Council on Competitiveness studies have demonstrated that small and medium-size businesses (SMBs) face major hurdles moving to HPC. Auto/aero supplier Swift Engineering proves how worthwhile the journey can be. HPCwire talked with Swift chief scientist Mark Page about the transition and its benefits.

HPCwire: What exactly does Swift Engineering do?

Mark Page: We specialize in high-performance aerodynamic vehicles, including race cars and UAVs [Unmanned Aerial Vehicles]. We also work with sophisticated composite structures for companies like Northrop Grumman, where we build a van-sized radome for the Global Hawk UAV, and ventral fins. On most projects we design complete vehicles, UAV's, race cars, or jets. So we don't just build these things. We do the configuration design and synthesis to the customer's specs.

HPCwire: Do you use a mix of physical and virtual prototyping?

Page: Swift uses both physical wind tunnels and virtual modeling with CFD. Each of these methods has its advantages, depending on what you're trying to do. With virtual modeling of a design, for example, you don't have to deal with the artificial effects of wind tunnel walls, model supports, or humidity changes. Wind tunnel testing has been a huge part of what we do, but each year more of the work moves to HPC. We see this trend not just at Swift, but everywhere. Wind tunnel testing is still best for some things, such as "mapping," but CFD can't be beaten for others.

HPCwire: Who are your principle customers?

Page: Much of our business is with the big OEMs, companies like Toyota, Northrop Grumman and Boeing. Some other customers need access to our wind tunnel and our other resources. They have a concept, and they use us to make the prototype. We help them design it, build the model, test it in our wind tunnel, and finally, we build the prototype. They may have an idea for an airplane with certain specs, but they may not know how to design composites, or how to interpret the FAA regulations, or they might not have the CAD software, etc. They want to build interest with a customer for their concept, and we help them.

Toyota came to us to help them get into NASCAR. For two-and-a-half years, we secretly developed the aerodynamic shape of the Tundra truck and then helped with the Camry for NASCAR. Northrop Grumman came to us to develop and build the BAT UAV, which we now manufacture and develop further. Eclipse Aviation asked us to secretly design a new, light business jet for them. We designed and built it in 200 days, that's from design to first flight. Eclipse provided us with sub-systems and we did the aerodynamics, wind tunnel testing, and built the airplane.

HPCwire: How closely related are the work you do for racing cars and your other work, such as for UAVs? Do they share many of the same applications and methodologies?

Page: The technical disciplines are very similar, so our technical team supports both types of work. Things begin to get specialized more at the architecting and program management levels. The design criteria and certifying agencies are very different, so we need specialty experts there. But from the standpoint of the basic disciplines, such as structures, aerodynamics and electronics, it's very similar.

HPCwire: What advantages do you have over a large OEM for doing this work? In other words, why would customers come to you for this?

Page: Northrop Grumman has world-class experts in every field, but the company is not organized for quick turnaround. We design a race car in one season that would take a big OEM four years, and there are cost benefits associated with our approach. But we make prototypes, not certified production vehicles like the ones Northrop and others build. That's why it takes four years.

HPCwire: How long has Swift Engineering been using HPC servers and software?

Page: We've been running CFD on a small cluster since 1997. We upgraded to a true HPC system in May 2010. It's 10 to 100 times more powerful than the small cluster, depending on the application.

HPCwire: Not long ago, HPC systems were too expensive and hard to use for most SMBs. How do you think this has changed?

Page: You could build a high performance platform with off-the-shelf hardware, but then you'd need intensive IT activity to maintain and upgrade that system. People are moving away from building clusters out of PCs and servers. Also, advances in cluster management software have made things much easier. Cost-effective, powerful processors and other hardware and software components have been packaged together by places liked Cray, so we don't need our own IT person to maintain the system. And ISVs design their software to run on these systems, so we don't need to make mods as we used to do to run the apps on the earlier cluster hardware.

HPCwire: Which HPC system do you use, and why did you choose it?

Page: We use the Cray CX1000, which is a rack-mounted system based on Intel Xeon processors. Cray has a proven HPC track record. We wanted that kind of support. As I said, we used a home-made cluster for years and said, "No more." The CX1000 is also designed to be expanded, so we have headroom for growth. Home-built clusters can also be expanded, but the components quickly become obsolete and then you have to figure what to do next. We're not in that business. We wanted an expert to do that for us.

We also have a Cray CX1. That unit has a visualization blade and a data storage blade. It has 12-core processors with 48GB of RAM. The storage blade has dual quad-core processors and 24GB of RAM. We use the CX1 for grid generation and for pre-and post-processing and storage.

