Cookbook » Executor

After you create a task dependency graph, you need to submit it to threads for execution. In this chapter, we will show you how to execute a task dependency graph.

Create an Executor

To execute a taskflow, you need to create an executor of type tf::Executor. An executor is a thread-safe object that manages a set of worker threads and executes tasks through an efficient work-stealing algorithm. Issuing a call to run a taskflow creates a topology, a data structure to keep track of the execution status of a running graph. tf::Executor takes an unsigned integer to construct with N worker threads. The default value is std::thread::hardware_concurrency.

tf::Executor executor1;     // create an executor with the number of workers
                            // equal to std::thread::hardware_concurrency
tf::Executor executor2(4);  // create an executor of 4 worker threads

An executor can be reused to execute multiple taskflows. In most workloads, you may need only one executor to run multiple taskflows where each taskflow represents a part of a parallel decomposition.

Understand Work-stealing in Executor

Taskflow designs a highly efficient work-stealing algorithm to schedule and run tasks in an executor. Work-stealing is a dynamic scheduling algorithm widely used in parallel computing to distribute and balance workload among multiple threads or cores. Specifically, within an executor, each worker maintains its own local queue of tasks. When a worker finishes its own tasks, instead of becoming idle or going sleep, it (thief) tries to steal a task from the queue another worker (victim). The figure below illustrates the idea of work-stealing:

Image

The key advantage of work-stealing lies in its decentralized nature and efficiency. Most of the time, worker threads work on their local queues without contention. Stealing only occurs when a worker becomes idle, minimizing overhead associated with synchronization and task distribution. This decentralized strategy effectively balances the workload, ensuring that idle workers are put to work and that the overall computation progresses efficiently.

Execute a Taskflow

tf::Executor provides a set of run_* methods, tf::Executor::run, tf::Executor::run_n, and tf::Executor::run_until to run a taskflow for one time, multiple times, or until a given predicate evaluates to true. All methods accept an optional callback to invoke after the execution completes, and return a tf::Future for users to access the execution status. The code below shows several ways to run a taskflow.

 1: // Declare an executor and a taskflow
 2: tf::Executor executor;
 3: tf::Taskflow taskflow;
 4:
 5: // Add three tasks into the taskflow
 6: tf::Task A = taskflow.emplace([] () { std::cout << "This is TaskA\n"; });
 7: tf::Task B = taskflow.emplace([] () { std::cout << "This is TaskB\n"; });
 8: tf::Task C = taskflow.emplace([] () { std::cout << "This is TaskC\n"; });
 9: 
10: // Build precedence between tasks
11: A.precede(B, C); 
12: 
13: tf::Future<void> fu = executor.run(taskflow);
14: fu.wait();                // block until the execution completes
15:
16: executor.run(taskflow, [](){ std::cout << "end of 1 run"; }).wait();
17: executor.run_n(taskflow, 4);
18: executor.wait_for_all();  // block until all associated executions finish
19: executor.run_n(taskflow, 4, [](){ std::cout << "end of 4 runs"; }).wait();
20: executor.run_until(taskflow, [cnt=0] () mutable { return ++cnt == 10; });

Debrief:

  • Lines 6-8 create a taskflow of three tasks A, B, and C
  • Lines 13-14 run the taskflow once and wait for completion
  • Line 16 runs the taskflow once with a callback to invoke when the execution finishes
  • Lines 17-18 run the taskflow four times and use tf::Executor::wait_for_all to wait for completion
  • Line 19 runs the taskflow four times and invokes a callback at the end of the fourth execution
  • Line 20 keeps running the taskflow until the predicate returns true

Issuing multiple runs on the same taskflow will automatically synchronize to a sequential chain of executions in the order of run calls.

executor.run(taskflow);         // execution 1
executor.run_n(taskflow, 10);   // execution 2
executor.run(taskflow);         // execution 3
executor.wait_for_all();        // execution 1 -> execution 2 -> execution 3
tf::Executor executor;  // create an executor

// create a taskflow whose lifetime is restricted by the scope
{
  tf::Taskflow taskflow;
  
  // add tasks to the taskflow
  // ... 

