have as much computational power as the top supercomputers of 15 years ago, that’s still too slow for traditional supercomputing applications. To escape the more limited progression of Moore’s Law, HPC has turned to multiplying those processors across ever-larger clusters. But is Moore’s Law too slow for a typical CAE/CAD user?
Since the cluster is the lens through which HPC practitioners look at computing problems, it’s no surprise they believe the technology is appropriate for most, if not all, technical computing problems. In his conference presentation, West acknowledged that mindset, pointing out that people in this community tend to view HPC as a “unalloyed good,” which can be applied to good effect nearly everywhere. “I think that’s not always helpful,” admitted West.
Intersect360 Research CEO Addison Snell, who has been following the HPC-manufacturing gap for the past couple of years, remarked that not every company is going to need the technology. According to him, the easiest converts will be those manufacturers who need to create innovative products, rather than just standard widgets that fit into a supply chain.
At the conference this week, their were three examples of such companies that made a successful leap to HPC: Simpson Strong Tie, which employs high fidelity FEA models for its structural engineering designs; Accio Energy, a wind energy start-up that is using HPC to design electrohydrodynamic (EHD) wind energy technology (no moving parts); and Intelligent Light, a software company that used its CFD software to help design a game-changing bicycle racing wheel for manufacturer Zipp Speed Weaponry. In all cases, these fit into the high-innovation-need category, where the engineering, by necessity, required a lot of design iterations.
Intel’s Bill Feiereisen got the last word at the conference with his HPC in Manufacturing presentation on Wednesday afternoon. He brought up the idea of creating a pilot project that offers a template for entry-level users interested in make the jump to HPC. He also saw outreach and education as ways of getting the HPC message out and creating a critical mass of qualified practitioners.
Ultimately though, Feiereisen believes that high performance computing has to become accessible enough to be a “pull” rather than a “push” technology. Obviously, there’s no magic bullet for that, but at least there seems to be pretty solid consensus in the community now that they need to find some new ways to connect the technology dots.