Most manufacturers, especially small and medium businesses (SMBs), mainly use desktop workstations for their daily R&D work, according to NCMS (the National Center for Manufacturing Sciences). Often they do prep work during the day and production runs over-night, resulting in one simulation job per day. In this article, we will investigate how breaking the CAE jobs free from the restrictions of the workstation environment and moving them to the cloud could benefit the CAE engineers as well as the CAE software vendors.
Let’s start with some HPC facts before we look at the Cloud
An NCMS study shows that only 15% of the US manufacturing companies are utilizing high-performance computing (HPC), where the rest, the 85%, are using technical workstations to do CAD and CAE.
In the same NCMS study 57% of the manufacturers said that they have problems that they can’t solve with their existing desktop computers. Clearly, they have a real need for more computing power and again clearly some barriers are preventing them from getting access to it.
Most important benefits for workstation users to upgrade to entry-level HPC are:
1. Enormous economic benefits (Alcoa reported a 98% cost reduction in product testing).
2. Optimized processes (Procter & Gamble saved millions by optimizing its Pringles line).
3. Shorter time to market (PING Golf cut its design cycle time by 68%).
The most severe barriers to HPC adoption of workstation users are: lack of application software, lack of sufficient talent, and cost constraints.
Therefore, buying their own high-performance multi-server clusters to speed up each simulation run, do more frequent CAE simulation runs, or to analyze larger geometries, finer meshes/more cells, or better physics, are simply out of reach for many companies. And here is where HPC in the Cloud could really level the playing field.
A Convincing CFD Use Case Scenario – Private versus Public Resources
It is quite common for R&D teams to follow an iterative CAD/CAE process for designing or modifying a product’s geometry by gradually modifying the physics (application software, set of input parameters, initial flow field), and performing many simulation runs on the engineer’s workstation to find improvements.
To illustrate our scenario, suppose the engineer selects a discrete granularity of 20 million cells (or finite elements) for his analysis and composes an application batch job which then runs 15 hours on his high-end $10K workstation (these days e.g., Intel Xeon E5-2670 dual 8-core, Sandy Bridge), which has just enough memory to host the whole problem. This means one job a day, five jobs a week. Because the workstation runs at its limits, there is no way to speed up these jobs, to run larger geometries and more complex physics; or any opportunity to improve the quality of the results.
Suppose the engineer believes that the quality of the simulation results would improve from a finer mesh decomposition, say by a moderate factor of 2 in each (x,y,and z) dimension, resulting in about 2*2*2 = 8 times more cells (or finite elements), 120 million, 8 times more memory need, and at least 8 times longer runs (8 * 15 = 120 hours) of the batch job on the workstation. But again since the workstation is the limiting factor there no way to perform such a job his workstation due to its memory and computing power limitations. And with the multi-server HPC being out of reach, our engineer has no where to turn.
Cloud computing offers a potential solution: no upfront capital expenditure, no lengthy and tedious purchasing procedure, no management approval necessary, no deep HPC expertise needed.
Getting back to our scenario above, and focusing on Amazon’s EC2 Compute Cluster Instances (CCIs) as an example, we find the Eight Extra Large CCIs comparable to our engineer’s workstation, each CCI equipped with dual 8-core Intel Xeon E5-2670, at $2.40 per CCI per hour, on demand cost about $0.50 on the Amazon spot market.
Submitting now the above mentioned 15-hour job to EC2 and running it on 20 CCIs would reduce the run time of the job to about 1 hour, at a total cost of 20 CCIs x 1 hour x $2.40 = $48.00 for the whole job. Let’s not forget the cost for the application licenses on demand, for 20 CCIs for 1 hour. It’s important to point out that we assume the software is able to make use of the available CPU’s and process the job in many parallel pieces efficiently, which is a quite safe assumption for CFD software.
Let’s now return to our scenario and see what happens if we apply the finer mesh decomposition, which will increase the number of cells by a factor of 8. On the workstation this job would run in 8×15= 120 hours (5 days) and on 20 CCI instances on EC2 the runtime would be around 8 hours for the cost of 8 hours x 20 instances x $2.40 per instance= $384.
This simple example demonstrates that a much bigger job, which was impossible to execute on a workstation, can be run on Cloud resources in a short time and at a reasonable price.
What does this all mean for the manufacturer and for the ISV?
Let’s look at the benefits of this new computing paradigm for the stakeholders. One important remark upfront: we don’t mean to suggest replacing the engineer’s workstation and ISV’s software license on that workstation with Cloud resources. These tools will continue to be invaluable for the engineer and his R&D projects. We view Cloud computing and on-demand software licenses as an additional, complementary benefit for the engineer’s daily work.
The Engineer’s benefits:
The benefits for the engineer in using Cloud resources have been demonstrated above: faster execution of the job; execution of the job in the Cloud while preparing the next job on his workstation; running more than one job at a time with (slightly) different parameters thus increasing the quality of the results; and running bigger jobs with finer geometries and better physics thus further increasing the quality of the results.
The ISV’s benefits:
An engineer turning to the Cloud because of the just mentioned scenarios and benefits will (have to) continue to use the application software license on his workstation, for business as usual, as demanded by the typical and standard R&D scenarios described above.
The Cloud enables additional opportunities for the engineer to do more, faster, better. For all these additional scenarios, the ISV sells additional licenses, on demand, paid by the hour, resulting in additional business, on top of the workstation license business.
Even for those companies which already have an HPC cluster in their computing center, bursting into a Cloud offers much higher efficiency and flexibility for the engineers, and results in additional license-on-demand business for the ISV.
Finally, ISVs offering application software on demand are able to attract new customers who are just getting started with computer simulations, and who would never think of buying a license for just a few simulations.
For all these good reasons, I strongly doubt that there will be any negative impact from cloud-based on-demand licensing for ISVs, as it is sometimes concluded in our community. On the contrary, if introduced with care, e.g., with some early incentives for their existing customer base, ISVs will be able to add new business right from the start.
What are we waiting for?
To explore the Cloud computing model in CAE, we have developed an experiment that brings together the four major stakeholders: manufacturing end users, resource providers, software providers, and subject matter experts.
The idea is to better understand the end-to-end process of using remote resources available in high-performance technical computing centers and in clouds, and figure out how to achieve the benefits of service-based delivery. We believe the technology is not the challenge anymore; rather the challenge is bringing the right people together.
The experiment is scheduled to begin later in July and will run for three months. At that point, the results will be made publicly available to the manufacturing and the HPC community.
Anyone interested in participating at this experiment can register at http://www.hpcexperiment.com. More information about the experiment is available at http://www.hpcwire.com/hpcwire/2012-06-28/the_uber-cloud_experiment.html.