A High Performance Inter-Thread Messaging Library
Michael Barker mikeb01@gmail.com
- Bug fix, race condition in SequenceGroup when removing Sequences and getting current value
- Bug fix, potential race condition in BlockingWaitStrategy.
- Bug fix set initial SequenceGroup value to -1 (Issue #27).
- Deprecate timeout methods that will be removed in version 3.
- Bug fix, correct OSGI metadata.
- Remove unnecessary code in wait strategies.
- Remove deprecated timeout methods.
- Added OSGI metadata to jar file.
- Removed PaddedAtomicLong and use Sequence in all places.
- Fix various generics warnings.
- Change Sequence implementation to work around IBM JDK bug and improve performance by ~10%.
- Add a remainingCapacity() call to the Sequencer class.
- Deprecate timeout methods for publishing.
- Add tryNext and tryPublishEvent for events that shouldn't block during delivery.
- Small performance enhancement for MultithreadClaimStrategy.
- Create new MultithreadClaimStrategy that works between when threads are highly contended. Previous implementation is now called MultithreadLowContentionClaimStrategy
- Fix for bug where EventProcessors weren't being added as gating sequences to the ring buffer.
- Fix range tracking bug in Histogram
- Artefacts made available via maven central repository. (groupId:com.googlecode.disruptor, artifactId:disruptor) See UsingDisruptorInYourProject for details.
- Changed construction API to allow user supplied claim and wait strategies
- Added AggregateEventHandler to support multiple EventHandlers from a single BatchEventProcessor
- Support exception handling from LifecycleAware
- Added timeout API calls for claiming a sequence when publishing
- Use LockSupport.parkNanos() instead of Thread.sleep() to reduce latency
- Reworked performance tests to better support profiling and use LinkedBlockingQueue for comparison because it performs better on the latest processors
- Minor bugfixes
- Introduced WorkerPool to allow the one time consumption of events by a worker in a pool of EventProcessors.
- New internal implementation of SequenceGroup which is lock free at all times and garbage free for get and set operations.
- SequenceBarrier now checks alert status on every call whether it is blocking or not.
- Added scripts in preparation for publishing binaries to maven repository.
- Bugfix for supporting SequenceReportingEventHandler from DSL. (issue 9)
- Bugfix for multi-threaded publishing to multiple ring buffers (issue 10)
- Change SequenceBarrier to always check alert status before entering waitFor cycle. Previously this was only checked when the requested sequence was not available.
- Change ClaimStrategy to not spin when the buffer has no available capacity, instead go straight to yielding to allow event processors to catch up.
- Changed RingBuffer and publisher API so any mutable object can be placed in the RingBuffer without having to extend AbstractEvent
- Added EventPublisher implementation to allow the publishing steps to be combined into one action
- DisruptorWizard has been renamed to Disruptor with added support for any subtype of EventProcessor
- Introduced a new Sequencer class that allows the Disruptor to be applied to other data structures such as multiple arrays. This can be a very useful pattern when multiple event processors work on the same event and you want to avoid false sharing. The Diamond for FizzBuzz is a good example of the issue. It is also higher performance by avoiding the pointer indirection when arrays of primitives are used.
- Further increased performance and scalability by reducing false sharing.
- Added progressive backoff strategy to the MultiThreadedClaimStrategy to prevent publisher getting into the claim cycle when the buffer is full because of a slow EventProcessor.
- Significantly improved performance to WaitStrategy.Option.BLOCKING
- Introduced SequenceGroup to allow dynamic registration of EventProcessors.
- Rework of "False Sharing" prevention which makes the performance much more predictable across all platforms. Special thanks to Jeff Hain for helping focus in on a solution.
- Renaming mistake for publishEventAtSequence should have been claimEventAtSequence
- Fixed bug in YieldingStrategy that was busy spinning more than yielding and introduced SleepingStrategy
- Removed code duplication in Unicast perf tests for expected result
- New API to reflect naming changes
- Producer -> Publisher
- Entry -> Event
- Consumer -> EventProcessor
- ConsumerBarrier -> DependencyBarrier
- ProducerBarrier has been incorporated into the RingBuffer for ease of use
- DisruptorWizard integrated for fluent API dependency graph construction
- Rework of sequence tracking to avoid false sharing on Java 7, plus avoid mega-morphic calls to make better use of the instruction cache
- Reduced usage of memory barriers where possible
- WaitStrategy.YIELDING initially spins for a short period to reduce latency
- Major performance improvement giving more than a 2X increase for throughput across most use cases.
- ProducerBarrier change to yield after busy spinning for a while. This may help the situation when the the number of producers exceeds the number of cores.
- Bug fix for setting the sequence in the ForceFillProducerBarrier.
- Code syntax tidy up.
- Bug fix for regression introduced inlining multi-thread producer commit tracking code. This was a critical bug for the multi-threaded producer scenario.
- Added new ProducerBarrier method for claiming a batch of sequences. This feature can give a significant throughput increase.
- Off by one regression bug in ProducerBarrier introduced in 1.0.9.
- Clarified the algorithm for initial cursor value in the ClaimStrategy.
- Added Apache 2.0 licence and comments.
- Small performance improvements to producers barriers and BatchConsumer.
- Bugfix for BatchConsumer sequence update when using SequenceTrackingHandler to ensure sequence is always updated at the end of a batch regardless.
- Factored out LifecycleAware interface to allowing consumers handlers to be notified when their thread starts and shuts down.
- Cache minimum consumer sequence in producer barriers. This helps make the performance more predictable on Nehalem processors and greater on earlier Core 2 processors.
- Removed Entry interface. All Entries must now extend AbstractEntry.
- Made setSequence package private on AbstractEntry for encapsulation.