Getting rid of the data types is not an option, in my opinion. I would be ok if we decided on a subset that includes something that includes basic data types and some sort of regular patterns based on these - which I believe represents a very large fraction of the application uses. I am NOT advocating going away from the general support we have for data types in MPI, just providing a way for implementers to know that under some use case scenarios (which I think are by far the common case) simpler and more efficient data type support can be provided. This also allows for implementations, if they choose to take advantage of h/w gather/scatter capabilities. At this stage the notion of subsetting is just that - a notion - and I don't think that as a group we have thought through all the implications. Rich On 2/29/08 10:02 AM, "Torsten Hoefler" <htor_at_[hidden]> wrote:

Bronis, for the record: I do *not* advocate to get rid of datatypes! I think datatypes are a great thing for some parallel applications and they certainly should be used as a high-level abstraction. I've implemented scatter/gather list-based optimizations for modern NICs (IB).

But on the other hand, there are many codes out there that do just not use datatypes. Codes that are only supposed to run in heterogeneous environments. Codes that use sockets instead of MPI. If we want to aim at this market, we need to simplify here. A simplification could be to use MPI_BYTE by default ;) but it would be better to get rid of the code and control-path overhead.

Just to clarify my opinion, Torsten