window lock functions

Nobutoshi Sagawa (sagawa@crl.hitachi.co.jp)
Thu, 30 Nov 95 10:26:16 JST

>Ah, now I get it; my mistake. So, rather than force applications to use their
>own (possibly non-portable) locks, you would like to provide a MPI function to
>enable/disable remote operations on a local window?

Exactly.

>I'm wondering how such a function would be implemented. If you use some sort of
>agent to manage remote requests, it seems pretty straightforward. But if you
>have a shared memory machine, things might get a bit messy.

I admit I haven't given enough thought about shared memory machines. I
appreciate some help on this front since I have little experience with
those machines.

>I also wonder if this provides enough functionality to be generally useful. As
>you point out, it cannot be used to implement general RMW operations.

'get'), we found that this kind of operation was bug-prone and
sometimes programmer hostile. I think it is better, at least for
debugging and tuning purposes, for a window to be equipped with a
memory protection feature. There may be some point we can make
compromise between generality, usefulness and ease of implementation.
My own assessment is that the local lock/unlock functions are "not too
general, useful for some, and cheap to implement (except for shared
memory machines)".

>Maybe we should just bite the bullet and try to come up with some sort of
>general semaphore interface instead? (Ouch.)

If it can be done, it's most welcome.

-N.Sagawa