Re: relaxed 1-sided model

Rolf Rabenseifner (Rabenseifner@RUS.Uni-Stuttgart.DE)
Thu, 20 Jun 1996 14:20:22 +0200 (DST)

Marc Snir wrote:
> An alternative is to specify that a put becomes visible to any other access,
> either local load, or get, only after the call to WINDOW_IN. We get a "delayed
> consistency" model for all window updates. [...]
>
> cons:
> We cannot use anymore windows for "shared memory" style third-party
> communication, where process A and B communicate (using put and get) thru
> memory at process C. Process C must be involved. [...]
>
> pros:
> We can have now a (low quality -:) MPI implementation with no agents: the
> transfer is always affected when the call to WINDOW_IN occurs. [...]

I believe, that chances for low quality MPI implementation
are minor arguments.
And I believe, that most applications fulfill the "delayed
consistency" because they do not make a "shared memory" style
third-party communication (but this is only a weak opinion).

Is there a chance for higher throughput or lower latency
on some hardware platforms if the MPI implementation knows that
the application fulfills the "delayed consistency" model on
a specific window?

Probably on workstation clusters it can be faster to implement
the target side inside of WINDOW_IN/_OUT if the application uses
only small pieces of data.

In those cases I would prefere an optimization hint saying
that it is enough that a put becomes visible to get operations
only after a call to WINDOW_IN rather than restricting MPI
1-sided to the "delayed consistency" model.

Rolf


Rolf Rabenseifner (Computer Center )
Rechenzentrum Universitaet Stuttgart (University of Stuttgart)
Allmandring 30 Phone: ++49 711 6855530
D-70550 Stuttgart 80 FAX: ++49 711 6787626
Germany rabenseifner@rus.uni-stuttgart.de