comments

Steve Huss-Lederman (lederman@cs.wisc.edu)
Wed, 17 Apr 1996 07:38:21 -0500

First, I would like to say I think the latest put/get semantics is
much better than the old way. Furthermore, the allowing of different
disp_units on each processor will likely take care of many programs
running in the heterogeneous case.

Second, about p. 14. There is a discussion that target_datatype can
be a derived datatype where it is made up of only one basic datatype.
I went back and looked at the 1.1 text. It is not so clear there. It
limits the datatypes to the ones listed for each predefined op.
However, it does not seem to exclude a derived datatype that combines
multiple datatypes listed. Thus, I think it may leave the door open
to a MPI_MAX on a derived datatype which has ints, floats, and longs.
Since you have the datatype at each node, this can work. However, I
was wondering if we really intended to allow this. Also, it is
somewhat different than Marc's one-sided text. This difference could
be argued since that datatype is only given at the origin in this
case. However, it is not clear to me this is what we intended.

Steve