Query Optimization and Execution - Pro SQL Server Internals - page 470

Figure 25-14. Parallel Execution: Query Plans
As you can see, the response (elapsed) time of the first query is much slower than that of the second query:
245 milliseconds versus 90 milliseconds. However, the total CPU time of the first query is much lower compared to
second query: 240 milliseconds versus 655 milliseconds. We are using CPU resources for parallelism management.
A parallel execution plan does not necessarily mean that all operators are executing in parallel. An execution
plan can have both parallel and serial execution zones. The parallel plan shown in Figure 25-14 runs a subquery in a
parallel zone and an outer count(*) calculation serially.
The Parallelism operator, sometimes called Exchange , manages parallelism during query execution, and it can
run in three different modes. It accepts the input data from one or more producer threads and distributes it across one
or more consumer threads. Let's take an in-depth look at those modes.
In Distribute Streams mode, the Parallelism operator accepts data from one producer thread and distributes it
across multiple consumer threads. This mode is usually the entry point to the parallel execution zone in the plan.
Figure 25-15 illustrates this concept.
Figure 25-15. Parallelism: Distribute Streams
In Gather Streams mode, the Parallelism operator merges the data from multiple producer threads and passes
it to a single consumer thread. This mode is usually the exit point from the parallel execution zone in the plan.
Figure 25-16 illustrates this idea.