I see. Well the reason for non-symmetry is that I'm using a finite difference discretization on quadtree AMR grids. When discretizing at a hanging node, i.e. a node that is missing a direct neighbor, I use interpolation from nodes of the parent cell. When it comes to discretizing at those nodes (i.e. nodes of the parent), however, they are on a coarser grid and will have direct neighbors so they would not see the hanging node in their discretization. This leads to a non-symmetric discretization matrix which I was using (structurally) to define the adjacency matrix of the graph. I think I can fix this by requiring the parent nodes to also have a link to the hanging node (although they are not used in the discretization anyway)
I'm just surprised that ParMetis does not complain about non-symmetric graphs at lower resolutions and partitions them just fine! As for the debugger, -start_in_debugger does not start the debugger for me :(. I'm using mpirun -np 2 ./a.out -start_in_debugger. Is this the correct way? On Wed, Apr 11, 2012 at 3:07 PM, Matthew Knepley <knepley at gmail.com> wrote: > On Wed, Apr 11, 2012 at 8:24 PM, Mohammad Mirzadeh <mirzadeh at > gmail.com>wrote: > >> I was reading the FAQ list when I came across the following: >> >> http://www.mcs.anl.gov/petsc/documentation/faq.html#key >> >> When calling MatPartitioningApply() you get a message Error! Key 16615 >> not found The graph of the matrix you are using is not symmetric. You >> must use symmetric matrices for partitioning. >> >> Is this a limitation on ParMetis side? I set up the adjacency matrix >> based on the discretization that I will be performing on the grid which is >> non-symmetric; both numerically and structurally. What's the solution here? >> Make an "approximate" adjacency matrix that sort of looks like >> (structurally) my discretization but is symmetric? What I don't understand >> is my matrix IS non-symmetric when the code runs on coarser grids! >> > > I don't quite understand how you can have a non-symmetric adjacency > description. But Metis/ParMetis partitions undirected graphs, which > by definition have symmetric adjacency matrices. Non-symmetric adjacency > would seem to imply a directed graph of some sort, or more plainly, > something is a adjacent to another thing which is not adjacent to it. That > is a very strange concept. > > >> Also, I was reading the FAQ hoping I can find something regarding using >> gdb in parallel. I found this: http://scicomp.stackexchange.com/a/410/485but >> I'm not sure how I should be using gdb in parallel. Could you (maybe >> Matt?) please explain a little bit? >> > > -start_in_debugger spawns a gdb windows for EVERY process and attaches it > > -debbuger_nodes a,b,c spawns gdb windows ONLY for ranks a, b, and c > > Thanks, > > Matt > > >> Thanks >> >> On Tue, Apr 10, 2012 at 12:08 AM, Mohammad Mirzadeh <mirzadeh at >> gmail.com>wrote: >> >>> Just built petsc-dev and it did not help. I'm going to look into the >>> code to see if my graph is "ill-formed" in some sense. Just hope the >>> problem is from my side not a real bug in ParMetis! >>> >>> >>> >>> On Mon, Apr 9, 2012 at 8:46 PM, Mohammad Mirzadeh <mirzadeh at >>> gmail.com>wrote: >>> >>>> ok. Thanks Jed. I'll try petsc-dev to see if it fixes the problem. >>>> >>>> Thanks everyone. >>>> >>>> >>>> On Mon, Apr 9, 2012 at 8:42 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote: >>>> >>>>> On Mon, Apr 9, 2012 at 22:37, Mohammad Mirzadeh <mirzadeh at >>>>> gmail.com>wrote: >>>>> >>>>>> Thanks Sean. I'm using Petsc 3.2-p6 along with ParMetis 4.0.2. Since >>>>>> this was not supported with 3.2-p6, and previous versions had bugs, I >>>>>> built >>>>>> parmetis myself and used --with-parmetis-include and --with-parmetis-lib >>>>>> flags to build petsc. >>>>>> >>>>>> Should I switch to petsc-dev? >>>>>> >>>>> >>>>> Yes, and use --download-metis --download-parmetis because the version >>>>> upstream has some bugs for which the patches have not been applied. >>>>> >>>> >>>> >>> >> > > > -- > What most experimenters take for granted before they begin their > experiments is infinitely more interesting than any results to which their > experiments lead. > -- Norbert Wiener > -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120411/70bf228c/attachment.htm>
