[petsc-users] KSP Residual for GMRES

[Travis C. Fisher]

I am running a Navier Stokes Solver that uses implicit Runge Kutta for 
timestepping. I am using SNES to perform preconditioned JFNK. As a 
sanity check, I am running in serial and using the FD generated Jacobian 
at the beginning of a time step as the preconditioner matrix for the 
linear solves, and superlu_dist as the preconditioner. This means that 
the linear solve (GMRES) for the first Newton iteration of the first 
stage should converge almost exactly. I do not observe this for large 
time steps. While it does converge to the Eisenstat-Walker convergence 
criterion, the KSP residual norm is not small. It only decreases four 
orders of magnitude. My question is whether or not this is the 
preconditioned residual? The solver spits out the same residual whether 
or not I use left or right preconditioning. I should also mention that 
in this case I have turned off all viscous cross terms and the problem 
is essentially 1-D, so I have taken as much care as possible to ensure 
that the FD jacobian is a proper representation of the system. 
Additionally, I observe the same problem in the inviscid case with both 
the FD jacobian and the exact jacobian that I have computed. I 
appreciate any insight you can give me into this observation.

Thanks,

Travis Fisher