  // run the taskflow
  executor.run(taskflow);

} // leaving the scope will destroy taskflow while it is running, 
  // resulting in undefined behavior

Similarly, you should avoid touching a taskflow while it is running.

tf::Taskflow taskflow;

// Add tasks into the taskflow
// ...

// Declare an executor
tf::Executor executor;

tf::Future<void> future = executor.run(taskflow);  // non-blocking return

// alter the taskflow while running leads to undefined behavior 
taskflow.emplace([](){ std::cout << "Add a new task\n"; });

You must always keep a taskflow alive and must not modify it while it is running on an executor.

Execute a Taskflow with Transferred Ownership

You can transfer the ownership of a taskflow to an executor and run it without wrangling with the lifetime issue of that taskflow. Each run_* method discussed in the previous section comes with an overload that takes a moved taskflow object.

tf::Taskflow taskflow;
tf::Executor executor;

taskflow.emplace([](){});

// let the executor manage the lifetime of the submitted taskflow
executor.run(std::move(taskflow));

// now taskflow has no tasks
assert(taskflow.num_tasks() == 0);

However, you should avoid moving a running taskflow which can result in undefined behavior.

tf::Taskflow taskflow;
tf::Executor executor;

taskflow.emplace([](){});

// executor does not manage the lifetime of taskflow
executor.run(taskflow);

// error! you cannot move a taskflow while it is running
executor.run(std::move(taskflow));

The correct way to submit a taskflow with moved ownership to an executor is to ensure all previous runs have completed. The executor will automatically release the resources of a moved taskflow right after its execution completes.

// submit the taskflow and wait until it completes
executor.run(taskflow).wait();

// now it's safe to move the taskflow to the executor and run it
executor.run(std::move(taskflow));

Likewise, you cannot move a taskflow that is running on an executor. You must wait until all the previous fires of runs on that taskflow complete before calling move.

// submit the taskflow and wait until it completes
executor.run(taskflow).wait();

// now it's safe to move the taskflow to another
tf::Taskflow moved_taskflow(std::move(taskflow));

Execute a Taskflow from an Internal Worker

Each run variant of tf::Executor returns a tf::Future object which allows you to wait for the result to complete. When calling tf::Future::wait, the caller blocks without doing anything until the associated state is written to be ready. This design, however, can introduce deadlock problem especially when you need to run multiple taskflows from the internal workers of an executor. For example, the code below creates a taskflow of 1000 tasks with each task running a taskflow of 500 tasks in a blocking fashion:

tf::Executor executor(2);
tf::Taskflow taskflow;
std::array<tf::Taskflow, 1000> others;

for(size_t n=0; n<1000; n++) {
  for(size_t i=0; i<500; i++) {
    others[n].emplace([&](){});
  }
  taskflow.emplace([&executor, &tf=others[n]](){
    // blocking the worker can introduce deadlock where
    // all workers are waiting for their taskflows to finish
    executor.run(tf).wait();
  });
}
executor.run(taskflow).wait();

To avoid this problem, the executor has a method, tf::Executor::corun, to execute a taskflow from a worker of that executor. The worker will not block but co-run the taskflow with other tasks in its work-stealing loop.

tf::Executor executor(2);
tf::Taskflow taskflow;
std::array<tf::Taskflow, 1000> others;

std::atomic<size_t> counter{0};

for(size_t n=0; n<1000; n++) {
  for(size_t i=0; i<500; i++) {
    others[n].emplace([&](){ counter++; });
  }
  taskflow.emplace([&executor, &tf=others[n]](){
    // the caller worker will not block but corun these
    // taskflows through its work-stealing loop
    executor.corun(tf);
  });
}
executor.run(taskflow).wait();

Similar to tf::Executor::corun, the method tf::Executor::corun_until is another variant that keeps the calling worker in the work-stealing loop until the given predicate becomes true. You can use this method to prevent blocking a worker from doing useful things, such as being blocked when submitting an outstanding task (e.g., a GPU operation).

taskflow.emplace([&](){
  auto fu = std::async([](){ std::sleep(100s); });
  executor.corun_until([](){
    return fu.wait_for(std::chrono::seconds(0)) == future_status::ready;
  });
});

