Last week, on April 12, I joined Alan Chalker from Ohio Supercomputer Center and Lee Margetts from simulation industry association NAFEMS to discuss the impact of affordable HPC (high performance computing).
The webcast is now available on-demand to listen and watch here.
The Premise: People often associate HPC or supercomputing with massive servers installed inside a climate-controlled room, located in the pristine glass-encased university research lab or the basement of a secretive government institution. But the emergence of on-demand HPC providers has eliminated the exclusivity. Now, whether you’re simulating the airflow inside inside a commercial airline’s cabin or rendering a photo-realistic animation that shows how a land excavator would work, you can tap into HPC resources available on-demand — without owning or buying HPC hardware. This democratization of HPC is expected to change the way engineers work.
Excerpts from the Discussion: Chalker recalled, “Time and again, I’d go to Ohio-based businesses and ask them, ‘What’s your bottleneck?’ Invariably they’d say, ‘No, we have no bottleneck.’ But when I ask them to explain their process, they’d say, ‘We design the model in SolidWorks or CATIA. We’ve gotta get the CAD model done by mid afternoon so I can get the [simulation] to start running, keep it running overnight on my desktop … to finish by the time we come in tomorrow.’ I’d say, ‘That’s the bottleneck.’ They don’t even realize they’re formatting an entire workflow around the need to run [simulation] overnight.”
Without in-house HPC hardware, businesses tend to schedule their compute-intense simulation jobs to run overnight so they won’t interfere with other operations that take place on the workstations. The strategy makes perfect sense when on-demand HPC is not an option; however, with the current option to access remote HPC resources and pay for them per usage, this strategy should be reexamined.
“With on-demand access to HPC, you can get your answers [simulation results] in a matter of hours, then you can do this routine multiple times a day,” Chalker pointed out.
Engineers also tend to simplify their models to reduce the amount of computing required to simulate the desired scenario. This, too, was a best practice prompted by the need to keep the computation burden manageable in the absence of onsite HPC. With on-demand HPC, engineers should rethink if simplification still a good idea.
Chalker recalled a plastic bottle manufacturer’s experience. “They want to remove every microgram of plastic they can [to make the product lighter and cheaper to manufacturer]. They have to model [the bottle], mesh it, and simulate it. They were working on hundred-hour-long simulation sessions with meshes that are in the order of a few hundred thousands elements. By increasing the number of elements to a few million, and also by adding more processor cores on-demand, they can get their answers in about 20 hours, working with higher resolution and higher fidelity,” he said.
NAFEMS’ Margetts said, “One of the things that frustrates me as a HPC user is when things don’t work. then I have deal with system administration, to try to find out why the [HPC setup] is not working. But when the technology is an on-demand app [HPC delivered remotely like an app], all of that is taken care of. It really improves usability.”
Margetts cautioned, however, that certain IP protection practices and data-storage regulations, especially in the EU countries, could present roadblocks to using on-demand HPC.
“In a simple transaction between a customer and a supplier, the legal obligations between the parties are easily established, making it easy to comply with the existing data protection laws,” Margetts said. “With cloud computing [a key component of on-demand HPC], it gets trickier. The roles of each party are not that straightforward.”
Audience Questions and Panelist Responses:
Question: I am interested in HPC software that can be run on Raspberry PI 3s. I have run OpenSUSE, SUSE Linux Enterprise Linux (SLES), Microsoft 2016 Nano and core, as well as Raspbian on the PI 3 platform. We use IBM’s open source clustering SW for compute nodes.
Chalker’s response: Technically pretty much any Linux-based HPC software package can be made to run on any Linux system, including a Raspberry PI. When we first configure a new HPC system, we start with a single node on which we build the environment, which then gets replicated across all nodes. The same could technically be done with a Raspberry PI; however, it would kind of be like trying to drive a go-cart at the Indy500 — the hardware just doesn’t have enough capabilities to really make it worth your while for anything other than an educational exercise.
Question: Do you feel like the primary driver toward open source has more to do with preventing vendor lock-in? By open-sourcing certain parts, it enables a standardization, which allows jobs to be portable between cluster vendors.
Chalker’s response: I haven’t seen any significant indication of concern regarding vendor lock-in. And I think the majority of HPC users aren’t trying to run on multiple clusters. Rather, they have a primary resource provider that they use almost exclusively. As I said in the webinar, I do think the main driver towards open-source is the fact that lots of users don’t need the full ‘Cadillac’ of features (and the corresponding costs and licenses), and thus are willing to settle for less features for lower cost. I also think some of it is due to the stereotypical “hacker” mentality the many computational scientists have — They want to be able to poke around inside inner workings of their tools and see what they can learn or improve.
Question: What software should we be looking at for SMBs and smaller enterprise customers that need to build out either internally, in the cloud, or as a hybrid environment?
Chalker’s response: To use another car analogy, asking ‘What software should we be looking at?’ is like asking ‘What car should I buy?’ There is no one generic solution that suits everyone. It really depends upon the specific needs and preferences of each organization (do you need the equivalent of a sports car, a minivan, or a hybrid vehicle?). User requirements should be collected, which will drive what types of software are needed. To get an idea of the possibilities, I’d recommend visiting the website of any of the big supercomputer centers such as mine, which usually detail out all the software we run on our systems.
Question: Do you feel like creating an app marketplace will allow crowdfunded hardware startups to justify simulation to prevent major design issues?
Chalker’s response: Yes! Having an app marketplace is just one aspect of making modeling and simulation more accessible, affordable and scalable. This can be to the benefit of startups who want to virtual test out something before trying to secure funding. I’ve had several similar companies approach OSC in recent years with exactly those types of queries, and I expect such clients will become more and more common.
Moderator Wong’s response: A simulation app is usually much easier to learn and use than a general purpose simulation software. Whereas someone who’s an expert in CPU behaviors and PC building can start using a CPU heat-sink simulator app within a couple of hours, the same person, if he or she is unfamiliar with simulation software, will need to invest months or years to master a finite element analysis (FEA) or computer fluid dynamic (CFD) program. So a startup with limited simulation software expertise may prefer an app.