Filippo Leonardi <[email protected]> writes: > I am actually having hard time figuring out where I am spending my time. > > Reading the report I am spending time on KSPSolve and PCApply (e+02). Since > the number of those operations is well under control. I guess is some > communication that is the bottleneck. > > The lines: > VecScatterBegin 4097 1.0 2.5168e+01 3.2 0.00e+00 0.0 2.9e+09 3.7e+01 > 0.0e+00 3 0 87 39 0 10 0100100 0 0 > VecScatterEnd 4097 1.0 1.7736e+02 1.1 0.00e+00 0.0 0.0e+00 0.0e+00 > 0.0e+00 25 0 0 0 0 88 0 0 0 0 0 > are probably what is slowing down the solution.
How large is your problem? > Also, times do not add up properly, especially in KSPSolve. Timings are inclusive, not exclusive. I don't know what is going on with the having the same event appear many times within the stage, but I remember fixing an issue that might have caused that years ago and I'd appreciate it if you would upgrade to the current version of PETSc. How large is your problem size and how many processors are you running on? > PS: until now I was using outputs in VTK. I guess it is better to output in > PEtsc binary? Yes. You can use the binary-appended VTK viewer (please upgrade to current version of PETSc) or the PETSc binary viewer. > Is it better to output from PETSC_COMM_SELF (i.e. each processor > individually)? No, use collective IO.
pgpYaEobmSSuR.pgp
Description: PGP signature
