Hmm, could be a bug, could be the algorithm. Run with -ksp_view and send the output
What problem are you solving, is it a PETSc example? Barry On Jan 17, 2011, at 8:31 PM, li.luo at siat.ac.cn wrote: > Hi, > > > I met a problem when testing some examples on PETSc-dev for GPU computing, > that is, if one proc pair(1GPU-1CPU) is used, the grid size can be enlarged > to even 2048*2048, the memory limitation; > however, if more than one proc pairs are use, for example 4GPU-4CPU, the grid > size is limited to about 200*200, if larger, the ksp solver would not > converge. The same problem happens to 8GPU-8CPU limited by 500*500 or other > size. > > I wonder whether you have the same problem? Any error happens to type > MPICUDA? > > Regards, > Li Luo > > >>> # Machine type: >>> CPU: Intel(R) Xeon(R) CPU E5520 >>> GPU: Tesla T10 >>> CUDA Driver Version: 3.20 >>> CUDA Capability Major/Minor version number: 1.3 >>> >>> # OS Version: >>> Linux console 2.6.18-128.el5 #1 SMP Wed Dec 17 11:41:38 EST 2008 x86_64 >>> x86_64 x86_64 GNU/Linux >>> >>> # PETSc Version: >>> #define PETSC_VERSION_RELEASE 0 >>> #define PETSC_VERSION_MAJOR 3 >>> #define PETSC_VERSION_MINOR 1 >>> #define PETSC_VERSION_SUBMINOR 0 >>> #define PETSC_VERSION_PATCH 6 >>> #define PETSC_VERSION_DATE "Mar, 25, 2010" >>> #define PETSC_VERSION_PATCH_DATE "Thu Dec 9 00:02:47 CST 2010" >>> >>> >>> # MPI implementation: >>> ictce3.2/impi/3.2.0.011/ >>> >>> # Compiler: >>> >>> >>> # Probable PETSc component: >>> run with GPU >>> # Configure >>> ./config/configure.py --download-f-blas-lapack=1 >>> --with-mpi-dir=/bwfs/software/mpich2-1.2.1p1 --with-shared-libraries=0 >>> --with-debugging=no --with-cuda-dir=/bwfs/home/liluo/cuda3.2_64 >>> --with-thrust-dir=/bwfs/home/liluo/cuda3.2_64/include/thrust >>> --with-cusp-dir=/bwfs/home/liluo/cuda3.2_64/include/cusp-library > > > > > >
