task_arena

[scheduler.task_arena]

A class that represents an explicit, user-managed task scheduler arena.

// Defined in header <oneapi/tbb/task_arena.h>

namespace oneapi {
    namespace tbb {

        class task_arena {
        public:
            static const int automatic = /* unspecified */;
            static const int not_initialized = /* unspecified */;
            enum class priority : /* unspecified type */ {
                low = /* unspecified */,
                normal = /* unspecified */,
                high = /* unspecified */
            };

            struct constraints {
                constraints(numa_node_id numa_node_       = task_arena::automatic,
                            int          max_concurrency_ = task_arena::automatic);

                constraints& set_numa_id(numa_node_id id);
                constraints& set_max_concurrency(int maximal_concurrency);
                constraints& set_core_type(core_type_id id);
                constraints& set_max_threads_per_core(int threads_number);

                numa_node_id numa_id = task_arena::automatic;
                int max_concurrency = task_arena::automatic;
                core_type_id core_type = task_arena::automatic;
                int max_threads_per_core = task_arena::automatic;
            };

            task_arena(int max_concurrency = automatic, unsigned reserved_for_masters = 1,
                    priority a_priority = priority::normal);
            task_arena(constraints a_constraints, unsigned reserved_for_masters = 1,
                    priority a_priority = priority::normal);
            task_arena(const task_arena &s);
            explicit task_arena(oneapi::tbb::attach);
            ~task_arena();

            void initialize();
            void initialize(int max_concurrency, unsigned reserved_for_masters = 1,
                            priority a_priority = priority::normal);
            void initialize(constraints a_constraints, unsigned reserved_for_masters = 1,
                            priority a_priority = priority::normal);
            void initialize(oneapi::tbb::attach);

            void terminate();

            bool is_active() const;
            int max_concurrency() const;

            template<typename F> auto execute(F&& f) -> decltype(f());
            template<typename F> void enqueue(F&& f);

            void enqueue(task_handle&& h);
        };

    } // namespace tbb
} // namespace oneapi

A task_arena class represents a place where threads may share and execute tasks.

The number of threads that may simultaneously execute tasks in a task_arena is limited by its concurrency level.

Each user thread that invokes any parallel construction outside an explicit task_arena uses an implicit task arena representation object associated with the calling thread.

The tasks spawned or enqueued into one arena cannot be executed in another arena.

Each task_arena has a priority. The tasks from task_arena with higher priority are given a precedence in execution over the tasks from task_arena with lower priority.

Note

The task_arena constructors do not create an internal task arena representation object. It may already exist in case of the “attaching” constructor; otherwise, it is created by an explicit call to task_arena::initialize or lazily on first use.

Member types and constants

static const int automatic

When passed as max_concurrency to the specific constructor, arena concurrency is automatically set based on the hardware configuration.

static const int not_initialized

When returned by a method or function, indicates that there is no active task_arena or that the task_arena object has not yet been initialized.

enum priority::low

When passed to a constructor or the initialize method, the initialized task_arena has a lowered priority.

enum priority::normal

When passed to a constructor or the initialize method, the initialized task_arena has regular priority.

enum priority::high

When passed to a constructor or the initialize method, the initialized task_arena has a raised priority.

struct constraints

Represents limitations applied to threads within task_arena.

Starting from C++20 this class should be an aggregate type to support the designated initialization.

numa_node_id constraints::numa_id

An integral logical index uniquely identifying a NUMA node. If set to non-automatic value, then this NUMA node will be considered as preferred for all the threads within the arena.

Note

NUMA node ID is considered valid if it was obtained through tbb::info::numa_nodes().

int constraints::max_concurrency

The maximum number of threads that can participate in work processing within the task_arena at the same time.

core_type_id constraints::core_type

An integral logical index uniquely identifying a core type. If set to non-automatic value, then this core type will be considered as preferred for all the threads within the arena.

Note

core type ID is considered valid if it was obtained through tbb::info::core_types().

int constraints::max_threads_per_core

The maximum number of threads that can be scheduled to one core simultaneously.

constraints::constraints(numa_node_id numa_node_ = task_arena::automatic, int max_concurrency_ = task_arena::automatic)

Constructs the constraints object with the provided numa_id and max_concurrency settings.

Note

To support designated initialization this constructor is omitted starting from C++20. Aggregate initialization is supposed to be used instead.

constraints &constraints::set_numa_id(numa_node_id id)

Sets the numa_id to the provided id. Returns the reference to the updated constraints object.

constraints &constraints::set_max_concurrency(int maximal_concurrency)

Sets the max_concurrency to the provided maximal_concurrency. Returns the reference to the updated constraints object.

constraints &constraints::set_core_type(core_type_id id)

Sets the core_type to the provided id. Returns the reference to the updated constraints object.

constraints &constraints::set_max_threads_per_core(int threads_number)

Sets the max_threads_per_core to the provided threads_number. Returns the reference to the updated constraints object.

Member functions

task_arena(int max_concurrency = automatic, unsigned reserved_for_masters = 1, priority a_priority = priority::normal)

Creates a task_arena with a certain concurrency limit (max_concurrency) and priority (a_priority). Some portion of the limit can be reserved for application threads with reserved_for_masters. The amount for reservation cannot exceed the limit.

