Webinar: Cloud Premium: Cluster Computing Arrives for Autodesk CFD

Following is the transcript from the recent webinar, “Cloud Premium: Cluster Computing Arrives for Autodesk CFD,” presented by James Neville of Autodesk.

My name is James Neville and I am a simulation engineering here at Autodesk. I want to welcome everyone here today to our webinar, and thank you all for coming out to learn a little bit about the Autodesk simulation. Today, in addition to giving a brief overview of Autodesk’s CDF, our flagship computational fluid dynamics program, I’m going to focus on a new feature Autodesk has recently unveiled called Cloud Premium.

My goal is to ensure that everyone on the call walks away with an understanding of how CDF can be used in the design process, how customers are being successful today, and ultimately how you all can apply our groundbreaking cloud technology to your own engineering challenges to solve faster, solve bigger, solve more. A little bit about myself before we get rolling. I’ve been working in the simulation field for a little over 12 years now. I started by career back in 2005 working for Blue Ridge Numerics, and Blue Ridge was a technology start-up with a mission of democratizing CFD for design engineers.

Back in 2011 Autodesk acquired Blue Ridge and added CFdesign to their growing portfolio of simulation products. And CFdesign serves as the foundation for all of Autodesk’s CFD products today. I’ve held various roles over the years ranging from tech support to application engineering to consulting services, and now I serve as a global CFD expert for Autodesk. I’m happy to be here today and to provide some insight into Autodesk’s simulation program. Regarding the webinar agenda today, we’ll start with a brief overview of our simulation tools. I want to give everyone a complete picture of our offerings, and then we’ll move on to a few CFD specific topics like cloud computing.

To help set the stage for Cloud Premium, I’m going to go over a bit of the history of our CFD product, and then we’ll get into the good stuff, and we’ll talk about the newly released technology. And then finally at the end we will wrap it up with some Q&A. To back up for just a minute, I think that when a lot of folks

hear the company name Autodesk they immediately think AutoCAD or Inventor, and while this may be correct, it’s only really a part of the picture of the company. Autodesk has built a complete portfolio of simulation products over the last 10 years, and we can group these products into two main categories you see here.
You got upfront engineering and advanced materials. These tools range from CFD and FDA solutions flow instructional mechanics on the left, all the way to advanced material simulation like plastic injection molding and composites on the right. Today we’ll be talking specifically about the first item on the slide here, flow and thermal analysis CFD. What is CFD? CFD stands for computational fluid dynamics. It sounds a lot more complicated than it is. Fluid refers to both liquids and gases, and dynamic simply refers to movement. A computer is used to perform these calculations, hence the term computational. In short, CFD uses computers to analyze fluid flow and heat transfer problems.

In real world terminology you can think of it as a virtual wind tunnel, or a virtual flow bench or a thermal test rig. There’s a large number of commercial CFD programs out there today, and the vast majority of them are designed for full-time CFD analysts. These programs are considered traditional CFD tools in that they were designed at the PhD level, and are intended to be used at the PhD level. Autodesk CFD is very clearly not a traditional CFD tool. While it solves the same equations as all of the traditional codes, Autodesk CFD is really intended for a different kind of user. Above all else, there is a sharp focus on automation and user accessibility to allow more users to leverage the benefits of CFD earlier on in product design.

The vast majority of our successful users wear multiple hats, and don’t have cycles in the day to run simulation tools full time. Autodesk CFD utilizes a design study environment, and this allows our users to analyze and compare a large number of design variance, instead of just looking at, say, one at a time, and there’s a sharp focus on automation within Autodesk CFD. And our meshing technology is really no exception here. Fully automated mesh, auto sizing and results based adaptation allow both novice and expert users to obtain accurate, repeatable results with really minimal effort. And finally, Autodesk CFD is the only upfront tool, only upfront CFD tool I should say, on the market today that provides infinite computing power through cloud access for design engineers, and I’ll be diving deeper into this topic here shortly.

You heard me say upfront a second ago. Upfront really refers to a certain point in time in the product development process. Earlier on in product development we have much more opportunity to affect change, and the associated cost of these changes is relatively low. As product development matures more and more items become set in stone, less can be changed, and the cost of any changed dramatically rises. Upfront simulation means leveraging powerful simulation tools in the design and engineering phase where they have the most value to us. From the ground up, Autodesk CFD is an upfront simulation tool.

