As we told in the last teleconference, a few lines about our work in the University of Extremadura related to Open MPI:
We are developing new (proof-of-concept) collective and mpool components for improving the collective operations performance. New collective algorithms require user buffers on a shared memory space to all the processes, that it is allocated at initialization time and bound to communicators. User needs to use MPI_Alloc_mem for allocating memory in this shared pool (with an MPI_Info object or MPI_INFO_NULL) instead of malloc. Mpool new component manages the allocation using Open MPI allocator services.
Broadcast and Reduce algorithms are implemented. As expected, performance is improved because of zero intermediate copy, and all processes collaborating in calculation directly on user buffers in Reduce.
Components are called CAS (Common Address Space). Perhaps it is not a good name.
Another work is about adding hierarchical capabilities to Tuned and SM (and CAS) collective components in Open MPI, which not care about hardware topology.
Please, we appreciate comments and your opinions. If you need more information we are glad to provide you (also, a poster in the EuroMPI 2012 will be presented in September).
I would like to thank Shiquing Fan for their help at HLRS in contact you and the procedure for contributing to Open MPI project.
Juan A. Rico
University of Extremadura.
Phone: +34 645 269 389