So far, we have been composing multiple threads of computation into safe concurrent programs. In doing so, we focused on ensuring their correctness. We saw how to avoid blocking in concurrent programs, react to the completion of asynchronous computations, and how to use concurrent data structures to communicate information between threads. All these tools made organizing the structure of concurrent programs easier. In this chapter, we will focus mainly on achieving good performance. We require minimal or no changes in the organization of existing programs, but we will study how to reduce their running time using multiple processors. Futures from the previous chapter allowed doing this to a certain extent, but they are relatively heavyweight and inefficient when the asynchronous computation in each future is short.

Data parallelism is a form of computation where the same computation proceeds...