Back in May 2019 we surveyed the CFD and meshing communities about the computer workstations on which they currently run their mesh generation software and on their need for meshing software to be supported on HPC platforms. It has taken a while, but the results have been compiled, charted, and are presented here as promised.
Before delving into the data it’s fair to state that these results aren’t statistically valid for several reasons. However, they do provide a little insight into what computer hardware people are using for their CFD meshing software.
Given that question #1 asked for the respondent’s email address (which we collect solely to ensure one response per participant), let’s proceed with the data. Recall that we asked about the main workstation the respondent used to generate meshes.
Q2: What mesh generation software do you use?
Nearly two-thirds of the respondents are Pointwise users. However, the results shown below are for all respondents regardless of what meshing software they use.
Q3: What OS does your workstation run?
Nearly twice as many respondents use Linux versus Windows. Historically, Pointwise users have predominantly used Windows so this is an interesting result for us. Also, Apple products still haven’t caught on much for CFD.
Q4: What version of Windows?
The Windows users predominantly use Windows 10 which really isn’t surprising. Those folks still running Windows 7 need to be aware that Microsoft is ending support in January 2020. And what other versions of Windows might people still be using? I have a laptop that still boots to Windows 95 but that’s just me.
Q5: Which Linux distribution?
While not judging the relative merits of one flavor of Linux over another, I was surprised by the popularity of Ubuntu.
Q6: What version of macOS?
The dataset of macOS respondents is so relatively small I’m not certain any conclusions can be drawn from these results.
Q7: How many compute cores?
Regardless of the operating system, it seems that a good portion of respondents have in excess of 16 compute cores available. This is good news for a multi-threaded application like Pointwise.
Q8: How much RAM?
Given that (in my opinion at least) RAM gives you the most benefit for your computing dollar, I’m surprised that one third of the respondents have less than 32 GB on their workstation. I notice also that 64 GB seems to be the median with half of the respondents having less, half more.
Q9: What graphics card?
NVIDIA is clearly the leader among the respondents. Given that we advise Pointwise customers to have a professional graphics card with hardware accelerated OpenGL, I can only assume the Intel graphics respondents use other meshing software. However, this all maybe changing soon as OpenGL evolves and macOS deprecates support for OpenGL in favor of their own graphics language.
Q10: How much GPU RAM?
Much the same as plain ol’ RAM, GPU RAM is probably the best bang for your computing dollar, especially if your software uses GPU acceleration. We did a pilot implementation about a decade ago that showed both good acceleration and an unacceptable degree of complication for the source code. We may go back and revisit that work given the advancements in how one programs for a GPU.
It’s also interesting that our literature survey for the paper at AIAA Aviation on the last five years of progress in mesh generation showed little use of GPUs by meshing software. A lot by CFD software, but little by meshing.
Q11: Do you want to mesh on HPC?
Question #11 segues from the workstation people are using now to potential future use of a high performance computing (HPC) platform for mesh generation.
About half the respondents have an HPC platform they could start using today.
There are basically two reasons why you’d want to use an HPC platform. You either want the speed (i.e. get your current jobs done faster) or you want the memory so you can run bigger jobs than you currently can (i.e. so-called “exascale” jobs).
More respondents indicated they wanted the speed.
(Note: The percentages don’t add to 100% because respondents could choose one or the other or both.)
We asked a couple of demographics questions to get an idea about where the respondents were coming from.
Statistics are like swim suits. What they reveal is interesting but what they conceal is important. Regardless, there are a few takeaways. Several conclusions are based, as my attorney friends would say, on “facts not in evidence.” I have supplemented the survey data with other information of which I am aware.
- Use of Linux platforms for meshing is rising.
- Use of Linux distributions other than Red Hat is rising.
- The next several years will reveal the future of graphics languages beyond OpenGL. This is exacerbated by the fact that Apple’s OpenGL support is years behind the current version and the fact that they’ve deprecated support for OpenGL in favor of their own language. How can software authors efficiently support multiple platforms?
- Opportunities abound for gaining performance through the use of threading.
- GPUs offer potential but it’s unclear why they are not more widely used for mesh generation software.
- Use of HPC platforms is motivated largely by the need for speed. Which tells me that a sufficiently fast workstation implementation may negate the need for HPC support. Of course, exascale meshes scream for HPC’s distributed memory.
What do you think we should ask the meshing world about next? I know what I want to ask but I’m hoping you have better ideas. Share them in the comments below.
Also, how do you think we might get more survey respondents? I realize mesh generation isn’t that exciting of a topic unto itself to get thousands of replies. Do we need to give away a prize or prizes? Are engineers really that motivated by t-shirts? Let me know in the comments below.
I think that asking a forum like CFD-online might get you a lot more hits.
I don’t think prizes are worth it but that’s just me.
WE are working with a GPU company on a CFD solver running on GPUs. We are also looking into GPU based meshing. The issue is that we would need a highly parallelizable algorithm to take advantage of hte 100s of GPU cores.
Has Pointwise looked into such algorithms
Thanks, Jim. I had not considered asking the CFD Online community to respond to our questions. As for GPU, our experience is limited to using them on one particular aspect of one particular algorithm. We achieved fairly good (40x) speedups but the code complexity would’ve been a maintenance nightmare. Because this was over 10 years ago, we want to go back and revisit how the programming languages have evolved.
Thanks for reading and commenting.
Reblogged this on Nguoidentubinhduong.
Pingback: This Week in CFD | Another Fine Mesh