Work Isolation

In oneAPI Threading Building Blocks (oneTBB), a thread waiting for a group of tasks to complete might execute other available tasks. In particular, when a parallel construct calls another parallel construct, a thread can obtain a task from the outer-level construct while waiting for completion of the inner-level one.

In the following example with two parallel_for calls, the call to the second (nested) parallel loop blocks execution of the first (outer) loop iteration:

// The first parallel loop.
oneapi::tbb::parallel_for( 0, N1, []( int i ) {
    // The second parallel loop.
    oneapi::tbb::parallel_for( 0, N2, []( int j ) { /* Some work */ } );
} );

The blocked thread is allowed to take tasks belonging to the first parallel loop. As a result, two or more iterations of the outer loop might be simultaneously assigned to the same thread. In other words, in oneTBB execution of functions constituting a parallel construct is unsequenced even within a single thread. In most cases, this behavior is harmless or even beneficial because it does not restrict parallelism available for the thread.

However, in some cases such unsequenced execution may result in errors. For example, a thread-local variable might unexpectedly change its value after a nested parallel construct:

oneapi::tbb::enumerable_thread_specific<int> ets;
oneapi::tbb::parallel_for( 0, N1, [&ets]( int i ) {
    // Set a thread specific value
    ets.local() = i;
    oneapi::tbb::parallel_for( 0, N2, []( int j ) { /* Some work */ } );
    // While executing the above parallel_for, the thread might have run iterations
    // of the outer parallel_for, and so might have changed the thread specific value.
    assert( ets.local()==i ); // The assertion may fail!
} );

In other scenarios, the described behavior might lead to deadlocks and other issues. In these cases, a stronger guarantee of execution being sequenced within a thread is desired. For that, oneTBB provides ways to isolate execution of a parallel construct, for its tasks to not interfere with other simultaneously running tasks.

One of these ways is to execute the inner level loop in a separate task_arena:

oneapi::tbb::enumerable_thread_specific<int> ets;
oneapi::tbb::task_arena nested;
oneapi::tbb::parallel_for( 0, N1, [&]( int i ) {
    // Set a thread specific value
    ets.local() = i;
    nested.execute( []{
        // Run the inner parallel_for in a separate arena to prevent the thread
        // from taking tasks of the outer parallel_for.
        oneapi::tbb::parallel_for( 0, N2, []( int j ) { /* Some work */ } );
    } );
    assert( ets.local()==i ); // Valid assertion
} );

However, using a separate arena for work isolation is not always convenient, and might have noticeable overheads. To address these shortcomings, oneTBB provides this_task_arena::isolate function which runs a user-provided functor in isolation by restricting the calling thread to process only tasks scheduled in the scope of the functor (also called the isolation region).

When entered a task waiting call or a blocking parallel construct inside an isolated region, a thread can only execute tasks spawned within the region and their child tasks spawned by other threads. The thread is prohibited from executing any outer level tasks or tasks belonging to other isolated regions.

The isolation region imposes restrictions only upon the thread that called it. Other threads running in the same task arena have no restrictions on task selection unless isolated by a distinct call to this_task_arena::isolate.

The following example demonstrates the use of this_task_arena::isolate to ensure that a thread-local variable is not changed unexpectedly during the call to a nested parallel construct.

#include "oneapi/tbb/task_arena.h"
#include "oneapi/tbb/parallel_for.h"
#include "oneapi/tbb/enumerable_thread_specific.h"
#include <cassert>


int main() {
    const int N1 = 1000, N2 = 1000;
    oneapi::tbb::enumerable_thread_specific<int> ets;
    oneapi::tbb::parallel_for( 0, N1, [&ets]( int i ) {
        // Set a thread specific value
        ets.local() = i;
        // Run the second parallel loop in an isolated region to prevent the current thread
        // from taking tasks related to the outer parallel loop.
        oneapi::tbb::this_task_arena::isolate( []{
            oneapi::tbb::parallel_for( 0, N2, []( int j ) { /* Some work */ } );
        } );
        assert( ets.local()==i ); // Valid assertion
    } );
    return 0;
}