October 03, 2012
NSF and NCSA recently commissioned a study to see whether improvements in the science inside applications and other factors could help industrial HPC users. Merle Giles, director of NCSA's Private Sector Program, discusses the findings with HPCwire.
HPCwire: What was the goal of the study?
Merle Giles: The main goal was to better understand the root causes that limit the realism and performance of today's HPC applications and to solicit ideas for improving the applications' performance and realism for industrial users.
HPCwire: Why is HPC use in industry important to NCSA and NSF?
Giles: Engineering and scientific research problems have a lot in common. They can be equally complex and are based on the same underlying science, so advances in science can often be leveraged across both domains for greater benefit to society. Industrial engineering has a special value because it can function as a living laboratory for applying new science and getting feedback under the tight time constraints of the product design window.
HPCwire: How are NSF and NCSA involved in helping industrial HPC users?
Giles: NSF and NCSA are heavily involved in supporting engineering as well as science. NSF's XSEDE program stands for Extreme Science and Engineering Discovery Environment. It's a five-year, $121 million initiative to support scientific and industrial researchers across the world with, quote, "the most advanced, powerful, and robust collection of integrated advanced digital resources and services in the world."
NCSA is the lead institution in this partnership, which involves more than a dozen major universities plus other organizations. In addition, NCSA formed the Private Sector Program that I'm in charge of, so we could make our HPC resources and the best research minds at the University of Illinois available to industry.
HPCwire: What prompted the new study?
Giles: Industrial users were telling us they needed better simulation-based science and engineering. They said the performance of some of their key applications in HPC environments was stuck and they needed better solutions so they could develop higher-quality, more competitive products in shorter timeframes. Industry has a time scale for innovation that the academic world often doesn't. We need both ends of this time scale.
Scientific researchers may need years or decades to advance the frontiers of knowledge in their fields. The industrial community, on the other hand, is driven by more immediate competitive pressures to build higher-quality products sooner. So, this study was prompted by industry's urgent need to know what is limiting the performance and realism of their simulations and what can be done to address the limitations.
HPCwire: Can you give me an example of an industrial user facing limitations?
Giles: Sure. One auto company said they've been pushing up against the limits of single-domain CFD and now need multi-domain simulations. We felt we owed it to users like this to document what's limiting the realism. We've learned a lot in this process.
Collaboration with academia will be crucial for advancing the applications. In many cases, the academics know how to do this but industry doesn't necessarily. There's a clear need for academia and industry to work together to address these limitations, and that will help both camps because academic scientists and industrial scientists have many of the same issues.
HPCwire: Who conducted the new study and what methodology did they use?
Giles: We hired IDC to carry it out. They conducted extensive interviews in person and in writing. They also organized a focus group to probe some of the issues more deeply.
HPCwire: For those that aren't getting the HPC performance they need today, did the study show how are they dealing with this inadequacy?
Giles: Yes. 30 percent of the organizations named specific ways in which they now have to dumb down their problems to complete the runs in reasonable amounts of time. The most popular strategies are using coarser meshes, employing fewer elements, not fully exploiting the known science, and employing fewer time steps or reducing the length of the investigated timeframe.
Industry uses HPC to test to a spec, but they don't have enough time to test to see where the point of failure is. They dumb the problem down to validate the spec. But without testing to failure, they can't see how much a product may be over-engineered. Over-engineering can cost a lot of money and reduce product efficiency and competitiveness. Advances in the HPC simulations could let them test to failure within the design window.
HPCwire: Who was interviewed?
Giles: We started earlier on with NCSA partners and IDC expanded from that base. Of the surveyed organizations, 17 percent were in government, 7 percent were academic, and the large majority of 76 percent came from industry or from organizations closely allied with industry.
The organizations ranged from a small business with 22 employees to one with 164,000 people. Revenue ranged from $240 million to $129 billion, and profits varied from $15 million to $5 billion. R&D investments ranged from $515 million to $5 billion. On average, the responding companies invested about 6.5 times more in R&D, as a percentage of their revenue, than all U.S. companies did as a group.
