Bulk operations: the evidence for thread wasted joining

[Doug Lea]

On 08/02/12 09:52, Aleksey Shipilev wrote:

> At this point, I can actually make a stronger argument. fjp-trace is now
> able to infer how many tasks there is in each worker queue. Now, you can
> cross-reference [3] and [5]. On [5] there is a color code for number of
> elements in worker queue. You can clearly see free subtasks are indeed
> are available up to the point when there are idle threads.

Thanks for more empirical arguments that JDK bulk support ought
to use completion-based processing. The main argument for doing so
is that no matter what you tell people, they will sometimes
block inside supplied functions, if only because the VM itself
is not non-blocking.

-Doug