You may ask what do our customers simulate with Autodesk CFD, and I’d like to rephrase it, I think a better question might be what don’t they simulate. As a general purpose tool, we find that our customers use Autodesk CFD in literally every industry across the globe. You’ve got microscopic medical devices, ceiling fans, Indy cars, stadium wind studies. The tool is really designed to be accessible to all users in all industries. All right, I have a handful of poll questions here and I’m curious to see everyone’s responses. I’ll share the poll results after everyone has a chance to chime in.First one here, in the event that your organization encounters say a flow or a thermal challenge, go ahead and pick which method best fits how that challenge would be addressed today. There’s four options. You’ve got hand calculations, building and testing. You know classic build, test, fail, built, test fail. Outsourcing, that could be outsourcing it just to solve a problem. Outsourcing it for CFD services somewhere else, or in house CFD. Go ahead and submit your answers and I’ll tally them up here in about 15 seconds or so. I’m going to pick in house CFD. All right, we’ll go ahead and take a look at those. That’s higher than I thought. That’s good. Equal votes for hand calcs, build test, outsource, and then 50% of the folks on the call do in house 50. That’s good to know.

I think more and more companies are using CFD these days, bringing it in house as the costs drop. I think Autodesk moving to a subscription based software delivery has helped that greatly. Thank you for answering that one. I’ve got another one here. Here’s the next one. This one’s a simple yes or no. Basically, does your organization use the cloud today for product development? That could be collaboration, asset storage, simulation processing. If you’re already using the cloud basically a simple yes or no here. Let me take a look at the answers here. This one’s a little surprising. I see the vast majority have reported back no. That’s very curious. I’d be interested to dig in with each one of you guys or gals about why that is or maybe help illustrate some of the benefits of cloud computing.

We’ll be chatting about a couple of those here in a second. This is my final poll question. The last one for you. If you think back over the last two years, or even maybe what you had projected for the coming year, about how many unique projects do you estimate could benefit from flow and thermal simulation? All right, we’ll tally these up here. It looks like 50% are at 1-2. The next one up there is 3-4. No one at more than 10. I actually put in a vote for more than 10, but I’m not sure my vote counts. Looks like most of it 1-2. We actually find when we look back over our data from all of our customers over a very long time span, the vast majority of our successful simulation users leverage CFD for four projects per year.

That doesn’t mean four simulations per year, but in general they’re using it for around four projects per year, and that’s using it maybe hot and heavy for a couple weeks or maybe a month or so and then putting it down for a little while. That’s really what the tool’s designed to be good at, and that’s what our data shows. Interesting. Looks like most folks out there have some projects that could benefit from upfront simulation though. All right, let’s get into cloud computing. I think it makes sense to go over some of the history of cloud computing at Autodesk to give Cloud Premium some context. When Autodesk acquired CFdesign in 2011, the CFD solving technology was fully optimized for local machine solves.

Whatever resources were available on a local machine, CFD would utilize to the fullest. You had 2 cores, 8 cores, 16 cores, it just didn’t matter. Once core counts started to reach 16 and higher we did start to see a bit of a diminishing return with additional resources, but in general CFD solving performance grew linearly with hardware. As users obtained faster hardware CFD would take advantage of it. If users say upgraded to faster clock speeds or more cores they would see solve speed benefits. However, with just one license, one CFD license, users were limited to solving one job at a time. But the only real way around this was to purchase additional licenses and additional hardware. That’s a pretty find solution for large companies, large customers who have leverage CFD to solve problems year round, but the vast majority of our customers used the tool in brief bursts when the need came up and therefore couldn’t justify those kind of costs.

