Filter
Top Products
Section: 5.4 Collective operations
Scali MPI Connect Release 4.4 Users Guide 50
4 pair4
5 pipe0
6 pipe1
7 safe
def 8 smp
By looping through these alternatives the performance of IS varies:
algorithm 0: Mop/s total = 95.60
algorithm 1: Mop/s total = 78.37
algorithm 2: Mop/s total = 34.44
algorithm 3: Mop/s total = 61.77
algorithm 4: Mop/s total = 41.00
algorithm 5: Mop/s total = 49.14
algorithm 6: Mop/s total = 85.17
algorithm 7: Mop/s total = 60.22
algorithm 8: Mop/s total = 48.61
For this particular combination of Alltoallv-algorithm and application (IS) the performance
varies significantly, with algorithm 0 close to doubling the performance over the default.
5.4.1 Finding the best algorithm
Consider the image processing example from Chapter 4 which containes four collective
operations. All of these can be tuned with respect to algorithm according to the following
pattern:
user% for a in <range>; do
\>; SCAMPI_<MPI-function>_ALGORITHM=$a \
\>;mpimon <application> -- <nodes> ><application>.out.$a; \
\>; done
For example, trying out the alternative algorithms for MPI_Reduce with two processes can be
done as follows (assuming Bourne Again Shell [bash]:
user% for a in 0 1 2 3 4 5 6 7 8; do
\>; SCAMPI_REDUCE_ALGORITHM=$a
\>; mpimon ./kollektive-8 ./uf256-8.pgm -- r1 r2;
\>; done
Given that the application then reports the timing of the relevant parts of the code a best choice
can be made. Note however that with multiple collective operations working in the same
program there may be interference between the algorithms. Also, the performance of the
implementations is interconnect dependent.