Changes between Initial Version and Version 1 of Ticket #18529, comment 85


Ignore:
Timestamp:
12/02/15 10:59:38 (4 years ago)
Author:
egourgoulhon
Comment:

Legend:

Unmodified
Added
Removed
Modified
  • Ticket #18529, comment 85

    initial v1  
    22> Okay, let's revert back to having `UniqueRepresentation` because it will work as a session/variable manager. I don't think we should spend a lot of time doing that as we can hack around it (even though it is in effect what we are doing/should do). I don't like it because it is a hack, but it seems like the best option for now. Unless you want to directly use Python's multiprocessing with shared memory for the corresponding parents to avoid pickling, which I would still somewhat prefer in a way so we don't have to do this hack. (Multithreading in Python has shared memory, but unfortunately it can't run on multiple processors because of the GIL.) Anyways, I leave the decision up to you.
    33
    4 Yes, I think it is reasonable to revert to `UniqueRepresentation` at this stage, leaving Python's multiprocessing without pickling for a future development, exploring meanwhile other ways of parallelization (IPython parallel framework ?). I've rerun this morning some benchmarks on the commit of #19209 based on `UniqueRepresentation` (commit [http://git.sagemath.org/sage.git/commit/?id=82f6f495bb6f729624dd8396208f5a3c67e2fa8d 82f6f495b]) merged into Sage 6.10.beta6: the gain in the computation of the Riemann tensor of a concrete metric (4-dimensional Kerr metric) is really significant: a factor of 4 when using 8 cores instead of 1. 
     4Yes, I think it is reasonable to revert to `UniqueRepresentation` at this stage, leaving Python's multiprocessing without pickling for a future development, exploring meanwhile other ways of parallelization (IPython parallel framework ?). I've rerun this morning some benchmarks on the commit of #19209 based on `UniqueRepresentation` (commit [http://git.sagemath.org/sage.git/commit/?id=82f6f495bb6f729624dd8396208f5a3c67e2fa8d 82f6f495b]) merged into Sage 6.10.beta6: the gain in the computation of the Riemann tensor of a non-trivial metric (4-dimensional [http://sagemanifolds.obspm.fr/examples/html/SM_Kerr.html Kerr metric]) is really significant: a factor of 4 when using 8 cores instead of 1 (30 s instead of 115 s on Xeon E5-2623 CPUs).