In 2012 Autodesk rolled out cloud solving for CFD, and in a nutshell this enabled users to off load their CFD analysis through the AWS cloud where the job would be solved remotely. A single user with only one license of CFD could solve simultaneously for 10, 50 jobs at one time. No additional license was required, no additional hardware was required. Each job would cost 15 cloud credits, and on demand

cloud computing really became the norm for us. Our CFD users essentially had access to on demand remote work stations to solve more jobs in the same amount of time. In the past year alone, I’ve got the numbers here for you, in the past year alone 23,700 jobs were submitted and solved on the cloud via Autodesk CFD.
Our product development team gave me some conservative numbers last night regarding the total number of simulations submitted since cloud computing arrived for Autodesk CFD five-ish years ago. The number is staggering. Over 191,000 jobs since 2012. It’s ridiculous. Acquirements like expensive local hardware, heavy laptops having to go back into the office to check on the status of a lengthy overnight simulation. I’ve done that many times. They’ve all gone away. On a personal note, as a CFD power user, up until this year I’ve always managed to have both a very powerful desktop computer and a powerful and heavy laptop for travel, and I’ve recently downsized greatly to now function of a single high end Ultrabook. I can do infinitely more work from this one device than I could do with maybe five or 10 local machines previously because of the power of cloud computing.

That brings us to now where Autodesk has just release Cloud Premium, which is more or less our industry leading cloud technology on steroids. Cloud Premium is a new option in Autodesk CFD 2018. Users can now select to have their jobs solved on the traditional cloud or on Cloud Premium. Standard cloud jobs are solved on a … Just give you some details on what’s going on in the background. But standard cloud jobs are solved on a single Amazon EC2 instance with 8 physical cores and provide performance comparable to say a high end local machine. Cloud Premium jobs are different. They’re solved and distributed on 16 EC2 instances, each with 16 visible cores. The list distributed computing network allows for a much faster simulation and enables users to solve jobs previously not possible.

Cloud CFD job costs are calculated from a really wide range of variables like model size, requested durations, problem physics, etc cetera. There’s a number of things that we look at. And Cloud Premium jobs will have a multiplier 4 X. They’ll involve a 4 X multiplier over the standard cloud job cost. I’m going to jump right in and share a few examples to illustrate some performance metrics. Starting with internal flow through a plug valve, I’ve got results from Cloud Standard and Cloud Premium this year. We’ll look at this same problem with a few different mesh sizes to show what kind of real world performance gains are to be had.

At a median mesh density of say 5 million elements solve time dropped from around three hours to just under one hour. That yields a 3.1 X speed up. When we bump the mesh count up to 11 million elements, which is starting to get pretty large, the solve time drops from 8.2 hours to an hour and a half. That’s a 5.4 X speed up. Now we’re starting to see some impressive gains. With Cloud Premium, this problem was just reduced from a complete work day of time down to basically an extended lunch break. We have similar performance gains of around 400%. They’re shown when the model size is increased to 14.7 million elements, that’s a pretty big model, we go from 10 hours down to two and a half hours. And then finally we arrive at a 30 million element job.

This is a very large job. It requires an entire day of Cloud Standard simulation time for completion. Cloud Premium manages to crank this simulation out 4.5 times faster at only five hours. The next set of simulations take this performance improvement to a whole nother level entirely. External aerodynamic studies, external flow, these simulations exhibit two properties that typically result in long run times. I think anyone who’s run external flow they’ve run into these. But the first is just the sheer model size.

Often times the mesh required for an accurate solution can be very, very large. The second is just the number of iterations required for convergence. Due to the number of the complex flow structures, the wake structures, many iterations are typically required to fully resolve the flow field.

Cloud Premium really excels at both of these challenges. At a relatively small mesh size of 1.6 million elements Cloud Premium shows a 5X speed up for this race car rear wing assembly. When we bump the mesh count up to 11.3 million elements Cloud Premium hit the sort of sweet spot here and solved this problem 27 X faster than Cloud Standard. That’s 27 times faster. We go from 30.2 hours in Cloud Standard, down to 1.1 hours in Cloud Premium, and all of this performance improvement was accompanied with a slight increase with simulation accuracy as well. Jumping up to 20 million elements, showed a still amazing 10 X speed up for Cloud Premium. And when we run the numbers on jobs like this, jobs that are this large, they just become staggering.

