What is the advantage of strand in boost asio?

Think of a system where a single io_service manages sockets for hundreds of network connections. To be able to parallelize workload, the system maintains a pool of worker threads that call io_service::run.

Now most of the operations in such a system can just run in parallel. But some will have to be serialized. For example, you probably would not want multiple write operations on the same socket to happen concurrently. You would then use one strand per socket to synchronize writes: Writes on distinct sockets can still happen at the same time, while writes to same sockets will be serialized. The worker threads do not have to care about synchronization or different sockets, they just grab whatever io_service::run hands them.

One might ask: Why can't we just use mutex instead for synchronization? The advantage of strand is that a worker thread will not get scheduled in the first place if the strand is already being worked on. With a mutex, the worker thread would get the callback and then would block on the lock attempt, preventing the thread from doing any useful work until the mutex becomes available.

I know it is too old, but hope it will help new users to understand with examples. Read comments in the code

#define BOOST_DATE_TIME_NO_LIB
#define BOOST_REGEX_NO_LIB

#include <boost/asio.hpp>
#include <boost/shared_ptr.hpp>
#include <boost/thread.hpp>
#include <boost/thread/mutex.hpp>
#include <boost/bind.hpp>
#include <iostream>

boost::mutex global_stream_lock;

void WorkerThread(boost::shared_ptr<boost::asio::io_service> iosvc, int counter) {
    global_stream_lock.lock();
    std::cout << "Thread " << std::this_thread::get_id() << ", " << counter << " Start.\n";
    global_stream_lock.unlock();

    iosvc->run();

    global_stream_lock.lock();
    std::cout << "Thread " << counter << " End.\n";
    global_stream_lock.unlock();
}

void async_send_handler(int number) {
    std::cout << "Number: " << number << ", threadID: " << std::this_thread::get_id() << std::endl;
}

int main(void) {
    boost::shared_ptr<boost::asio::io_service> io_svc(
        new boost::asio::io_service
    );

    boost::shared_ptr<boost::asio::io_service::work> worker(
        new boost::asio::io_service::work(*io_svc)
    );

    boost::asio::io_service::strand strand(*io_svc);

    global_stream_lock.lock();
    std::cout << "The program will exit once all work has finished.\n";
    global_stream_lock.unlock();

    boost::thread_group threads;
    for( int i = 1; i <= 5; i++ )
        threads.create_thread(boost::bind(&WorkerThread, io_svc, i));

    boost::this_thread::sleep(boost::posix_time::milliseconds(500));

    // Imagine you are invoking async_send on tcp or udp socket several times
    // and you want the handlers of this async_send call to be invoked sequentially

    // This code is almost equal to calling handlers of socket.async_send.
    // The handlers are invoked concurently and the order might be arbitrary
    io_svc->post(boost::bind(&async_send_handler, 1));
    io_svc->post(boost::bind(&async_send_handler, 2));
    io_svc->post(boost::bind(&async_send_handler, 3));
    io_svc->post(boost::bind(&async_send_handler, 4));
    io_svc->post(boost::bind(&async_send_handler, 5));

    // This code will do what you exactly want;
    // It will execute the handlers sequentially in that order
    strand.post(boost::bind(&async_send_handler, 1));
    strand.post(boost::bind(&async_send_handler, 2));
    strand.post(boost::bind(&async_send_handler, 3));
    strand.post(boost::bind(&async_send_handler, 4));
    strand.post(boost::bind(&async_send_handler, 5));

    worker.reset();

    threads.join_all();

    return 0;
}