HPCwire: What do you think were the most important findings?
Giles: It haunts me that only about one in six organizations said that their applications as now written would meet their requirements for the next five years. As I said earlier, 30 percent said they have to dumb down their problems today. Another important finding is that users are okay with today's HPC systems but many need retraining for next-generation systems. It will take more hands-on training and education to move industry forward.
HPCwire: What was the most surprising finding?
Giles: The similarity of responses regardless of company size. There's a mistaken impression that the small and medium-sized companies have fewer challenges. The study confirmed that companies of all sizes need to reduce uncertainty. This is often due to the limitations of the science embedded into the applications today. To increase their confidence levels, they need more HPC capacity and capability.
HPCwire: You asked the surveyed organizations what they could do with unlimited computing power applicable to their problems. What did they say?
Giles: Some organizations provided very specific examples. An aerospace company said they could simulate the complete turbine assembly of a jet engine or optimize an entire new engine design. A life sciences firm would be able to predict the function of an ensemble of living things in a sample of soil based entirely on the DNA sequences found in that soil sample. That's called metagenomics. A manufacturer said they could gain knowledge faster and sooner in the product development cycle, and ultimately be able to design by analysis and validate the designs prior to first build and use first build to further validate the simulation model.
HPCwire: Did they give you a wish list of high-priority things they needed?
Giles: Yes. We asked about improvements that could be accomplished in one to two years. Most of these near-term improvements fell into a few categories. These included higher-resolution meshes, improved mathematical models and algorithms, improvements to the underlying physics, and better methods for data integration and analysis.
Another frequent desire is to be able to do multi-disciplinary, multi-scale simulations. Large-scale data integration is related to this challenge and is also on the wish list. In fact, multi-disciplinary, multi-scale simulations would be helped by every other item on the wish list.
HPCwire: If they got the items on their wish list, what difference could it make?
Giles: The wish list is needed to exploit the science we already know today. A large majority of the respondents believe that today's known science could support a moderate or a large amount of additional realism in the applications. More than a third of the respondents said that taking the next step, advancing the known science, could add even more realism to their key applications.
HPCwire: Could NSF and NCSA play a role in advancing applications capabilities for industrial organizations? Would industry want to work with you on this?
Giles: About two-thirds of the organizations said yes to having NCSA and/or NSF heavily involved in the process of advancing their important applications. The respondents were saying that we have really good tools out there and today's science is good, but NCSA and NSF can help the user community to exploit the tools better and to advance the science and the models underlying the applications.
This confirms what we've been hearing separately from this study. Companies are looking for expert organizations like NSF and NCSA to play a more consultative role in helping them to use advanced science and engineering tools. More than one-third of Fortune 50 companies have chosen to work with NCSA's Private Sector Program.
HPCwire: Is this study publicly available?
Giles: It will be publicly available in a few months.
10/30/2013 | Cray, DDN, Mellanox, NetApp, ScaleMP, Supermicro, Xyratex | Creating data is easy… the challenge is getting it to the right place to make use of it. This paper discusses fresh solutions that can directly increase I/O efficiency, and the applications of these solutions to current, and new technology infrastructures.
10/01/2013 | IBM | A new trend is developing in the HPC space that is also affecting enterprise computing productivity with the arrival of “ultra-dense” hyper-scale servers.
Ken Claffey, SVP and General Manager at Xyratex, presents ClusterStor at the Vendor Showdown at ISC13 in Leipzig, Germany.
Join HPCwire Editor Nicole Hemsoth and Dr. David Bader from Georgia Tech as they take center stage on opening night at Atlanta's first Big Data Kick Off Week, filmed in front of a live audience. Nicole and David look at the evolution of HPC, today's big data challenges, discuss real world solutions, and reveal their predictions. Exactly what does the future holds for HPC?