Intel Threading Building Blocks

Achieving Parallelism

Coordinating people around the job of preparing and mailing the envelopes is easily expressed by the following two conceptual steps:

1. Assign people to tasks (and feel free to move them around to balance the workload).

2. Start with one person on each of the six tasks, but be willing to split up a given task so that two or more people can work on it together.

The six tasks are folding, stuffing, sealing, addressing, stamping, and mailing. We also have six people (resources) to help with the work. That is exactly how Threading Building Blocks works best: you define tasks and data at a level you can explain and then split or combine data to match up with resources available to do the work.

The first step in writing a parallel program is to consider where the parallelism is. Many textbooks wrestle with task and data parallelism as though there were a clear choice. Threading Building Blocks allows any combination of the two that you express.

If you are lucky, your program will be cleanly data-parallel only. To simplify this work, Threading Building Blocks requires you only to specify tasks and how to split them. For a completely data-parallel task, in Threading Building Blocks you will define one task to which you give all the data. That task will then be split up automatically to use the available hardware parallelism. The implicit synchronization will often eliminate the need for using locks to achieve synchronization.

People have been exploring decomposition for decades, ...