[Eman.]

Hello everyone,

when I try to restart the adjoint solver in parallel, it seems as if it ignores the adjoint restart file and recomputes the solution all over again (the residuals are in the order of 10E-2). However, when I run it in serial using "SU2_CFD_AD" (or using "parallel_computation.py -n1 -f config.cfg") everything seems to be alright and the initial residuals are around 10E-9. Any ideas what the problem could be?

Many thanks in advance.

[fpalacios]

Hi, thanks for using SU2!

SU2 counts with a very sophisticated Automatic Differentiation (the entire code has been differentiated ) using CoDIPack that also includes an Adjoint MPI

http://www.scicomp.uni-kl.de/software/codi/

However, because of the intrinsic nature of the AD it is not possible to perform a perfect restart at the boundaries of the different MPI Domains... so, when you restart it takes several iterations to converge the domain boundaries again.

However, I have seen several improvements in the develop branch related with this topic. If I were you, I'll try the same problem with the develop branch version of SU2.

Thanks,

Francisco

[Eman.]

Originally Posted by fpalacios

Hi, thanks for using SU2!

SU2 counts with a very sophisticated Automatic Differentiation (the entire code has been differentiated ) using CoDIPack that also includes an Adjoint MPI

http://www.scicomp.uni-kl.de/software/codi/

However, because of the intrinsic nature of the AD it is not possible to perform a perfect restart at the boundaries of the different MPI Domains... so, when you restart it takes several iterations to converge the domain boundaries again.

However, I have seen several improvements in the develop branch related with this topic. If I were you, I'll try the same problem with the develop branch version of SU2.

Thanks,

Francisco

Thanks a lot for the perfect explanation. I'll run the same simulation with SU2-devel.