On 2014-11-14 15:33, Jørgen Kvalsvik wrote:
I have an update and a few questions regarding my project. For
introduction and details, please see
http://www.opm-project.org/pipermail/opm/2014-October/000664.html
Thanks a lot for the update. I haven't digested the message yet and I
don't feel qualified to answer just yet so I'll comment on just one point.
Using my SparseMatrixBuilder, now modified to rely on std::map instead
of std::vector
As a general rule I usually distrust std::map<> when it comes to
handling anything that has to do with computational performance and
inner loops. Std::map<> is to a very great extent a "pointer focused"
container (tree-like structure internally). What's more discouraging
still is that std::unordered_map<> isn't all that much better though
there is some improvement. You can look up
Chandler Carruth CppCon 2014
on YouTube for some stories. Mr. Carruth is a Clang and LLVM developer
at Google.
Also, the conversion from coordinate format to CSR (or CSC for that
matter) is a solved problem that is O(nnz) complexity--including
optionally generating sorted rows (columns) without repeated indices.
You don't actually need the std::map<> at all, just the triplets (in
whatever form). This just means that there is further potential for
performance improvements. That's good.
Still
I am able to remove all allocation code in the
IncompFlowSolverHybrid. I consider this a win, because it
significantly simplifies setting up sparse matrices and leaks fewer
implementation details
That's good too! As the person who perpetrated that code in the first
place I can attest that it was rather painful to write and difficult to
extend.
I am able to reduce the running time of
upscale-benchmark-relperm. I attribute this to a more efficient
allocation and instantiation of the matrices, as the BCRSMatrix now can
be build row-wise and not in the more inefficient random mode.
Do keep in mind, however, that per-row construction leads to a data
structure in which the matrix values are stored in separate containers
per row whereas the original scheme uses a single contiguous array for
the values. That may or may not matter in any particular application.
Bård
_______________________________________________
Opm mailing list
[email protected]
http://www.opm-project.org/mailman/listinfo/opm
