Re: [Pypar-developers] pypar bug+patch

[Ole N. <ole...@gm...>]

Hi Jim

Thanks again for taking interest in pypar. I looked closer at you
suggestion, incorporated it and ran all the tests and examples.
It is now part of the latest release  (pypar_2.1.0_49). It incorporates


   - Race-condition patch from Jim Bosch
   - Functionality and demo for Bsend by C Makassikis
   - Deprecated obsolete functions
   - Migrated Mandelbrot demos to numpy
   - Fixed demo3.py for version 2 thanks to Felix Richter

Let me know how it goes.
Cheers and thanks
Ole Nielsen
Canberra, Australia



On Fri, Apr 18, 2008 at 6:53 AM, Jim Bosch <jf...@uc...> wrote:

> I've found (and fixed, I think, with the attached patch) a bug in pypar:
> in the current version (SVN), when sending a generic python object, pypar
> actually passes two MPI messages (the "control info" message, and the actual
> object).  However, it does not require that both of these come from the same
> node, so if a third node sends a message that is received between the
> "control info" message and the data message, pypar gets confused.
>
> The attached patch adds an option to "receive_control_info" that causes it
> to return the source node, and this is used to require that the data message
> come from the same node.  It also contains some pickle error trapping and
> re-raising, which I found useful to add a little more information to aid in
> debugging those sorts of errors.
>
> Thanks!
>
> Jim Bosch
>
>
>
>