On Jan 5, 2009, at 10:09 AM, Brian W. Barrett wrote:
I think this sounds reasonable, if (and only if) MPI_Accumulate is properly handled. The interface for calling the op functions was broken in some fairly obvious way for accumulate when I was writing the one-sided code. I think I had to call some supposedly internal bits of the interface to make accumulate work. I can't remember what they are now, but I do remember it being a problem.
Coolio; I'll look into it.
Of course, unless it makes mpi_allreduce on one double-sized floating point number using sum go faster, I'm not entirely sure a change is helpful ;).
From my (admittedly limited) understanding, since there are memory registration and/or copy in/out issues with GPUs, the operation has to be "big enough" and/or already located in GPU memory for the GPU to outperform the CPU. It is my assumption that the component-ized CUDA/ OpenCL/whatever code will need to make a decision whether it should perform the operation at run-time or pass it back to a fallback [probably CPU-based] implementation, analogous to how "tuned" picks the right coll algorithm.
I'm told that there's some researchy middleware working on exactly this kind of problem (determining if a given operation is suitable to run on the GPU or the main CPU). So in a best-case scenario, OMPI can just link against and use that middleware rather than implementing all the logic in the component itself. We'll see how it plays out.
My goal is to give these guys the infrastructure that they need in OMPI to play with these kind of concepts and see what they can accomplish in terms of real performance. FWIW: a few SC08 attendees thought that they could avoid writing much CUDA/CL/whatever code if MPI_REDUCE did the work for them (particularly if paired with the proposed MPI_REDUCE_LOCAL function, https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/24) . [shrug] We'll see!
-- Jeff Squyres Cisco Systems
