To get the most value out of engineering simulation, ANSYS customers often take advantage of high performance computing (HPC). In simple terms, HPC enables you to apply a group of computers running in parallel to solve larger problems and/or reduce the solution time for a given problem. Unlike “embarrassingly parallel” applications like genomics or graphics rendering, all of the compute cores involved in a single Computational fluid dynamics (CFD) simulation need to communicate with each other during the solution process. That places significant demands on the network fabric used to connect the machines. Cloud computing can certainly provide computing capacity at a vast, global scale, but can it provide the desired HPC performance?
It is accepted as fact in the engineering simulation community that, if you want to get serious about HPC scalability, you need to have a low latency, high-bandwidth network fabric. Otherwise, performance suffers at high core counts, and throwing more cores at your problem won’t reduce the overall solution time because of the overhead associated with interprocess communication. There’s truth in this conventional wisdom, but before you decide it’s Infiniband or bust, it’s worth knowing just how far you can get with more conventional 10 GigE networking. Is there a price/performance sweet spot where you can still take advantage of HPC using lower-cost “commodity” infrastructure, but still gain the benefits of HPC?
ANSYS, in partnership with Amazon Web Services (AWS), recently ran a series of tests to demonstrate the scalability of ANSYS CFD on the AWS Elastic Compute Cloud (EC2) using the current generation of compute-optimized instances that support Enhanced Networking. Here, the instances used were AWS c4.8xlarge instances; these instances use high frequency Intel Xeon E5-2666 v3 (Haswell) processors optimized specifically for EC2 and offer 18 physical CPU cores and 60 GB of RAM. The data below shows results for a standard ANSYS CFD benchmark problem which simulates the flow around a Formula 1 race car with 140 million cells. The results may alter your opinion about whether or not you can do “real HPC” on AWS. The results show near-ideal scalability well past 1000 cores and a reduced overall solution time even beyond 2000 cores.
While an InfiniBand-backed HPC cluster would show scaling closer to ideal, AWS has advantages when you consider price along with performance. AWS offers a global reach and economy of scale that is unmatched. Where AWS really gets interesting for HPC workloads is when you consider Amazon EC2 Spot instances. Spot instances allow you to bid for cloud computing resources that are currently unused and pay a market-driven price that is typically much lower than on-demand pricing. As long as your bid price exceeds the market price, you will have the use of the cloud instance you requested.The market is variable, depending on the region, instance type and of course market demand. With Spot instances, we were able to run the benchmarks above using c4.8xlarge instances in the US West (Oregon) region for roughly 1/4th of the on-demand price.
On September 27th, ANSYS will host a free webinar to talk more about its new, cost-effective options for cloud-based engineering simulation. I encourage you to register here to learn more about this economical spot market pricing model.
While it is true that the very high end of the HPC space is still best served by clusters with an InfiniBand network fabric, this data shows that the AWS offers HPC performance which merits serious consideration for the overwhelming majority of ANSYS customers. It’s worth noting that, for ANSYS CFD, HPC scalability is problem-dependent; the most important factor is mesh size. Experience shows that good scalability can be achieved on 10 GigE networking provided that you stay above about 50,000 cells per core. That means, for example, that a 20 M cell problem shows good scaling up to about 400 cores; beyond that, interprocess communication latency begins to dominate and scaling degrades.
Using HPC for engineering simulation has proven itself many times over. If you need more reasons for moving your engineering simulation to the cloud, read my blog post, “Cloud Offers Much More Than Unlimited Computing Power.”
I’ve shown here that AWS can provide excellent HPC price/performance, especially if you follow these suggestions for HPC best practices. ANSYS Enterprise Cloud offers a turnkey virtual simulation data center running on AWS EC2 that delivers the HPC performance I’ve highlighted here in an easy-to-deploy package
To learn more, read the articles I’ve linked to in this post, and view the on-demand webinar recording that discusses the economical spot pricing and pay-per-use licensing models.