Re: [Nek5000-users] Proper setup for AMG solver
Dear Nek experts, I wanted to add some remarks on my performance tests conducted last week mostly for the turbulent pipe flow at Re_b=5300 - The bad performance of the AMG solver was indeed due to an old version of amg_hypre, as Stefan mentioned already. For Re_b=5300 XXT and AMG showed the same performance. - Additionally, I could increase the performance by about 20% using the following settings: -- lxd=10 (instead of 12, for lx1=8) -- lx2=lx1-0 (PN-PN) -- lower tolerances for p=1e-5 (instead of 1e-8) and 1e-6 for velocity and passive scalars -- turning of projection except for the fields of low Prandtl number - Going from BDF3/EXT3 with a variable DT and targetCFL=0.5 to BDF2/OIFS with variable DT and targetCFL=0.5 resulted in a longer time / timestep (3x) but a also a larger average DT (6x). Assuming the collected statistics require a similar averaging time in wash-outs, BDF2/OIFS is advantageous (even for 10 passive scalars). Before this discussion, I was running at constant DT and collected statistics every 10th step. Now with a larger DT using characteristics, should I collect statistics for each step? - Regarding the filterWeight and filterCutoffRatio, I found out (with Stefan's help) that the divergence error in L2 norm (when using PN-PN) is affected by filterWeight. For the setup at Re_b=5300 a filterWeight of 27 seemed to be OK considering a divergence error in L2 norm of L2=3e-2, whereas a weight of 54 resulted in L2=1e-1. How these settings will affect the statistics, I am testing now. Thank you all for your help and suggestions. Cheers, Steffen
Hi Steffen, Thanks for the note. I would recommend fixed dt at a value near your target CFL. There are several reasons - 1. I doubt you gain that much with variable dt 2. I personally don't like Nek's auto-dt selection scheme 3. You have to reorthogonalize the projection basis every time dt changes (expensive) Paul ________________________________ From: Nek5000-users <[email protected]> on behalf of [email protected] <[email protected]> Sent: Monday, September 10, 2018 9:19:00 AM To: [email protected] Subject: Re: [Nek5000-users] Proper setup for AMG solver Dear Nek experts, I wanted to add some remarks on my performance tests conducted last week mostly for the turbulent pipe flow at Re_b=5300 - The bad performance of the AMG solver was indeed due to an old version of amg_hypre, as Stefan mentioned already. For Re_b=5300 XXT and AMG showed the same performance. - Additionally, I could increase the performance by about 20% using the following settings: -- lxd=10 (instead of 12, for lx1=8) -- lx2=lx1-0 (PN-PN) -- lower tolerances for p=1e-5 (instead of 1e-8) and 1e-6 for velocity and passive scalars -- turning of projection except for the fields of low Prandtl number - Going from BDF3/EXT3 with a variable DT and targetCFL=0.5 to BDF2/OIFS with variable DT and targetCFL=0.5 resulted in a longer time / timestep (3x) but a also a larger average DT (6x). Assuming the collected statistics require a similar averaging time in wash-outs, BDF2/OIFS is advantageous (even for 10 passive scalars). Before this discussion, I was running at constant DT and collected statistics every 10th step. Now with a larger DT using characteristics, should I collect statistics for each step? - Regarding the filterWeight and filterCutoffRatio, I found out (with Stefan's help) that the divergence error in L2 norm (when using PN-PN) is affected by filterWeight. For the setup at Re_b=5300 a filterWeight of 27 seemed to be OK considering a divergence error in L2 norm of L2=3e-2, whereas a weight of 54 resulted in L2=1e-1. How these settings will affect the statistics, I am testing now. Thank you all for your help and suggestions. Cheers, Steffen _______________________________________________ Nek5000-users mailing list [email protected] https://lists.mcs.anl.gov/mailman/listinfo/nek5000-users
participants (1)
-
nek5000-users＠lists.mcs.anl.gov