Re: caching on MPI handles
Rajeev Thakur (thakur@mcs.anl.gov)
Fri, 18 Oct 1996 16:15:25 -0500
> From: "Eric Salo" <salo@mrjones.engr.sgi.com>
>
> I'll pass on the question of file handles, since I don't claim to
> understand
> anything about the I/O chapter. But in general, adding dedicated
> routines for
> each object class to support attribute caching feels like a much better
> solution to me, so this is an approach that I would personally favor.
>
> Regarding datatypes, it does seem like this would be a useful bit of
> functionality, but I'm a bit concerned about the effect on derived
> datatypes.
>
> Consider the following sequence of events:
>
> 1) An attribute is cached on Datatype A
> 2) Datatype A is used to create Datatype B
> 3) The value of the attribute on Datatype A is altered.
>
> What is the defined behavior? Perhaps there is a simple solution to this
> problem, but if so it's not yet obvious to me.
>
> Let's continue this thread on mpi-external...
The behavior is what we define it to be. I will happily define it to
be the simplest possible: The first time Datatype B is created, the
behavior is defined by the copy function specified with the
attribute. From then on, any changes to the attribute on Datatype A do
not automatically get reflected on Datatype B.
Rajeev