Re:Many Core processors: Everything You Know (about Parallel...
by tomp on Monday January 16, 2012 @05:02PM
(#3472619)
Which is why every good numeric library should have a "correct" reduce functions implementations that guarantee minimal errors. in case of a sum of an floating point array, that would be a tree-based sum, possibly coupled by sorting the array beforehand.
Also, if you do some kind of map-reduce style work distribution where different nodes perform same operations on different data, it pays off to use a deterministic (pseudo-)random algorithm for work distribution, so that your results can be reproduced.
Also, if you do some kind of map-reduce style work distribution where different nodes perform same operations on different data, it pays off to use a deterministic (pseudo-)random algorithm for work distribution, so that your results can be reproduced.