Sunday, September 03, 2006

Latches in Java

David Holmes provides an excellent description of the difference between java.util.concurrent.CountDownLatch and java.util.concurrent.CyclicBarrier on the concurrency interest mailing list. I have posted it here for those of you who don't subscribe to the list and may not have a good grasp of these classes. And since I use CountDownLatch in code I've posted in the past and in code I'm planning on posting I'm happy to have such a clear explanation that I can link to in future entries.

... As Brian describes one notion of state "latching" is that it progresses to a terminal value from whence it no longer changes. It is permanently "latched". A synchronization object that behaves in this way is the CountDownLatch - once "open" it never closes and can't be reset.

More generally though latches can be reset - consider a digital flip-flop such as a "gated D-latch", the flip-flop latches the value of the data when the gate is pulsed/strobed; if you change the data without changing the gate then the latched value is unchanged, but change the gate and the latched value is updated.

In synchronization object terms a latch is sometimes called a "gate" - the connotation being that if the gate is open anyone/everyone can pass through; while if it is shut no one can pass through. The CountDownLatch operates this way, but a more general "gate" is the CyclicBarrier which can also be reset (and automatically does so). Of course the semantics of CountDownLatch and CyclicBarrier are somewhat different. CyclicBarrier is what can be called a "weighted gate" or "weighted bucket" - it is set up to expect N threads to arrive, when they arrive they have sufficient "weight" to open the gate, or tip the bucket - in this case the gate/bucket is spring-loaded and closes/rights-itself as soon as the threads leave, so it is ready for the next set of threads to use. CountDownLatch on the other hand is like a gate with multiple padlocks - when the last padlock is removed, the gate opens and it stays open. Aren't these analogies quaint :-) We could have defined CountDownLatch to allow reset but reset semantics are messy and usually not needed for CDL usage, in contrast barrier designs typically always use the barrier multiple times.

It seems the database folk are using the term "latch" for lightweight lock, which is an uncommon usage from a synchronization perspective and a poor choice in my view, though arguably there is an analogy between "locking a door/window" and just "latching it shut". In that sense "latching" is a weaker form of "locking". But I don't like the usage.

3 comments:

  1. Nice to see that blogposts are made based on my question :)

    I'll read the blog this evening.

    Yesterday I have posted the first entry on my blog. I have just switched from company.

    If you are interested, you can find it here:
    http://pveentjer.wordpress.com

    the old one can be found here:
    http://peter.jteam.nl

    ReplyDelete
  2. I would rather post a link to the entire thread so readers get the big picture but the mailing list doesn't have am addressable archive that doesn't require an account.

    ReplyDelete