On Sat, Aug 01, 2009 at 10:54:20AM -0500, Rajeev Thakur wrote:

Torsten, What does libNBC do in this case? hah, I forgot to say that :). LibNBC generates a collective schedule that encodes all steps in the init (MPI_I*) call. Thus, it simply copies the arguments in the schedule.

This technique allows for tree-like (or butterfly) methods to be used in vector variants reducing the needed space per process to O(log(P)). However, they can become arbitrarily slow (log(P) large messages can be serialized at one process) with an adversarial vector distribution. Thus, I implemented the less scalable method that needs O(P) space per process (it'll eliminate all zeros) which can not loose parallelism or degenerate. Best, Torsten -- bash$ :(){ :|:&};: --------------------- http://www.unixer.de/ ----- Torsten Hoefler | Postdoctoral Fellow Open Systems Lab | Indiana University 150 S. Woodlawn Ave. | Bloomington, IN, 474045, USA Lindley Hall Room 135 | +01 (812) 855-3608