What do you run simulations on?
 

Hello all - I'm trying to understand what type of hardware CFD-onliners run their simulations on.

Could you please fill out the poll, and if desired, describe your hardware setup with a post? If you have any issues with your current setup, that would be very useful to know as well. Thank you! :)

I think you need to add another category, or even two. There is a world of difference between a 5 node cluster, a 500 node cluster, and a 50,000 node cluster. Much, much more difference than the difference between a 4 core desktop and an 8 core desktop, which your poll does differentiate between. I don't think anyone would consider a 5 node cluster a "large" cluster.

You're correct, kyle - my error. Is there any way to change the poll?

My hypothesis was that most simulations run on HPC desktops or small clusters...do you think that holds?

For this forum audience I definitely think that is true. There seems to be a lot of academics and people in industries like automotive and manufacturing. It's people using either open source or commercial codes. For some reason you rarely find people here working for the big aerospace and defense firms, which all have their own massive clusters and proprietary codes.

Personally I'd like to know how many people are using cloud based hardware for real industrial work. I tried really hard to make it work for my business, but the economics and logistics just don't make any sense. The cost per CPU-hour is just way, way too high even on Amazon's cheapest offerings. Latency between nodes is typically ridiculously high, severely limiting how large you can scale. The huge lag and cost of uploading and downloading data seems to never be mentioned.

It seems to me that businesses or individuals either do so little CFD that hiring a consulting firm is probably their best option, or they do so much that they could pay off the cost of a cluster in less than a year. There is a very narrow range of customers for which these cloud services make sense, yet for some reason there are at least a half dozen startups trying to cater to them.

Good points, kyle. And I wish I could fix that poll! :)

From what I've heard of the cloud, the main argument seems to be about flexibility and "pay by the hour". Using publicly available data from both cloud and on-premise computing vendors, and counting ALL the costs of installing and operating an in-house cluster, the "cloud premium" does not seem that high.

That's actually part the problem I have with how the cloud vendors market their business model. They are very careful to count ALL the costs of maintaining your own hardware while neglecting all logistical benefits, but then pretend that the only cost of using their hardware is the hourly CPU rate.

I ran the calculations for my business. I figured renting CPU time would be the way to go, but it wasn't even close. The simulations I wanted to do would have cost over $100,000 in CPU time. I built a $12000 cluster in a week that was able to do the calculations in less than a month. Yes I spent a few hundred dollars on electricity, and that week of my time is probably worth $2000, but it is still no where in the same ballpark as the cloud cost. Plus we dismantled and sold the hardware six months later for something like $7000. Can't do that with something you rented.