The discrete adjoint method uses too large memory.

Yminjo · April 26, 2016, 12:31

Dear SU2 developers,

Hi, I'm testing the discrete adjoint method implimented in SU2 v4.1.
When I try to use the method for RANS computation with k-w SST turbulence model and a grid with a total of 0.6 million points, the computation uses over 130Gb memory, so it is directly stopped as the amount is over the physical memory.
The pure CFD computation uses about 20Gb, and the continuous adjoint method uses about 27Gb.

Is that a normal situation, because the discrete adjoint is based on the external AD?
Or, is there any memory leak?

Thank you in advance.

talbring · April 27, 2016, 10:52

Hi Yeongmin,

the discrete adjoint solver takes about 4-7 times more memory than the flow solver.

But 20GB for 0.6 Mio points for the pure CFD computation seems to be way to high. Are you sure about these numbers ?

Tim

April 26, 2016, 12:31	The discrete adjoint method uses too large memory.	#1
Yminjo New Member Yeongmin Jo Join Date: Dec 2013 Posts: 10 Rep Power: 12	Dear SU2 developers, Hi, I'm testing the discrete adjoint method implimented in SU2 v4.1. When I try to use the method for RANS computation with k-w SST turbulence model and a grid with a total of 0.6 million points, the computation uses over 130Gb memory, so it is directly stopped as the amount is over the physical memory. The pure CFD computation uses about 20Gb, and the continuous adjoint method uses about 27Gb. Is that a normal situation, because the discrete adjoint is based on the external AD? Or, is there any memory leak? Thank you in advance. Last edited by Yminjo; April 26, 2016 at 23:25.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Problems launching adjoint	max_	SU2	2	March 12, 2016 11:32
Help for learning about Adjoint Method in optimization	EmadTandis	Main CFD Forum	0	March 15, 2013 04:55
Adjoint method for shape optimization	Edris	Main CFD Forum	6	November 1, 2012 08:10
Anybody has CFD codes used Discrete Vortex Method	LuckyJ	Main CFD Forum	7	August 6, 2004 07:48
Discrete element method?	Ki Ha Hong	Main CFD Forum	1	May 28, 1999 09:46

April 27, 2016, 10:52		#2
talbring Super Moderator Tim Albring Join Date: Sep 2015 Posts: 195 Rep Power: 11	Hi Yeongmin, the discrete adjoint solver takes about 4-7 times more memory than the flow solver. But 20GB for 0.6 Mio points for the pure CFD computation seems to be way to high. Are you sure about these numbers ? Tim