>> At the last meeting, I was the sole dissenter in a number of the
>> formal votes on the one-sided communication chapter. A colleague has
>> since pointed out the essence of my unease, which is that we are
>> voting to standardize an interface that has not been implemented (even
>> partially) and which is not closely related to common practice.
>> I know of no implmentation of the proposed MPI interface. My
>> understanding is that at this stage in MPI-1, there was an
>> implementation that tracked the current spec and provided substantial
> I have documentation that is dated December '95 for an MPI-2 one sided
> communications implementation that EPCC has done for the Cray T3D also.
Indeed, we have implemented a subset of the MPI2 One-sided proposal
that was current prior to SuperComputing '95. This is available as an
extension of the CRI/EPCC MPI for Cray T3D software.
I admit to not having followed all of the subsequent changes to the
proposal (RMC becomes RMA right?). At the time, the subset omissions
were broadly as follows:
* Basic MPI datatypes only for window declaration
* Window, origin and target datatypes must match in Put/Get
* (I)Accumulate, (I)Rmw, and Type_global omitted
* Handler-related sections omitted
Counter requests have been implemented.
Unsurprisingly this fits snugly on CRI's SHMEM interface.
The software has been installed and used at various T3D sites. We have
had limited feedback, but not much to report. I hope to discuss this
with attendees of a European Cray MPP Workshop to be held here in
Edinburgh in July.
-=-=- A. Gordon Smith -=- Edinburgh Parallel Computing Centre -=-=-
=- http://www.epcc.ed.ac.uk/~smith -=- Phone +44 (0)131 650 6712 -=