Fault-tolerant UDP

Lecture 03

UDP is not reliable

  • a packet can be lost in the IP network
  • the packet order may be modified by the IP network
    •

    Observation of faults ?

  • remote server (hard to set up)
  • locally, almost no losses

    • We use a proxy to simulate faults in the network.

  • it is an intermediary between client and server
  • it transfers requests from client to server
  • it transfers answers from server to client
  • it can introduce drops of packets, swaps, ...
    •

    Proxy (1/12)

    Proxy (2/12)

    Proxy (3/12)

    Proxy (4/12)

    Proxy (5/12)

    Proxy (6/12)

    Proxy (7/12)

    Proxy (8/12)

    Proxy (9/12)

    Proxy (10/12)

    Proxy (11/12)

    Proxy (12/12)

    Fault-tolerance

    What can a client who did not receive its answer do?

  • His request was lost ?
  • The answer was lost ?
  • The answer/request was simply delayed ?
    •

    Only one mechanism : resend the request after a certain waiting time.

  • How long should we wait before sending again ?
  • Why not ask the server to send again ?
    •

    UDP client in Java

    The method datagramChannel.receive() is blocking by default.

    If we are blocked waiting for an answer, we cannot resend.

    We need two threads:

  • one thread to send
  • one thread to receive
    •

    Warning: obey the rules and principles seen in concurrent programming

    Producer / consumer

  • A thread safe queue, used by both threads
  • Consumer waits until there is something to consume in the queue
  • Producer waits if there is no room in the queue
    •

    Avoids active waiting and limit the memory usage.

  • Producer = listener inserts the received packet in the queue.
  • Consumer = sender (main) polls (with a timeout) the received answers from the queue
    •

    The key is the ability to put a timeout on the consumer.

    BlockingQueue

    Throws exception Special value Blocks Times out
    Insert add(e) offer(e) put(e) offer(e, time, unit)
    Remove remove() poll() take() poll(time, unit)
    Examine element() peek() not applicable not applicable

    Waiting with a timeout before resending

    Important points

    Which implementation of BlockingQueue

  • ArrayBlockingQueue? LinkedBlockingQueue?
  • Maximal size ?
    •

    Do not share the same memory via the queue = data race?

    Two threads, one DatagramChannel: which thread halts the other? which thread closes the DatagramChannel ?