From: John P. <pet...@cf...> - 2007-03-24 15:42:28
|
Hi, I recently checked in code to the Mesh partitioners to set the processor IDs of the nodes. Are you working with the CVS libmesh? 65536 is just a default which means 'not set'. -John Ingo Schmidt writes: > My idea is to check the processor ID > ((mesh.node(res_node_no-1)).processor_id()) of the result node and > compare it with libMesh::processor_id() before writing the nodal data to > the output file (using the previous version of ex8). > Stupidly the processor ID of the node is 65536. Does libMesh actually > assign the proc ID to the DofObject? > > regards > > Ingo > > Ingo Schmidt schrieb: > > Hi, > > > > at first I was glad that the bug fix is working (i.e. compiling and > > running the code in parallel) but yesterday I checked the benefit of > > parallelisation. > > The benefit is unfortunately is an impairment. Increasing the number of > > CPU's slows down the computation time in a considerable measure. > > E.g. > > example 8: 1 CPU approx. 9 seconds > > 2 CPU's approx. 14 seconds > > > > That's pretty bad... > > > > best regards > > > > Ingo > > > > Roy Stogner schrieb: > > > >> On Mon, 19 Mar 2007, Roy Stogner wrote: > >> > >> > >>> I'll commit a fix to CVS this afternoon, but hopefully that's enough > >>> info to get your own code back on track. > >>> > >> I may not be familiar enough with MPI or our NumericVector API, but I > >> don't actually see how to write an efficient, simple fix. So, I've > >> committed an inefficient and convoluted fix to CVS instead. > >> > >> I'll also see about making it more convenient to run these examples in > >> parallel; I can't believe that nobody has ever run ex8 in parallel > >> under METHOD=dbg before. My application Makefiles use a LIBMESHRUN > >> environment variable to find my own "mpirun -whatever" invocation, and > >> a LIBMESHOPTIONS environment variable to hold extra PETSc options to > >> be added to the end of commands; if there aren't any objections, I'll > >> make the example Makefiles check the same variables, so there will be > >> at least one person regularly trying "make run_examples" in parallel. > >> --- > >> Roy > >> > > > > > > -- > Dipl.-Ing. Ingo Schmidt > Institute of Modelling and Computation > > Hamburg University of Technology tel: +49/(0)40/42878-4483 > Denickestr.17 fax: +49/(0)40/42878-4353 > Building L/ room: 3032 e-mail: ingo.schmidt@TUHH.de > 21075 Hamburg > > http://www.mub.TU-Harburg.de > |