So is there any reason OMPI should not auto-detach buffers at Finalize?
I understand technically we don't have to but there are false
performance degradations incurred by us not detaching thus making OMPI
look significantly slower compared to other MPIs for no real reason. So
unless there is some really good reason we shouldn't detach I would
think detaching would make sense.
George Bosilca wrote:
> There is no such statement in the MPI Standard. In fact one of the
> example use exactly this: automatic detach at MPI_Finalize (example on
> page 310 on MPI 2.2). However, as the standard doesn't enforce a
> specific behavior, each MPI implementation can interpret/implement it
> differently. Therefore, by expecting the buffer detach at Finalize the
> user open itself to "inconsistent" behaviors depending on the MPI
> library used. On the opposite, i.e. by explicitly calling detach, the
> behavior is well defined in all cases.
> On Nov 18, 2009, at 15:41 , Eugene Loh wrote:
>> George Bosilca wrote:
>>> The proper practice based on the MPI Standard will be to call the
>>> detach function before finalize.
>> I don't find this described anywhere in the standard. To what
>> chapter/verse should I point a user to convince them that detach
>> before finalize is the proper thing to do?
>> devel mailing list
> devel mailing list