Requests are performance-critical objects and thus are allocated and managed internally by MPICH. You may not have access to it. It is possible to have user pre-allocated the request objects, but it will add much complications that we'd very much like to avoid. Anyway, I pretty much support that we not to bother with the cases of allreduce on different input/output layout. I don't think it is working with any implementations, so technically it is not even breaking anything. If we want to support this, the straightforward approach is to offer an extension that allows allreduce with different in/out datatypes and assume per-element operations. -- Hui ________________________________ From: Jeff Hammond <[email protected]> Sent: Friday, January 6, 2023 2:17 AM To: Zhou, Hui <[email protected]> Cc: MPICH <[email protected]> Subject: Re: [mpich-devel] code path for non-blocking operations that are deferred until synchronization Thanks. This is helpful. In order to avoid ABI issues with the status object, I was thinking instead we could define attributes on requests. That is sufficient to solve all sorts of problems, because one can just get the attributes off the request and decide to clean up as appropriate. Do you think adding attributes to requests causes any implementation headaches? My expectation is that it does not slow down any critical path because the only time the attribute fields are touched are when the user accesses them. The internal object behind a request would need an additional field, but that's a small change to MPIR_Request. However, if we end up standardizing a 256-bit status object in MPI-5 (hypothetically), then we have an option to put hidden fields for this in status. It's debatable if MPI_Status_get_buffer_address(MPI_Status * status) is significantly better than a request attribute lookup. In any case, I figured out that I can do the deferred implementation using generalized requests. It requires Pthreads and MPI_THREAD_MULTIPLE but I can live with that. Jeff On Fri, Jan 6, 2023 at 12:39 AM Zhou, Hui <[email protected]<mailto:[email protected]>> wrote: Hi Jeff, The nonblocking cases can also do the same support by copying to and from contiguous buffers if we stick the temporary buffer pointer in the request object and free it upon completion. But if you are looking at this from the Fortran binding layer, which does not have access to MPICH internals, then it is difficult to do. The straight-forward solution would be ask MPICH to add an extension, so the user can stick the buffer pointer somewhere, for example, the status object. I wonder if there is a code path in MPICH associated with nonblocking communication that is deferred until the request is synchronized. For example, is there code that merely copies all the arguments passed to the call into some object and then processes the whole operation when the request is synchronized? I am not exactly sure what do you mean here, but the persistent request most resembles your idea of merely hold all the arguments and only process it at start time. But again, if you are not allowed to access MPICH internals, then that is out-of-reach. -- Hui ________________________________ From: Jeff Hammond via devel <[email protected]<mailto:[email protected]>> Sent: Thursday, January 5, 2023 4:50 AM To: MPICH <[email protected]<mailto:[email protected]>> Cc: Jeff Hammond <[email protected]<mailto:[email protected]>> Subject: [mpich-devel] code path for non-blocking operations that are deferred until synchronization This is related to https://github.com/mpi-forum/mpi-issues/issues/663 but there are other scenarios where nonblocking communication with non-contiguous Fortran subarrays is extremely difficult to implement. The blocking cases are easy to support by copying to and from contiguous buffers. I wonder if there is a code path in MPICH associated with nonblocking communication that is deferred until the request is synchronized. For example, is there code that merely copies all the arguments passed to the call into some object and then processes the whole operation when the request is synchronized? Critically for my case, the Fortran arguments would need to be copied, meaning the CFI_cdesc_t pointer not just the associated base_addr, which I doubt is there based on what I've read so far. Thanks, Jeff -- Jeff Hammond [email protected]<mailto:[email protected]> http://jeffhammond.github.io/ -- Jeff Hammond [email protected]<mailto:[email protected]> http://jeffhammond.github.io/