Our big machine, the CX1000, has 18 nodes of Westmere dual quad-cores for the compute side, with 24GB of RAM each node. That's a lot of RAM.

HPCwire: What about software? What OS and applications do you use?

Page: Our OS is CentOS 5.4, and our cluster management software is Platform Cluster Manager. The main app we use for grid generation is MIME from Metacomp. We also use CFD++ from Metacomp. We've started using Sculptor morphing software from Optimal Solutions for 3D design. It gives us a way to deform shapes smoothly, without re-gridding. A few years back Formula-1 cars became ultra-swoopy, and this was the influence of Sculptor. We have also looked at Pointwise for CFD grid generation. On the structures side, we use RADIOSS, Altair's Optistruct, and NASTRAN.

HPCwire: What benefits are you seeing from using the HPC system? Is it enabling you to be more innovative and competitive? What can you do that you couldn't do before?

Page: The Cray CX1000 has totally transformed our processes. It's been the biggest single step forward since I've been at Swift. We now explore aerodynamic design spaces 10 to 100 times faster than before, and ask questions we never thought of asking at the design stage. We're getting good answers, so the line of users with questions is long.

HPCwire: What was your design workflow like before you started using HPC, and how has this changed?

Page: It's completely changed. In the old days, we had a designer for the wind tunnel model, and then we went to a specialized machinist to make the parts, and then to the wind tunnel to help us decide which way to go. Now we only have to worry about the outer structure, and we can skip the fabrication and test. Instead of using a team of mechanics in series, we have the aerodynamics guy hand off to the surfacing guy, and then it goes directly to our CFD guy. We've cut three-quarters of the people out of the loop. The turnaround time is 5 to 10 times faster.

With HPC, as with wind tunnel testing, you need to build a formal process to queue up the work, because the demand to use the system is great. We have had to make process changes to satisfy all of our internal customers.

HPCwire: What has the move to HPC done for your business?

Page: It definitely improves our ability to get business. The Cray has an iconic name and being able to say, "We have a Cray," really impresses people. Much of our business is referral.

HPCwire: How difficult was the transition to HPC for Swift Engineering? Were there any surprises?

Page: In general, it was not difficult. It takes some time to prep the building for installation, especially the cooling to accept this more powerful machine.

HPCwire: What advice would you give to other small and medium-size organizations that are contemplating a move to HPC?

Page: Our transition wasn't difficult because my colleague Dr. Winkler spent time assessing the alternatives. Because the system arrived as a pre-integrated package, all the interfaces were handled. With our home-brewed cluster we had to spend many months getting it to work at all. It's worth taking the time to figure out what will be most cost-effective for you. You need to ask what processor speed you need, because more-powerful processors cost more. The key is the CFD or other software you need to use. This will determine which way to go. You need to understand the software licensing model. Some ISVs charge by the core, and fewer, more-powerful cores could save money. It's also good to benchmark software packages to see which software scales well.

And don't forget the infrastructure costs, which in our experience are 15 to 25 percent of the machine cost. We chose to have a dedicated room with dedicated power, surge protection, and an uninterruptible power supply. These cost money to install. We also needed to install a dedicated cooling system.

HPCwire: Was it worth the trouble?

Page: Absolutely. We increased our capability 10 to 100 times, depending on the problem. The machine's fully booked around the clock. We aren't just doing the same things faster. We're doing way more things we haven't had the ability to do before. So, our costs are 20-30 percent less than before, and we're doing three to four times more work.

But it's more than that. You also get to understand the problems much better. Our productivity in aero-physics has increased because everyone now has their own little digital wind tunnel. We're doing Reynolds number surveys on our UAVs, correlation exercises with the broader industry, time-dependent rotating reference systems for windmills, and a lot more new things.

HPCwire: Is there anything important we haven't covered yet?

Page: Yes. As with anything that provides an answer, the user is ultimately responsible for the veracity of the solution. The hardware and software vendors are giving us these awesome resources. You always want to do a truth test when you go into a new area. Find a known answer to calibrate against, because you're doing pioneering work and you will get an answer even if you, as a user, have made a colossal error. You need to be on top of this.

The software vendors have made huge strides in visualizing input and output. The next step would be for them to create a dashboard that does a visual play-back of the inputs against simple standards. For example, if all the inputs are A-OK except for viscosity, there may be no alert that such a combination can't exist in our atmosphere. This is challenging because sometimes you're looking at flow in a pump where the pressures are way beyond atmospheric, or maybe the flow isn't even air. But for each fluid state there are combinations that are "normal." The dashboard would need to incorporate a lot of intelligence and best practices.