Yup, curl-curl is a big challenge for normal preconditioners. There are two choices that work reliably: edge-based smoothers (Hiptmair; Arnold, Falk, and Winther) and auxiliary space preconditioners. ML and Hypre have support for different variants of the latter, but they require you to provide some custom operators. PETSc hasn't had many direct requests for Maxwell, but we could add an interface without much work (providing a common interface to the auxiliary space methods from ML and Hypre).
Is this frequency-domain? Do you have a purely negative shift or is there damping? The high-frequency low-damping case is notoriously difficult. Jack Poulson and Paul Tsuji have done some of the best recent work on that problem. Their approach is pretty reliable, but it's heavy-weight and requires decomposition into layers with PML boundary conditions. On Thu, Sep 13, 2012 at 6:11 PM, Dmitry Karpeev <[email protected]> wrote: > The curl-curl operator is notoriously difficult because of the form of its > nullspace that the AMG algorithms have to carefully take into account. > From what I understand, ML tries to do something intelligent about > curl-curl, but as far as I know that requires going around the plain > vanilla API that PETSc taps into. Also, currently there is no libMesh API > to take in a nullspace and serve it to PETSc (or Trilinos?), although > that's something that can be easily fixed. > > Jed should be able to comment on the issues with curl-curl and AMG in more > detail. > > Dmitry. > > > On Thu, Sep 13, 2012 at 5:09 PM, Roy Stogner <[email protected]>wrote: > >> >> On Thu, 13 Sep 2012, Derek Gaston wrote: >> >> > I haven't looked at that problem and don't know much about curl-curl >> spaces... but if the matrix would respond to multigrid at all I highly >> > recommend compiling PETSc with Hypre support and using: >> > -pc_type hypre -pc_hypre_type boomeramg >> > >> > For problems where it works it can't be beat... >> >> "Where it works" is the tricky part. It seems to fly on the curl-curl >> problem, which is great; it failed to converge on a navier-stokes >> problem I just tried, which rules it out as a default LIBMESH_OPTIONS >> for me but still makes it a reasonable candidate for further >> consideration... >> >> except that the very *first* time I tried PETSc with boomeramg, like 5 >> years ago, it was "converging" to inaccurate answers on some of the >> runs I tried. I'm 99% sure they'll have fixed whatever was wrong by >> now, but once bitten twice shy... >> --- >> Roy >> >> >> ------------------------------------------------------------------------------ >> Live Security Virtual Conference >> Exclusive live event will cover all the ways today's security and >> threat landscape has changed and how IT managers can respond. Discussions >> will include endpoint security, mobile security and the latest in malware >> threats. http://www.accelacomm.com/jaw/sfrnl04242012/114/50122263/ >> _______________________________________________ >> Libmesh-users mailing list >> [email protected] >> https://lists.sourceforge.net/lists/listinfo/libmesh-users >> > > ------------------------------------------------------------------------------ Everyone hates slow websites. So do we. Make your web apps faster with AppDynamics Download AppDynamics Lite for free today: http://ad.doubleclick.net/clk;258768047;13503038;j? http://info.appdynamics.com/FreeJavaPerformanceDownload.html _______________________________________________ Libmesh-users mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/libmesh-users
