Pages

LRU Queue Device - general view

Implementing the Least Recently Used pattern for ZeroMQ was a longish but not complicated job. Here things are getting tougher, since we want to create a device to rule an exchange of messages from a bunch of clients to a bunch of workers, again using the LRU pattern.

This example is based un the Request-Replay Message Broker described in the ZGuide. I have reworked it a bit, both to adapt it to the Windows environment (no ipc protocol supported there) and both to make it clearer, at least from my point of view. I used 0MQ 2.2.0, its standard light-weight C++ wrapper, and Visual C++ 2010 as developing environment.

As they put it in the ZGuide, "Reading and writing multipart messages using the native ØMQ API is like eating a bowl of hot noodle soup, with fried chicken and extra vegetables, using a toothpick", and since that is what we are actually going to do, don't be suprised if sometime it could look to you overwhelmingly complex and messy. Be patient, try to work your way though the example one step after the other, and in the end you will find out that it is not so bad, after all.

The idea is that we want to have a class, QueueDevice, that is going to be used in this way:
void lruQueue(int nWorkers, int nClients)
{
    QueueDevice device;

    device.start(nClients, nWorkers); // 1
    device.poll(); // 2
}
1. We specify the number of clients and workers insisting on our device when we start it.
2. Then we poll on the device till some condition is reached.

The QueueDevice class is going to have a bunch of private data members:
class QueueDevice
{
    // ...
private:
    zmq::context_t context_; // 1
    zmq::socket_t backend_; // 2
    zmq::socket_t frontend_;
    std::queue<std::string> qWorkers_; // 3
    boost::thread_group thWorkers_; // 4
    boost::thread_group thClients_;

    // ...
};
1. First of all, a ZeroMQ context.
2. A couple of sockets, both ROUTER as we'll see in the ctor, one for managing the backend, AKA the connection to the workers, the other for the frontend, meaning the clients.
3. I am going to use a queue to temporary store in the device which worker is available to the clients. We want to implement an LRU pattern, so a queue is just the right data structure.
4. We are going to spawn some threads for the workers and other for the clients. Actually we could have used just one thread_group, but I guess it is more readable having two of them.

There are going to be a few public methods in the QueueDevice class:
  • A constructor to initialize the sockets.
  • A destructor to join all the threads.
  • A start() method to actually start the worker and client threads.
  • A poll() method to, well, poll on the backend and frontend sockets.
The QueueDevice start() method is going to need to access a worker() and a client() fuction representing the jobs executed on the frontend and backend sides of the device.

Finally, we'll have a bunch of dumping functions to be used by about all the many threads that are going to constitute our application to print some feedback for the user on the standard console. As usual, being std::cout a shared resource, we'll need to use a mutex to avoid unpleasent mixing up.

That should be all. In the next posts I'll go through the details. The full C++ source code for this example is available on github.

No comments:

Post a Comment