Solution: Load imbalance: SMxV

Prepare code:

$ source ~/Tools/env.sh
$ cd ~/Tools/SMXV
$ unset ${!SCOREP_*}

Build code:

$ make smxv-omp

Exercise 1: Run benchmark with number of threads between 1 and 16:

$ for t in {1..16}; do OMP_NUM_THREADS=$t ./smxv-omp yax_large.bin; done | nl
     1  Runtime: 8.506669 sec
     2  Runtime: 5.182805 sec
     3  Runtime: 4.042448 sec
     4  Runtime: 3.167305 sec
     5  Runtime: 2.837955 sec
     6  Runtime: 2.741896 sec
     7  Runtime: 2.482917 sec
     8  Runtime: 2.316289 sec
     9  Runtime: 2.110901 sec
    10  Runtime: 1.974172 sec
    11  Runtime: 1.849776 sec
    12  Runtime: 1.802529 sec
    13  Runtime: 1.732268 sec
    14  Runtime: 1.617819 sec
    15  Runtime: 1.604507 sec
    16  Runtime: 1.564077 sec

Ideal scaling would be 0.5 sec for 16 threads. Real scaling only ~34 % parallel efficiency with 16 threads. Most number of threads with >80 % parallel efficiency: 2 (82 %)

Exercise 3: Validate effect of optimization:

$ for t in {1..16}; do OMP_NUM_THREADS=$t OMP_SCHEDULE=dynamic,512 ./smxv-omp yax_large.bin; done | nl
     1  Runtime: 8.512925 sec
     2  Runtime: 4.504325 sec
     3  Runtime: 3.240145 sec
     4  Runtime: 2.416859 sec
     5  Runtime: 2.073413 sec
     6  Runtime: 2.006542 sec
     7  Runtime: 1.659394 sec
     8  Runtime: 1.563580 sec
     9  Runtime: 1.470130 sec
    10  Runtime: 1.407530 sec
    11  Runtime: 1.332805 sec
    12  Runtime: 1.316528 sec
    13  Runtime: 1.267057 sec
    14  Runtime: 1.259499 sec
    15  Runtime: 1.250060 sec
    16  Runtime: 1.228585 sec

~43 % with 16 threads. Most number of threads with >80 % parallel efficiency: 5 (82 %)