I have a vector<int> with 10,000,000 (10 million) elements, and that my workstation has four cores. There is a function, called ThrFunc, that operates on an integer. Assume that the runtime for ThrFunc for each integer in the vector<int> is roughly the same.
How should I determine the optimal number of threads to fire off? Is the answer as simple as the number of elements divided by the number of cores? Or is there a more subtle computation?
Editing to provide extra information
- No need for blocking; each function invocation needs only read-only
access
The optimal number of threads is likely to be either the number of cores in your machine or the number of cores times two.
In more abstract terms, you want the highest possible throughput. Getting the highest throughput requires the fewest contention points between the threads (since the original problem is trivially parallelizable). The number of contention points is likely to be the number of threads sharing a core or twice that, since a core can either run one or two logical threads (two with hyperthreading).
If your workload makes use of a resource of which you have fewer than four available (ALUs on Bulldozer? Hard disk access?) then the number of threads you should create will be limited by that.
The best way to find out the correct answer is, with all hardware questions, to test and find out.