The new outputs are attached. The same problem was run with arch-master-opt
The overall time was cut in half, but %T remain roughly the same...
Thanks,
Antoine
Hi Sir,
MatScale in "Main Stage" is indeed called 6 times for 0% run time. In stage "Stiff_Adj" though, we get :
MatScale 8192 1.0 7.1185e+01 1.0 3.43e+10 1.0 0.0e+00 0.0e+00 0.0e+00 50 46 0 0 0 80 98 0 0 0 482
MatMult is indeed expensive (23% run time) and should be improved, but MatScale in "Stiff_Adj" is still taking 50% run time
I was a little surprised that MatScale gets only 450 MFlops. However, it looks like you are running the debugging version of PETSc. Could you configure a version without debugging:
$PETSC_DIR/$PETSC_ARCH/lib/petsc/conf/reconfigure-$PETSC_ARCH.py --with-debugging=0 --PETSC_ARCH=arch-master-opt
and rerun the timings?
Thanks,
Matt Thanks,
Antoine
MatMult 9553 1.0 3.2824e+01 1.0 3.54e+10 1.0 0.0e+00 0.0e+00 0.0e+00 23 48 0 0 0 61 91 0 0 0 1079 MatScale 6 1.0 5.3896e-02 1.0 2.52e+07 1.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 467
Though the flop rate of MatScale is not so high (467) it is taking very little (0 percent of the run time while MatMult takes 23 percent of the time).
So the main cost related to the matrices is MatMult because it has a lot of operations 9553, you might think about your algorithms you are using and if there improvements.
It looks like you are using some kind of multigrid and solve 6 problems with 1357 total iterations which is 200 iterations per solve. This is absolutely HUGE for multigrain, you need to tune the multigrid for you problem to bring that down to at most a couple dozen iterations per solve.
Barry
Hi,
See attached files for both outputs. Tell me if you need any clarification. It was run with a DMDA of 33x17x17 nodes (creating 32x16x16=8192 elements). With 3 dof per nodes, problem has a total of 28611 dof.
Note : Stage "Stiff_Adj" is the part of the code modifying Mat K. PetscLogStagePush/Pop was used.
Regards,
Antoine
Hi,
I'm working with a 3D DMDA, with 3 dof per "node", used to create a sparse matrix Mat K. The Mat is modified repeatedly by the program, using the commands (in that order) :
MatZeroEntries(K)
In a for loop : MatSetValuesLocal(K, 24, irow, 24, icol, vals, ADD_VALUES) MatAssemblyBegin(K, MAT_FINAL_ASSEMBLY)
MatAssemblyEnd(K, MAT_FINAL_ASSEMBLY)
MatDiagonalScale(K, vec1, vec1)
MatDiagonalSet(K, vec2, ADD_VALUES)
Computing time seems high and I would like to improve it. Running tests with "-log_view" tells me that MatScale() is the bottle neck (50% of total computing time) . From manual pages, I've tried a few tweaks : - DMSetMatType(da, MATMPIBAIJ) : "For problems with multiple degrees of freedom per node, ... BAIJ can significantly enhance performance", Chapter 14.2.4
- Used MatMissingDiagonal() to confirm there is no missing diagonal entries : "If the matrix Y is missing some diagonal entries this routine can be very slow", MatDiagonalSet() manual
- Tried MatSetOption()
- MAT_NEW_NONZERO_LOCATIONS == PETSC_FALSE : to increase assembly efficiency
- MAT_NEW_NONZERO_LOCATION_ERR == PETSC_TRUE : "When true, assembly processes have one less global reduction"
- MAT_NEW_NONZERO_ALLOCATION_ERR == PETSC_TRUE : "When true, assembly processes have one less global reduction"
- MAT_USE_HASH_TABLE == PETSC_TRUE : "Improve the searches during matrix assembly"
According to "-log_view", assembly is fast (0% of total time), and the use of a DMDA makes me believe preallocation isn't the cause of performance issue.
I would like to know how could I improve MatScale(). What are the best practices (during allocation, when defining Vecs and Mats, the DMDA, etc.)? Instead of MatDiagonalScale(), should I use another command to obtain the same result faster?
Something is definitely strange. Can you please send the output of
-log_view -info :mat
Thanks,
Matt Thank you very much!
Antoine Côté
-- 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
-- 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
|
vecscale.patch