<div dir="ltr">Hi all,<div><br></div><div>I am trying to solve the Helmholtz equation for temperature T:</div><div><br></div><div>(C I + Div D grad) T = f</div><div><br></div><div>in IBAMR, in which C is the spatially varying diagonal entries, and D is the spatially varying diffusion coefficient. I use a matrix-free solver with matrix-based PETSc preconditioner. For the matrix-free solver, I use gmres solver and for the matrix based preconditioner, I use Richardson ksp + Jacobi as a preconditioner. As the simulation progresses, the iterations start to increase. To understand the cause, I set D to be zero, which results in a diagonal system:</div><div> </div><div>C T = f.</div><div><br></div><div>This should result in convergence within a single iteration, but I get convergence in 3 iterations.</div><div><br></div><div><p style="margin:0px;font-stretch:normal;font-size:15px;line-height:normal;font-family:Monaco;color:rgb(242,242,242);background-color:rgba(0,0,0,0.85)"><span style="font-variant-ligatures:no-common-ligatures">Residual norms for temperature_ solve.</span></p>
<p style="margin:0px;font-stretch:normal;font-size:15px;line-height:normal;font-family:Monaco;color:rgb(242,242,242);background-color:rgba(0,0,0,0.85)"><span style="font-variant-ligatures:no-common-ligatures"><span class="gmail-Apple-converted-space"> </span>0 KSP preconditioned resid norm 4.590811647875e-02 true resid norm 2.406067589273e+09 ||r(i)||/||b|| 4.455533946945e-05</span></p>
<p style="margin:0px;font-stretch:normal;font-size:15px;line-height:normal;font-family:Monaco;color:rgb(242,242,242);background-color:rgba(0,0,0,0.85)"><span style="font-variant-ligatures:no-common-ligatures"><span class="gmail-Apple-converted-space"> </span>1 KSP preconditioned resid norm 2.347767895880e-06 true resid norm 1.210763896685e+05 ||r(i)||/||b|| 2.242081505717e-09</span></p>
<p style="margin:0px;font-stretch:normal;font-size:15px;line-height:normal;font-family:Monaco;color:rgb(242,242,242);background-color:rgba(0,0,0,0.85)"><span style="font-variant-ligatures:no-common-ligatures"><span class="gmail-Apple-converted-space"> </span>2 KSP preconditioned resid norm 1.245406571896e-10 true resid norm 6.328828824310e+00 ||r(i)||/||b|| 1.171966730978e-13</span></p>
<p style="margin:0px;font-stretch:normal;font-size:15px;line-height:normal;font-family:Monaco;color:rgb(242,242,242);background-color:rgba(0,0,0,0.85)"><span style="font-variant-ligatures:no-common-ligatures">Linear temperature_ solve converged due to CONVERGED_RTOL iterations 2</span></p></div><div><br></div><div>To verify that I am indeed solving a diagonal system I printed the PETSc matrix from the preconditioner and viewed it in Matlab. It indeed shows it to be a diagonal system. Attached is the plot of the spy command on the printed matrix. The matrix in binary form is also attached. </div><div><br></div><div>My understanding is that because the C coefficient is varying in 4 orders of magnitude, i.e., Max(C)/Min(C) ~ 10^4, the matrix is poorly scaled. When I rescale my matrix by 1/C then the system converges in 1 iteration as expected. Is my understanding correct, and that scaling 1/C should be done even for a diagonal system?</div><div><br></div><div>When D is non-zero, then scaling by 1/C seems to be very inconvenient as D is stored as side-centered data for the matrix free solver. </div><div><br></div><div>In the case that I do not scale my equations by 1/C, is there some solver setting that improves the convergence rate? (With D as non-zero, I have also tried gmres as the ksp solver in the matrix-based preconditioner to get better performance, but it didn't matter much.)</div><div><br></div><div><br></div><div>Thanks,</div><div>Ramakrishnan Thirumalaisamy</div><div>San Diego State University.</div></div>