Building a multithreaded work-queue (consumer/prod

问题:

I have the following scenario: I have a single thread that is supposed to fill a container with pairs of integers (in essence, task descriptions), and I have a large number of worker threads (8-16) that should take elements from this container and perform some work.

I thought the problem could be easily solved by a blocking queue -- e.g. on item-removal, threads synchronize access to the queue, and sleep if there is no data available.

I (perhaps wrongly) assumed that something like this should exist in the STL or in boost, but I was unable to find anything.

Do I actually have to implement that thing myself ? It seems like such a common scenario...

回答1:

If you do implement it yourself, the implementation should be a fairly straightforward combination of a semaphore, a mutex, and a queue object.

Here's some pseudo-code:

Produce{
    pthread_mutex_lock(&mutex);
    queue.push_back(someObjectReference);
    pthread_mutex_unlock(&mutex);
    sem_post(&availabilitySem);
}

Consume{
    sem_wait(&availabilitySem);
    pthread_mutex_lock(&mutex);
    queue.pop_front(someObjectReference);
    pthread_mutext_unlock(&mutex);
}

回答2:

If you are on windows take a look at the agents library in VS2010 this is a core scenario.

http://msdn.microsoft.com/en-us/library/dd492627(VS.100).aspx

i.e.

//an unbounded_buffer is like a queue
unbounded_buffer<int> buf;

//you can send messages into it with send or asend
send(buf,1);

//receive will block and wait for data
int result = receive(buf)

you can use threads, 'agents' or 'tasks' to get the data out... or you can link buffers together and convert your blocking semantic producer / consumer problem to a data flow network.

回答3:

If you are on Windows and want a queue that is efficient in terms of how it manages the threads that are allowed to run to process items from it then take a look at IO Completion Ports (see here). My free server framework includes a task queue implementation that's based on IOCPs and that may also be of interest if you intend to go down this route; though it's possibly too specialised for what you want.

回答4:

I think message_queue from boost::interprocess is what you want. The second link has a usage example.

回答5:

You should take a look at ACE (Adaptive Communication Environment) and the ACE_Message_Queue. There's always boost's message_queue, but ACE is where it's at in terms of high performance concurrency.

回答6:

If you're on OSX Snow Leopard, you might want to look at Grand Central Dispatch.