Simulation time dropped from almost 50 hours, that’s over two days, to just five hours, and all of this with equal or superior result accuracy. The plotting these two examples here show how large model sizes previously resulted in lengthy run times of entire work days or even multiple days. They reduced to just a, reduced, excuse me, to just a few hours with Cloud Premium. It’s clear that Cloud Premium has the most benefit for simulations with large element counts. Gains are expected at most model sizes, but the true value of Cloud Premium is realized when users can reduce their simulation run times from days to just hours. In general, users can expect a four to five X speed up for most general purpose simulations.
Certain applications, like the one I just went over, external flow, they show dramatically greater performance improvements on the order of 10 X or more. And our benchmark testing has shown that Cloud Premium simulation accuracy is as good or better than Cloud Standard, and users can basically expect the same great simulation fidelity that they’re used to from Autodesk CFD. Where are we going with this? Cloud Premium is a very exciting technology for us here at Autodesk. We’re excited to roll it out to our customers, and frankly we’re excited to use it ourselves. It really wasn’t that many years ago when CFD was limited to in [inaudible 00:20:48] jobs and memory limitations and scenarios where customers weren’t sure if they were going to get their simulation work done in the desired time frame.

Cloud solving technology is really removed all of those barriers. One of the key advantages to Cloud Premium, one of the advantages that it holds over local solving is that the hardware and software can evolve in real time. This technology has been available for, I’ll give you an example, this technology has been available for customer usage for only a few weeks now, and we’ve already rolled out several performance improvements where we found gains to be had. As the software evolves, as new technologies are added, as more functionality makes it way into Cloud Premium, users will have access to all of this with their current subscription to Autodesk CFD.

As hardware evolves and the cloud network speeds are improved we do expect dramatically improved performance even over what I’ve shared here today. Since that solver technology resolves in the cloud no client updates are necessary to take full advantage of it. In summary, Cloud Premium and Autodesk CFD is the first in product high performance computing cloud solution period. With Cloud Premium users have on demand access to 16 cloud machines each with 16 cores, that’s 256 cores total, all working together to solve a single simulation. And users can expect to see performance gains of at least 400%, and they’ll be able to simulation jobs far larger than what is currently possible on a single local machine or in Cloud Standard.

With that, I’d like to open it up for some Q&A. It looks like we have a number of questions submitted already.

Q: First up, how do I know how many credits I have?

The number of credits you have, the number of cloud credits available to each user, you can find them in your simulation job manager. You can also find them in your Autodesk account. I think it’s manage.autodesk.com. That’ll show you how many credits you have available. If you get low you can always purchase more at any time. Between the simulation job manager and your Autodesk account it will give you an idea of how many credits you have to solve.

Q: What is the number of time steps required to reach convergence, or why is that different between Cloud Standard and Cloud Premium case?

Great question. Probably refers to some of those metrics that I shared most specifically with the race car rear wing assembly. What we did with our … Not we, because I didn’t do it. But what our product development team did was they took out the best of the best from our Cloud Standard solver and rewrote that for a massively parallel architecture. That’s distributing this load over many, many works and many cores on each one of those workers. It is truly a new solver. It is a different solver. It’s solving the same equation that we solved previously, but it is a different approach and a different solver.

The number of iterations required will be different than say when you’re comparing Cloud Standard to Cloud Premium. What we’re finding though, especially for some of those challenging problems in Cloud Standard or a traditional local solve where it might take many hundreds or even thousands of iterations to reach a nice convergence, Cloud Premium is really ideally suited for those problems because it can solve them in a fraction of the number of iterations. It is a different solver. You will likely see different solver iteration counts in your convergence plot. But as far as the equations we’re solving and the level of fidelity and accuracy that is as good or better.

Q: How long does my data stay on the cloud?

Good question. The data that we send up for the cloud, it’s model specific data contains basically your mesh and set up conditions and it’s very similar to what we would solve on your local machine. We don’t send a CAB model up, but we do send your simulation mesh and the associated boundary conditions and all good stuff. As soon as the job is complete it comes right back down to your analysis. In fact, this is a unique way that Cloud Premium works, but when you solve a job in Cloud Premium, as soon as you hit solve, it’ll actually spool up all 16 of those machines. It usually takes around four to six minutes.

You’ll get a little notification in your output window that says, “Spooling jobs takes X number of minutes.” And then as soon as that job is done all of the data sent back to your local machines, and those instances, those AWS instances, are then shut down. All data is removed from them as soon as the job is completed.

