I found a similar problem solving Laplace's equation: using MatCreate took forever, whereas using MatCreateAIJ instead the time for MatSetValues was essentially negligible. I set up MatCreate with a call to MatSetSizes. ...Peter
On May 17, 2013, at 2:50 PM, Matthew Knepley <[email protected]> wrote: > On Fri, May 17, 2013 at 3:40 PM, Joon hee Choi <[email protected]> wrote: > My petsc version is 3.3.5. > And I made nnz using the number of each row. When I wrote the matrix using > MatSetValues, I wrote from low row and column to high row and column. > However, when I changed the order writing the matrix, it took much more time > (up to 3hrs). So I am concerned that the slowness is because the big size of > matrix (2.6*10^7, 1.248*10^15) is related to reading from or writing to the > cache, ram, or hard. > > No, it is bad preallocation. This should be simple to fix. First, turn on > errors > > MatSetOption(MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE) > > and second run with -info. > > Thanks, > > Matt > > Anyway, the following code is the part that I read the data from a file and > set up tuples and nnz: > > FILE *fp = fopen("data.txt", "r"); > while (fscanf(fp, "%d %d %d %d", &x, &y, &z, &v) == 4) > { > tups.push_back(std::tr1::make_tuple (x-1, z-1, y-1, v)); > nzrow[i-1] += 1; > if (x > X) X = x; > if (y > Y) Y = y; > if (z > Z) Z = z; > } > fclose(fp); > PetscMalloc(X*sizeof(PetscInt), &nnz); > memset(nnz, 0, X); > for (itnz=nzrow.begin(); itnz!=nzrow.end(); ++itnz) { > nnz[itnz->first] = itnz->second; > } > sort(tups.begin(), tups.end()); > > If my code is wrong, then please let me know. > > Thank you, > Joon > > ----- Original Message ----- > From: "Jed Brown" <[email protected]> > To: "Joon hee Choi" <[email protected]> > Cc: [email protected] > Sent: Friday, May 17, 2013 3:24:09 PM > Subject: Re: [petsc-users] 3-dimension to matrix > > Joon hee Choi <[email protected]> writes: > > > Thank you for your fast reply. Your last comments looks like my first > > code. The full size of the matrix is (X, Y*Z). Also, I used SEQAIJ as > > matrix type and (X, Y) as block size. Also, I implemented > > SeqAIJpreallocating with nnz. Nevertheless, it was very slow. The > > Matrix-Set-Up part of the first code is as follows: > > > > sort(tups.begin(), tups.end()); > > MatCreate(PETSC_COMM_SELF, &A); > > MatSetType(A, MATSEQAIJ); > > MatSetSizes(A, PETSC_DECIDE, PETSC_DECIDE, X, Y*Z); > > MatSetBlockSizes(A, X, Y); > > MatSeqAIJSetPreallocation(A, PETSC_DEFAULT, nnz); > > > > sz = tups.size(); > > for (i=0; i<sz; i++) { > > x = std::tr1::get<0>(tups[i]); > > y = std::tr1::get<2>(tups[i]) + std::tr1::get<1>(tups[i])*Y; > > val = std::tr1::get<3>(tups[i]); > > MatSetValues(A, 1, &x, 1, &y, &val, INSERT_VALUES); > > } > > MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY); > > MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY); > > > > > > I used tuple (x, z, y, value) and vector of c++. I didn't get any > > errors from this code. However, it took about 9 minutes in this > > part. > > What version of PETSc? Your preallocation was almost certainly not > sufficient. > > > > -- > 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 ________________ Peter Lichtner Santa Fe, NM 87507 (505) 692-4029 (c) OFM Research/LANL Guest Scientist
