| Thread Previous • Date Previous • Date Next • Thread Next |
On 02/03/2012 10:07 AM, Johan Hake wrote:
On Thursday February 2 2012 18:16:20 Garth N. Wells wrote:On 2 February 2012 13:09, Marie E. Rognes<meg@xxxxxxxxx> wrote:Thanks for starting to take care of this Johan! However, it does not completely resolve the issue for me: Building the most recent 1.0.x and trunk on my laptop, by clearing the build directory, and running cmake.local gave the following (rough) timings for the Python stage: 1.0.x (aka 'old'): ~3 min 21 s trunk (aka 'new'): ~ 3 min 23 s These timings are for not doing anything else at the same time except staring at the screen.Did you use 'make -jX'? I would expect building with one process to take about the same time, but I see is that I can now build the SWIG interface files in parallel. I haven't timed the difference.
Yes: cmake.local uses make -j3
I use same configurations as in cmake.local (make -j3). My timings are: 1.0.x (aka 'old'): ~1 min 40 s trunk (aka 'new'): ~ 1 min 55 s And max resources are as Garth reports.
Ok, but it still stalls everything else I'm trying to do at the same time.
So no, it is not that of an improvement, but I think we have a framework to continue working on. Next step would be to split the modules even further. But we probably have to end up with splitting the directories of la and mesh to accomplish that.
One big improvement is that only the modules that need to be rebuilt will be each time, right? This is a more frequent scenario than building dolfin from scratch, so that should save some. -- Marie
JohanThe 'new' still uses so much memory that doing anything else at the same time is very painful.I see less< 1GB for each module, whereas before I think that it was more like> 2GB. Garth-- Marie On 02/01/2012 05:23 PM, Johan Hake wrote:Hello! I have now splitted the SWIG interface into 6 combined modules. common, io, mesh, function, fem, la The largest .cxx file generated now is 2MB (for la with all linear algebra backends installed), instead of 6MB. That is still pretty big and take some 1GB of memory during compilation. But the interface can now be generated in several threads, reducing compile time considerable. The runtime memory print, read from a simple top, is not much more than before (some 100MB), which is good. They new compiled modules all resides under: dolfin.cpp.foo All types and functions are now automatically included into: dolfin.cpp which means that most modules under site-packages/dolfin still works fine. There are now more types included from the compiled interface by default. If this is a problem we can change that. But now everything is automated. No handpicking done. The logic in the new SWIG interface is described in: dolfin/swig/README One thing to be aware of is that any reference to the compiled interface from another combined module have to be preluded with the name of the module. Example, in la/post.i we call dolfin_error, which is defined in common. Then we do: common.dolfin_error(...) instead of previous: dolfin_error(...) Have fun and let the bug reports come raining! Johan _______________________________________________ Mailing list: https://launchpad.net/~dolfin Post to : dolfin@xxxxxxxxxxxxxxxxxxx Unsubscribe : https://launchpad.net/~dolfin More help : https://help.launchpad.net/ListHelp_______________________________________________ Mailing list: https://launchpad.net/~dolfin Post to : dolfin@xxxxxxxxxxxxxxxxxxx Unsubscribe : https://launchpad.net/~dolfin More help : https://help.launchpad.net/ListHelp_______________________________________________ Mailing list: https://launchpad.net/~dolfin Post to : dolfin@xxxxxxxxxxxxxxxxxxx Unsubscribe : https://launchpad.net/~dolfin More help : https://help.launchpad.net/ListHelp
| Thread Previous • Date Previous • Date Next • Thread Next |
