3.3.3 Conclusion: four different ways to accomplish the same task, but use with caution!¶
The bottom line with these examples is that when we have added the complexity of 2 possible ways for trng to decide which thread gets which number from the stream with two possible ways of decomposing the repetitions of a loop, we end up with four ways to achieve the same ultimate task.
Which method to use depends on the situation you are faced with. As a general rule, we often use equal chunks for decomposition because it matches the sequential version and the work being done inside the loop is roughly the same for each iteration. However, if the amount of work varies in each loop iteration, then “chunks of one” can be more effective. We have to combine it with a mechanism for letting more threads do more iterations if they are ready for work. This is called dynamic thread assignment.
When it comes to doling out the random numbers to the threads, the choice is somewhat important. Block splitting makes sense when the iterations can be equally divided by the treads (what happened when you chose 3 threads?). Leapfrog works even if each thread does not do exactly the same number of iterations of the loop. This is often the case, so most of the time the leapfrog method becomes the proper choice.
Command line code for reference¶
This code block below has code for handling the command line arguments for the parallel versions above. You can study it if you want to see how it was done.