Q: How many jobs can I solve at the same time, and can I shut my machine down while solving on the cloud?

To the first one, “How many jobs can I solve at the same time?” That is infinite. There’s probably … I don’t know if we’ve run into a limit with the number of AWS instances that we can spool up at a given time. But most users I think have launched more than one job at a time. But I think the most that I’ve done is around 30. I definitely haven’t hit a limit yet. As far as I know there’s no limit to how many you can solve assuming you have enough cloud credits to support the request.

The second one there, “Can I shut my machine down while solving on the cloud?” Absolutely. Very similar to what happens on a local solve, let’s say you have an analysis. You’re going to get it off on your laptop of your local work station, you can actually close the interface at any time. The interface and the solver are separate entities. They communicate with each other, but they’re not necessary. You don’t have to have the interface up for the solver to do it’s work. It’s just like that in the cloud too. It works the same way with Cloud Standard and Cloud Premium.

Once that job’s submitted you could close your interface down. You could close your lid on your laptop. You could pull the plug. You can turn off your network card. It just doesn’t matter. Once you reconnect your laptop to the internet it will reconnect to the job. If the job’s done it’ll download. If it’s not done it’ll download interim results and as soon as the job completes then the rest of the data will be sent back. You can fully remove your local work station or laptop from the cloud solve. That’s super handy because if you launch a job at 4:55 and you rush home to make dinner or to get dinner or whatever, then you plug your computer back in at 7:30 or eight o’clock then you can check your jobs from really location as long as you have an interact location, or internet access.

Q: Could you comment a little bit about the pricing in general?

Yeah, sure. The pricing tiers, we introduce pricing tiers recently in CFD 2018. Previously, all jobs would cost 15 credits in Cloud Standard. This year there’s a multi tier … There’s a calculation that basically determines which tier you’re in. Basically tier one is 10 credits. Almost all of the jobs that are submitted on the cloud are 10 credit jobs. There’s a medium tier of 30 credits, and then an upper tier of 100 credits, and that’s for Cloud Standard. When you launch a Cloud Premium job there’s a 4 X multiplier on the 10, 30 and 100 credit tiers.

Q: What’s the speed of post processing results on the cloud and sending graphics as well as other information to the local machine?

It’s exactly the same as Cloud Standard. The only thing that changes to the user is that it’s when you hit the solve button and there’s a pull down for whether you want to solve locally, whether you want to solve on Cloud Standard or Cloud Premium. That’s the only difference to the user, the extra option in the pull down step. Everything else is the same. Model set up is the same. Results interaction is the same. If you’re familiar with CFD already, it’ll preform exactly the same way as it did previously.

Q: Do I need to download any other software to use this?

You do not. You do have to have the latest update for CFD 2018, which should be available through your Autodesk download manager. You can always get it on I think it’s on your manage.autodesk.com website. The 2018.1 release grants access to Cloud Premium.

Q: Who has access to my data, and what data is up there?

You have access to your data. We do not. The data that we … We have statistical data. We have log files that contain information about where the machine originated, where it’s solving, information about whether the solve was successful or whether it failed. If a job fails in the cloud we do not charge you for those cloud credits. That’s the sort of data that we gather. But the data, it’s accessible only to you in your local account.

Q: Can a user choose any number of cores, say between 8-256 when using Cloud Premium?

Currently no. That’s a good question. Currently the number of cores that we’re using in Cloud Premium is 256. We recognize that there might be some problems that are more optimal for say 512 cores. There might be some problems that are more optimal for 64 or 128. But as of right now we’re sticking with 256 cores for all Cloud Premium jobs. And what we’re finding is that jobs that are very small, say a million elements, under a million elements, things that would solve relatively quickly on a local machine, they’re not ideally suited to be solved on Cloud Premium.

For one, it’s four times more expensive for probably no performance gains. And two, when you spread out a small job over that many cores, that many systems, it just doesn’t function well with so much communication between all the machines that you don’t really get any performance gain. What we’re finding is that 256 cores is a nice sweet spot for medium to large to very large problems right now.

Speak Your Mind