Task Parallelism Overview¶
Characterizing Chapel Tasks¶
In Chapel, a task is a unit of computation that can and should be executed in parallel with other tasks. Tasks are the basis of all parallel execution within Chapel.
Chapel programs begin execution as a single task. As the program runs, additional tasks can be created to introduce parallel computation. Chapel does not automatically parallelize programs, which is to say, tasks are not automatically created on the programmer’s behalf. Rather, the programmer must create all tasks, either directly (using the constructs described in this Task Parallelism section), or indirectly (through the use of Data Parallelism abstractions that create the tasks).
A Chapel task may be short-lived, or it may run for the program’s duration. It may be defined via a single statement or using as much code as is required to specify its behavior. Tasks may also be nested, meaning that tasks may themselves create additional tasks.
Chapel tasks may share data with other tasks through traditional lexical scoping. Unlike many task-parallel programming models, Chapel programs are not guaranteed to be correct if tasks are executed serially and to completion. Specifically, Chapel tasks may synchronize with one another. They may also block or busy-wait while waiting for actions to be taken by other tasks. When tasks do not synchronize with one another correctly, Chapel programs may deadlock or livelock. Tasks may also race with one another, either in benign or problematic ways.
Task Parallelism vs. Data Parallelism¶
Chapel uses the term task parallel to describe a coding style in which tasks are explicitly specified to perform a parallel computation (e.g., “This task should do this thing while a second task does this other thing.” Or “These k tasks should all do this thing.”). This is in contrast to data parallel programming, in which parallelism is specified by computations on collections of indices or data.
Tasks vs. Threads¶
We tend to avoid the term thread in defining the Chapel language. In Chapel, threads are system-level resources (e.g., POSIX threads) that execute a program’s tasks. How tasks are mapped to threads is not defined by the Chapel language. A Chapel implementation may support distinct tasking layers, each with its own policies for executing tasks on the target system’s resources. For detailed information about the tasking layers available in our implementation of Chapel, refer to Chapel Tasks.
“Will My Tasks Execute in Parallel?”¶
Practically speaking, whether or not a task actually does run in parallel with other tasks depends on:
the number of simultaneous tasks created by the Chapel program,
the available system resources, and
details of the runtime library’s tasking implementation which are beyond the scope of this introductory description.
Generally speaking, if the number of concurrently executing tasks does not exceed the number of available processors in the target hardware, the tasks will be executed in parallel. So, if you are running on a 4-core laptop and never have more than four tasks running at any given time, all your tasks should run in parallel.
When the number of tasks created by a Chapel program exceeds the amount of hardware parallelism available, tasks will typically be multiplexed across the available hardware over time. The specifics depend on how tasks are implemented by the runtime library; again, see Chapel Tasks for further details if you’re curious to understand task scheduling in more detail.