Caution

If max_concurrency and reserved_for_masters are explicitly set to be equal and greater than 1, oneTBB worker threads will never join the arena. As a result, the execution guarantee for enqueued tasks is not valid in such arena. Do not use task_arena::enqueue() with an arena set to have no worker threads.

task_arena(constraints a_constraints, unsigned reserved_for_masters = 1, priority a_priority = priority::normal)

Creates a task_arena with a certain constraints(a_constraints) and priority (a_priority). Some portion of the limit can be reserved for application threads with reserved_for_masters. The amount for reservation cannot exceed the concurrency limit specified in constraints.

Caution

If constraints::max_concurrency and reserved_for_masters are explicitly set to be equal and greater than 1, oneTBB worker threads will never join the arena. As a result, the execution guarantee for enqueued tasks is not valid in such arena. Do not use task_arena::enqueue() with an arena set to have no worker threads.

If constraints::numa_node is specified, then all threads that enter the arena are automatically pinned to corresponding NUMA node.

task_arena(const task_arena&)

Copies settings from another task_arena instance.

explicit task_arena(oneapi::tbb::attach)

Creates an instance of task_arena that is connected to the internal task arena representation currently used by the calling thread. If no such arena exists yet, creates a task_arena with default parameters.

Note

Unlike other constructors, this one automatically initializes the new task_arena when connecting to an already existing arena.

~task_arena()

Destroys the task_arena instance, but the destruction may not be synchronized with any task execution inside this task_arena. It means that an internal task arena representation associated with this task_arena instance can be destroyed later. Not thread-safe for concurrent invocations of other methods.

void initialize()

Performs actual initialization of internal task arena representation.

Note

After the call to initialize, the arena parameters are fixed and cannot be changed.

void initialize(int max_concurrency, unsigned reserved_for_masters = 1, priority a_priority = priority::normal)

Same as above, but overrides previous arena parameters.

void initialize(constraints a_constraints, unsigned reserved_for_masters = 1, priority a_priority = priority::normal)

Same as above.

void initialize(oneapi::tbb::attach)

If an internal task arena representation currently used by the calling thread, the method ignores arena parameters and connects task_arena to that internal task arena representation. The method has no effect when called for an already initialized task_arena.

void terminate()

Removes the reference to the internal task arena representation without destroying the task_arena object, which can then be re-used. Not thread safe for concurrent invocations of other methods.

bool is_active() const

Returns true if the task_arena has been initialized; false, otherwise.

int max_concurrency() const

Returns the concurrency level of the task_arena. Does not require the task_arena to be initialized and does not perform initialization.

template<F>
void enqueue(F &&f)

Enqueues a task into the task_arena to process the specified functor and immediately returns. The F type must meet the Function Objects requirements described in the [function.objects] section of the ISO C++ standard. The task is scheduled for eventual execution by a worker thread even if no thread ever explicitly waits for the task to complete. If the total number of worker threads is zero, a special additional worker thread is created to execute enqueued tasks.

Note

The method does not require the calling thread to join the arena; that is, any number of threads outside of the arena can submit work to it without blocking.

Caution

There is no guarantee that tasks enqueued into an arena execute concurrently with respect to any other tasks there.

Caution

An exception thrown and not caught in the functor results in undefined behavior.

template<F>
auto execute(F &&f) -> decltype(f())

Executes the specified functor in the task_arena and returns the value returned by the functor. The F type must meet the Function Objects requirements described in the [function.objects] section of the ISO C++ standard.

The calling thread joins the task_arena if possible, and executes the functor. Upon return it restores the previous task scheduler state and floating-point settings.

If joining the task_arena is not possible, the call wraps the functor into a task, enqueues it into the arena, waits using an OS kernel synchronization object for another opportunity to join, and finishes after the task completion.

An exception thrown in the functor will be captured and re-thrown from execute.

Note

Any number of threads outside of the arena can submit work to the arena and be blocked. However, only the maximal number of threads specified for the arena can participate in executing the work.

void enqueue(task_handle &&h)

Enqueues a task owned by h into the task_arena for processing.

The behavior of this function is identical to the generic version (template<typename F> void task_arena::enqueue(F&& f)), except parameter type.

Note

h should not be empty to avoid an undefined behavior.

Example

The example demonstrates task_arena NUMA support API. Each constructed task_arena is pinned to the corresponding NUMA node.

#include "oneapi/tbb/task_group.h"
#include "oneapi/tbb/task_arena.h"

#include <vector>

int main() {
    std::vector<oneapi::tbb::numa_node_id> numa_nodes = oneapi::tbb::info::numa_nodes();
    std::vector<oneapi::tbb::task_arena> arenas(numa_nodes.size());
    std::vector<oneapi::tbb::task_group> task_groups(numa_nodes.size());

    for (int i = 0; i < numa_nodes.size(); i++) {
        arenas[i].initialize(oneapi::tbb::task_arena::constraints(numa_nodes[i]));
    }

    for (int i = 0; i < numa_nodes.size(); i++) {
        arenas[i].execute([&task_groups, i] {
            task_groups[i].run([] {
                /* executed by the thread pinned to specified NUMA node */
            });
        });
    }

    for (int i = 0; i < numa_nodes.size(); i++) {
        arenas[i].execute([&task_groups, i] {
            task_groups[i].wait();
        });
    }

    return 0;
}

See also: