Imagine you've just finished a visit to your favorite Scandinavian furniture store. You've strolled around, had some coffee and hot dogs, and even bought a glass jar with that strange pickled herring to try out at home. And you've finally purchased that new large and comfy corner couch that you've always wanted, and now you need to get everything home.
But there's no chance you can fit this all into your sedan. What do you do?
You might now ask yourself: why is he talking about furniture? I came here to read about CAE modeling and the cloud! I know and bear with me because the above situation is a perfect metaphor for a well-known phenomenon in physics modeling: your model has become too large for your computer.
Why cloud-based HPC is much like renting a car
The couch problem has two solutions. You can ask the furniture store to deliver it to you. But you'll have to coordinate with them to find a delivery time, and you don't know if they'll take proper care of the couch while transporting it. It'll probably also be cost-intensive since they have expenses to cover. In short, it can be a hassle.
Or you could rent a van or even a truck, and fetch the couch when it fits you. After you've finished, you hand the car back and go back using your sedan, which is practical for your everyday drives. And this is the way HPC in the cloud works: for your non-demanding tasks, you use your laptop or workstation, and when you need more computing power, it's just a mouse click away.
Case study: acoustics
Let's take a look at one specific area: acoustics. When modeling acoustics with the finite element method (FEM), the frequency that you can model directly correlates with the size of your mesh. The usual rule of thumb is that you need at least five elements per wavelength in each direction. A finer mesh can resolve higher frequencies, something that is crucial when you model, e.g., loudspeakers or headphones.
Keeping this in mind, we see that modeling higher frequencies become computationally intensive very fast. Add to this the aspect that you might want to compute the effects in larger spaces, such as a living room, and you will hit the limits of your local workstation or laptop very fast. Your local computer has turned into the sedan in the furniture example.
So you decide to rent more computational power. You can swiftly get cloud-based machines with more than 700 GB each and connect these into a cluster to compute models that need of 20-30 TB of RAM, and even more.
So stop feeling limited by your local machines and start dreaming about what you could do if you use the cloud.
And keep in mind that this benefit isn't limited to acoustic simulations. Mechanical, fluid, and chemical simulations can also benefit from more available memory.
But aren't cloud clusters only suitable for parametric sweeps?
We went through parametric sweeps in an earlier blog post and explained why the cloud is an excellent solution for parameter studies. Historically, this was also one of the few CAE areas for which a cloud-based cluster was feasible. Lack of high-speed low-latency networks made distributed memory models hard to compute.
Note that I said historically. Nowadays, you can get access to dedicated high-performance networks on several cloud providers. Amazon's AWS has its Elastic Fabric Adapter (EFA), and Microsoft's Azure has Accelerated Networking (AN). Both of these networks are very well suited to high-performance computing, and software like ANSYS has been showed to scale very well.
Combining this with the easy-to-deploy UberCloud Containers yields a versatile and flexible system. The old limitations of the cloud are long gone, and your on-premise systems are now what limits you.
How can I try it out?
If you want to try COMSOL, or any other software, we at UberCloud can help you get going. We'll work with your IT department to make your license server cloud-ready, and we'll work with you to make sure your first steps into the HPC in the cloud are successful.
Get in touch today!