The current MPI standard defines collective operations in a blocking
way. However, it was shown by multiple groups that a non-blocking
interface to collective operations would be beneficial. Benefits are for
example: possible communication and computation overlap, mitigation of
the effects of pseudo-synchronization on large communicators (with high
OS noise) and of course the semantic advantages of non-blocking
operations.
A proposal is attached to this email and is intended to serve as a
discussion base for the meetings. Comments are always welcome!
Best,
Torsten
--
bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ -----
Indiana University | http://www.indiana.edu
Open Systems Lab | http://osl.iu.edu/
150 S. Woodlawn Ave. | Bloomington, IN, 474045-7104 | USA
Lindley Hall Room 135 | +01 (812) 855-3608
Attachment:
nbcoll.pdf
Description: Adobe PDF document