Thread Safety of Executor

All run_* methods of tf::Executor are thread-safe. You can safely invoke these methods from multiple threads to run different taskflows concurrently. However, the execution order of the submitted taskflows is non-deterministic and determined by the runtime scheduler.

tf::Executor executor;
for(int i=0; i<10; ++i) {
  std::thread([i, &](){
    // ... modify my taskflow at i
    executor.run(taskflows[i]);  // run my taskflow at i
  }).detach();
}
executor.wait_for_all();

Query the Worker ID

Each worker thread in a tf::Executor is assigned a unique integer identifier in the range [0, N), where N is the number of worker threads in the executor. You can query the identifier of the calling thread using tf::Executor::this_worker_id. If the calling thread is not a worker of the executor, the method returns -1. This functionality is particularly useful for establishing a one-to-one mapping between worker threads and application-specific data structures.

std::vector<int> worker_vectors[8];       // one vector per worker

tf::Taskflow taskflow;
tf::Executor executor(8);                 // an executor of eight workers

assert(executor.this_worker_id() == -1);  // master thread is not a worker

taskflow.emplace([&](){
  int id = executor.this_worker_id();     // in the range [0, 8)
  auto& vec = worker_vectors[worker_id];
  // ...
});

Observe Thread Activities

You can observe thread activities in an executor when a worker thread participates in executing a task and leaves the execution using tf::ObserverInterface – an interface class that provides a set of methods for you to define what to do when a thread enters and leaves the execution context of a task.

class ObserverInterface {
  virtual ~ObserverInterface() = default;
  virtual void set_up(size_t num_workers) = 0;
  virtual void on_entry(tf::WorkerView worker_view, tf::TaskView task_view) = 0;
  virtual void on_exit(tf::WorkerView worker_view, tf::TaskView task_view) = 0;
};

There are three methods you must define in your derived class, tf::ObserverInterface::set_up, tf::ObserverInterface::on_entry, and tf::ObserverInterface::on_exit. The method, tf::ObserverInterface::set_up, is a constructor-like method that will be called by the executor when the observer is constructed. It passes an argument of the number of workers to observer in the executor. You may use it to preallocate or initialize data storage, e.g., an independent vector for each worker. The methods, tf::ObserverInterface::on_entry and tf::ObserverInterface::on_exit, are called by a worker thread before and after the execution context of a task, respectively. Both methods provide immutable access to the underlying worker and the running task using tf::WorkerView and tf::TaskView. You may use them to record timepoints and calculate the elapsed time of a task.

You can associate an executor with one or multiple observers (though one is common) using tf::Executor::make_observer. We use std::shared_ptr to manage the ownership of an observer. The executor loops through each observer and invoke the corresponding methods accordingly.

#include <taskflow/taskflow.hpp>

struct MyObserver : public tf::ObserverInterface {

  MyObserver(const std::string& name) {
    std::cout << "constructing observer " << name << '\n';
  }

  void set_up(size_t num_workers) override final {
    std::cout << "setting up observer with " << num_workers << " workers\n";
  }

  void on_entry(tf::WorkerView w, tf::TaskView tv) override final {
    std::ostringstream oss;
    oss << "worker " << w.id() << " ready to run " << tv.name() << '\n';
    std::cout << oss.str();
  }

  void on_exit(tf::WorkerView w, tf::TaskView tv) override final {
    std::ostringstream oss;
    oss << "worker " << w.id() << " finished running " << tv.name() << '\n';
    std::cout << oss.str();
  }

};

int main(){

  tf::Executor executor(4);

  // Create a taskflow of eight tasks
  tf::Taskflow taskflow;

  auto A = taskflow.emplace([] () { std::cout << "1\n"; }).name("A");
  auto B = taskflow.emplace([] () { std::cout << "2\n"; }).name("B");
  auto C = taskflow.emplace([] () { std::cout << "3\n"; }).name("C");
  auto D = taskflow.emplace([] () { std::cout << "4\n"; }).name("D");
  auto E = taskflow.emplace([] () { std::cout << "5\n"; }).name("E");
  auto F = taskflow.emplace([] () { std::cout << "6\n"; }).name("F");
  auto G = taskflow.emplace([] () { std::cout << "7\n"; }).name("G");
  auto H = taskflow.emplace([] () { std::cout << "8\n"; }).name("H");

