[concurrency-interest] Class striped/ordered Thread Pool
dl at cs.oswego.edu
Sun May 13 07:10:23 EDT 2012
On 05/12/12 14:29, √iktor Ҡlang wrote:
> Well, the entire point is to have stripes serialized, isn't it. The problem is
> that the consistent-hashing approach will make execution serial for all stripes
> that hash the same. Which is slightly different.
If it were not for the per-producer-fifo constraint, this could
be solved by using incremental table expansion and self-adjusting
hashes, as is done in Striped64 (LongAdder etc) as well as new
Exchanger and FJ submission queues.
It is possible but not easy, and probably not fast enough,
to adapt this to preserve fifo by tracking whether any thread
has unclaimed elements, and if so, inhibiting its hash adjustment.
More information about the Concurrency-interest