Re: we should wait for 1sided implementations

Koichi Konishi (konishi@research.nj.nec.com)
Wed, 22 May 1996 14:34:23 EDT

hello,

> From: bright@cs.sandia.gov (Ron Brightwell)
> Date: Wed, 22 May 96 10:56:57 MDT

> Our implementation currently ignores the target counter argument to
> the put functions (which will require a handler or something at the
> target), so I voted against having this functionality at the last
> meeting.

Now I am worried about this, though I have voted for keeping the
counter.

NEC Cenju has a non-MPI library including PUT/GET functions whose
implementation relying on interrupt handling. With that in my mind, I
have taken the role of MPI PUT/GET to provide a rich functionality
with a moderate performance.

If the intention is to provide the best possible performance on any
particular platform, then the counter feature does not look right.

I imagine people trying to introduce Window_expose/deliver must have
had the latter intention, as cache-coherence could be handled by
interrupt handlers, without using exporse/deliver.

Koichi Konishi konishi@research.nj.nec.com
NEC Research Institute Office: 609-951-2628
FAX: 609-951-2488