Peter Firmstone wrote:
Gregg Wonderly wrote:
Peter Firmstone wrote:
Gregg Wonderly wrote:
I have been looking into some seemingly slow responses in several clients running simultaneously, and I see in some stack traces that there are synchronization points in DynamicPolicyProvider.implies() that seem to be heavily contended. We probably need to revisit this class and rewrite it to use copy on write mutation so that reads (the majority of activity) are completely uncontended.

Any thoughts or experience with this issue?
This sounds like a job for java.util.concurrent.ReentrantReadWriteLock! Da dat, da dat, da dat, da da! Requires Java 5, works well, the javadoc is clear too. Can you submit this as an issue on Jira?

We don't actually want to lock, we just want to use a copy on write update strategy that does lock but set volatile references to the new contents.
If you have multiple references containing object state, I'd suggest using an immutable wrapper object (no setters) containing implicit references to the objects, where it is read or replaced using a single AtomicReference. The problem you have then is whether the state your wrapping has visibility elsewhere or not, it is likely that these will all need to be created by using defensive copies in your constructor, each time you wish to update state.

In other words you want an AtomicReference, the objects being de-referenced must be accessed by getting the referent for every read, it also must not be published (an implicit reference allowed to escape) after a read. When the AtomicReference is updated it is guaranteed to be done atomically, however if the referent has escaped, any escaped (implicit) references will still refer to the old object. This isn't as easy as it sounds.

Use the compareAndSet() method, in case another write occurs, if the referent isn't the one expected (it just got updated), you can retry it.

I haven't had time to look into the details so can't comment on whether this is appropriate or not. You might want to try this and the ReentrantReadWriteLock and compare performance before deciding. The contention write lock's cause might be negligible, for code, much easier to protect, read and understand later on.

Cheers,

Peter.

Gregg Wonderly




Reply via email to