  // create an observer
  std::shared_ptr<MyObserver> observer = executor.make_observer<MyObserver>(
    "MyObserver"
  );

  // run the taskflow
  executor.run(taskflow).get();

  // remove the observer (optional)
  executor.remove_observer(std::move(observer));

  return 0;
}

The above code produces the following output:

constructing observer MyObserver
setting up observer with 4 workers
worker 2 ready to run A
1
worker 2 finished running A
worker 2 ready to run B
2
worker 1 ready to run C
worker 2 finished running B
3
worker 2 ready to run D
worker 3 ready to run E
worker 1 finished running C
4
5
worker 1 ready to run F
worker 2 finished running D
worker 3 finished running E
6
worker 2 ready to run G
worker 3 ready to run H
worker 1 finished running F
7
8
worker 2 finished running G
worker 3 finished running H

It is expected each line of std::cout interleaves with each other as there are four workers participating in task scheduling. However, the ready message always appears before the corresponding task message (e.g., numbers) and then the finished message.

Modify Worker Property

You can change the property of each worker thread from its executor, such as assigning thread-processor affinity before the worker enters the scheduler loop and post-processing additional information after the worker leaves the scheduler loop, by passing an instance derived from tf::WorkerInterface to the executor. The example demonstrates the usage of tf::WorkerInterface to affine a worker to a specific CPU core equal to its id on a linux platform:

// affine the given thread to the given core index (linux-specific)
bool affine(std::thread& thread, unsigned int core_id) {
  cpu_set_t cpuset;
  CPU_ZERO(&cpuset);
  CPU_SET(core_id, &cpuset);
  pthread_t native_handle = thread.native_handle();
  return pthread_setaffinity_np(native_handle, sizeof(cpu_set_t), &cpuset) == 0;
}

class CustomWorkerBehavior : public tf::WorkerInterface {

  public:
  
  // to call before the worker enters the scheduling loop
  void scheduler_prologue(tf::Worker& w) override {
    printf("worker %lu prepares to enter the work-stealing loop\n", w.id());
    
    // now affine the worker to a particular CPU core equal to its id
    if(affine(w.thread(), w.id())) {
      printf("successfully affines worker %lu to CPU core %lu\n", w.id(), w.id());
    }
    else {
      printf("failed to affine worker %lu to CPU core %lu\n", w.id(), w.id());
    }
  }

  // to call after the worker leaves the scheduling loop
  void scheduler_epilogue(tf::Worker& w, std::exception_ptr) override {
    printf("worker %lu left the work-stealing loop\n", w.id());
  }
};

int main() {
  tf::Executor executor(4, tf::make_worker_interface<CustomWorkerBehavior>());
  return 0;
}

When running the program, we see the following one possible output:

worker 3 prepares to enter the work-stealing loop
successfully affines worker 3 to CPU core 3
worker 3 left the work-stealing loop
worker 0 prepares to enter the work-stealing loop
successfully affines worker 0 to CPU core 0
worker 0 left the work-stealing loop
worker 1 prepares to enter the work-stealing loop
worker 2 prepares to enter the work-stealing loop
successfully affines worker 1 to CPU core 1
worker 1 left the work-stealing loop
successfully affines worker 2 to CPU core 2
worker 2 left the work-stealing loop

When you create an executor, it spawns a set of worker threads to run tasks using a work-stealing scheduling algorithm. The execution logic of the scheduler and its interaction with each spawned worker via tf::WorkerInterface is given below:

for(size_t n=0; n<num_workers; n++) {
  create_thread([](Worker& worker)

    // pre-processing executor-specific worker information
    // ...

    // enter the scheduling loop
    // Here, WorkerInterface::scheduler_prologue is invoked, if any
    worker_interface->scheduler_prologue(worker);
    
    try {
      while(1) {
        perform_work_stealing_algorithm();
        if(stop) {
          break;
        }
      }
    } catch(...) {
      exception_ptr = std::current_exception();
    }

    // leaves the scheduling loop and joins this worker thread
    // Here, WorkerInterface::scheduler_epilogue is invoked, if any
    worker_interface->scheduler_epilogue(worker, exception_ptr);
  );
}