SU2 RANS Performance
I have recently been playing around with SU2 and have a couple of questions about the RANS cases.
First of all, how long does the turbulent oneram6 case supposed to take using the turb_ONERAM6.cfg file? For me it took about 13,000 iterations to drop the residual down to 1e-6, or about 3.75 hours of wall time on a single processor. It looks like there was no multigrid used however. Does the multigrid help much in this case? Are they any other options that would significantly speed up with calculation or is this performance typical?
For comparison, I ran a 128,000 cell M6 grid though SUmb (JST scalar, SA, using 3w MG without MG startup), on a single processor down to 1e-6 and that took about 35 minutes on a single processor.
Considering that SU2 grid has about 40,000 cells, that would make SUmb, over an order of magnitude faster. Is this sort of performance differential expected?
Thanks and keep up the great work!
Thanks a lot your your kind email. As you know, to find the right parameters in a CFD solver is an Art. For this particular problem, please find attached my convergence and the config file that I'm using. On a single processor, it took me 642 iterations and 11.6 min of wall time to drop the residual to 1e-6 using the current developers version of SU2 (https://github.com/su2code).
An apples to apples comparison with other solvers is very important for us, and we will participate in a special session during the next SciTech in which SU2 will be compared with the most important solvers in US.
Anyway, shape design, it multidisciplinarity capability, and flexibility are the most important characteristics of SU2. Its object oriented structure is very useful despite the performance could be inferior to a pure C code. For that reason we have a collaboration with Intel to improve the code vectorization and exploit all the capabilities of an Intel Xeon Phi platform (http://www.intel.com/content/www/us/...hi-detail.html).
In short, as far I know, the algorithmic implemented in SU2 is as good (or better) than the one implemented in SUMB. As you know SUMB uses pure C and a structured data storage (that is a plus in performance). However, as you know, when you are interested in complex geometries a structured grid could be orders of magnitude bigger than an unstructured one (assuming that it is possible to generate the multiblock structure grid without overset).
Thanks a lot for using our code and your encouraging words,
|All times are GMT -4. The time now is 12:29.|