asio::io_service and thread_group lifecycle issue

Question

Looking at answers like this one, we can do stuff like:

boost::asio::io_service ioService;
boost::thread_group threadpool;
{
    boost::asio::io_service::work work(ioService);
    threadpool.create_thread(boost::bind(&boost::asio::io_service::run, ioService));
    threadpool.create_thread(boost::bind(&boost::asio::io_service::run, &ioService));
    ioService.post(boost::bind(...));
    ioService.post(boost::bind(...));
    ioService.post(boost::bind(...));
}
threadpool.join_all();

However, in my case I want to do something like:

while (condition)
{
    ioService.post(boost::bind(...));
    ioService.post(boost::bind(...));
    ioService.post(boost::bind(...));
    threadpool.join_all();

    // DO SOMETHING WITH RESULTS
}

However, the boost::asio::io_service::work work(ioService) line is out of place, and as far as I can see, I cannot recreate it without needing to create every thread in the pool again.

In my code, the thread creation overhead seems negligible (and actually better performance than previous mutex-based code), but is there a cleaner way to do this?

Answer 1

while (condition)
{
    //... stuff
    threadpool.join_all();

    //... 
}

Doesn't make any sense, because you can only join threads once. Once joined, they are gone. You don't want to be starting new threads all the time (use a thread pool + task queue¹).

Since you don't want to actually stop the threads, you probably don't want to destruct the work. If you insist, a shared_ptr<work> or optional<work> works nicely (just my_work.reset() it)

¹ Update Suggestion:

• simple thread_pool with task queue: (in boost thread throwing exception "thread_resource_error: resource temporarily unavailable")
• A queue based on io_service itself (using work) c++ work queues with blocking

UPDATE

A simple extension to "SOLUTION #2" would make it possible to wait for all tasks to have been completed, without joining the workers/destroying the pool:

  void drain() {
      unique_lock<mutex> lk(mx);
      namespace phx = boost::phoenix;
      cv.wait(lk, phx::empty(phx::ref(_queue)));
  }

Note that for reliable operation, one needs to signal the condition variable on de-queue as well:

      cv.notify_all(); // in order to signal drain

CAVEATS

1. It's an interface inviting race conditions (the queue could accept jobs from many threads, so once drain() returns, another thread could have posted a new task already)

2. This signals when the queue is empty, not when the task is completed. The queue cannot know about this, if you need this, use a barrier/signal a condition from within the task (the_work in this example). The mechanism for queuing/scheduling is not relevant there.

DEMO

Live On Coliru

#include <boost/thread.hpp>
#include <boost/phoenix.hpp>
#include <boost/optional.hpp>

using namespace boost;
using namespace boost::phoenix::arg_names;

class thread_pool
{
  private:
      mutex mx;
      condition_variable cv;

      typedef function<void()> job_t;
      std::deque<job_t> _queue;

      thread_group pool;

      boost::atomic_bool shutdown;
      static void worker_thread(thread_pool& q)
      {
          while (auto job = q.dequeue())
              (*job)();
      }

  public:
      thread_pool() : shutdown(false) {
          for (unsigned i = 0; i < boost::thread::hardware_concurrency(); ++i)
              pool.create_thread(bind(worker_thread, ref(*this)));
      }

      void enqueue(job_t job) 
      {
          lock_guard<mutex> lk(mx);
          _queue.push_back(std::move(job));

          cv.notify_one();
      }

      void drain() {
          unique_lock<mutex> lk(mx);
          namespace phx = boost::phoenix;
          cv.wait(lk, phx::empty(phx::ref(_queue)));
      }

      optional<job_t> dequeue() 
      {
          unique_lock<mutex> lk(mx);
          namespace phx = boost::phoenix;

          cv.wait(lk, phx::ref(shutdown) || !phx::empty(phx::ref(_queue)));

          if (_queue.empty())
              return none;

          auto job = std::move(_queue.front());
          _queue.pop_front();

          cv.notify_all(); // in order to signal drain

          return std::move(job);
      }

      ~thread_pool()
      {
          shutdown = true;
          {
              lock_guard<mutex> lk(mx);
              cv.notify_all();
          }

          pool.join_all();
      }
};

void the_work(int id)
{
    std::cout << "worker " << id << " entered\n";

    // no more synchronization; the pool size determines max concurrency
    std::cout << "worker " << id << " start work\n";
    this_thread::sleep_for(chrono::milliseconds(2));
    std::cout << "worker " << id << " done\n";
}

int main()
{
    thread_pool pool; // uses 1 thread per core

    for (auto i = 0ull; i < 20; ++i) {
        for (int i = 0; i < 10; ++i)
            pool.enqueue(bind(the_work, i));

        pool.drain(); // make the queue empty, leave the threads
        std::cout << "Queue empty\n";
    }

    // destructing pool joins the worker threads
}