MPI_Reduce Called outside Main

[Kyle Klein]

Hi all,

As I was comparing our code I came across something extremely strange. Both teams call MPI_Reduce exactly the same way the same number of times (verified with TAU), however we called MPI_Reduce within DDOT, and the other team called DDOT and then with the return values called MPI_REDUCE. The strange thing is that our code is about 25% slower until we did the same thing and moved our MPI_Reduce out of the ddot, at which point it becomes slightly faster. So to give a more concrete example:

Faster:

ddot()

MPI_Reduce()

Slower:

ddot() //MPI_Reduce call inside ddot

Anyone have a clue what could cause this?

best,

Kyle

--
Kyle Klein 

Ph.D. Student 

Department of Computer Science UC Santa Barbara

http://www.cs.ucsb.edu/~kyleklein