[concurrency-interest] LinkedTransferQueue in ForkJoinPool
dl at cs.oswego.edu
Thu Jan 12 07:27:47 EST 2012
On 01/11/12 17:32, Dr Heinz M. Kabutz wrote:
> On my Mac 1.7.0 build 223, the ForkJoinPool still contains a
Sheesh. That was more than a year out of date even for JDK7. I hope
that OpenJDK for OSX starts putting out regular releases soon.
The handling of pool submissions has always been a difficult
engineering issue -- in classic ForkJoin use, one would expect
relatively few external submissions that recursively decompose,
which favors a low-overhead single array-based queue. However,
it turns out that a lot of people are using them as "ForkPools"
or plain ExecutorServices rather than ForkJoinPools, with tons
of small non-recursive submissions. This often works OK as is,
but can become highly contended (plus, is hostile to IO-bound tasks).
So I'm now in the midst of replacing with a distributed
random-multi-lane queue, and reworking stealing and scheduling
to improve throughput in these and other cases.
I'll post more about these upcoming changes as they get closer
(hopefully soon). In the mean time, if anyone is using
ForkJoinPools in this way and is feeling brave or curious,
I've been putting snapshots of jsr166.jar
Placing this in -Xbootclasspath/p:jsr166.jar (plus possibly
adding to CLASSPATH) replaces the j.u.c versions.
More information about the Concurrency-interest