This library provides implementations of concurrent FIFO queues (for both general boxed and primitive unboxed values) that are fast, perform well under contention, and offer a Chan-like interface. The library may be of limited usefulness outside of x86 architectures where the fetch-and-add instruction is not available.
We export several variations of our design; some support additional
functionality while others try for lower latency by removing features or
making them more restrictive (e.g. in the
Unagi: a general-purpose near drop-in replacement for
Unagibut specialized for primitive types; this may perform better if a queue grows very large.
Unagi.Bounded: a bounded variant with blocking and non-blocking writes, and other functionality where a notion of the queue's capacity is required.
Unagi.NoBlocking: lowest latency implementations for when blocking reads aren't required.
Unagi.NoBlockingbut for primitive types.
Some of these may be deprecated in the future if they are found to provide little performance benefit, or no unique features; you should benchmark and experiment with them for your use cases, and please submit pull requests for additions to the benchmark suite that reflect what you find.
Here is an example benchmark measuring the time taken to concurrently write and read 100,000 messages, with work divided amongst increasing number of readers and writers, comparing against the top-performing queues in the standard libraries. The inset graph shows a zoomed-in view on the implementations here.
- support new criterion and GHC 7.8.3
- small performance improvement to boxed unagi
- implement a bounded variant (See issue #1)
- address issue with stale tickets when running in GHCi
- conditionally use tryReadMVar (as before) when GHC >= 7.8.3
- set proper CPP flags when running tests
- fixed build on GHC 7.6 (thanks @Noeda)
Unagi.Unboxedis now polymorphic in a new
UnagiPrimclass, which permits an optimization; defined instances are the same
- add new NoBlocking variants with reads that don’t block, omiting some overhead
- these have a new
Streaminterface for reads with even lower overhead
- these have a new
- revisited memory barriers in light of https://github.com/rrnewton/haskell-lockfree/issues/39, and document them better
tryReadChanfunctions to all variants
- get rid of upper bounds on
- fix upper bounds on atomic-primops again (made as revision to cabal metadata for 0.3.0.0
- fix some docs
- re-bump atomic-primops version; should now support 7.10
- fix missing other-modules for test suite
- fix getChanContents for GHC 7.10 (see GHC Trac #9965)
tryReadChannow returns an
(Element a, IO a)tuple, where the
sndis a blocking read action
- depend atomic-primops >= 0.8
- add non-atomic
estimatedLength, thanks to danclien