Per-thread Haskell thread list/numbering (remove global lock from thread allocation)
|Reported by:||ezyang||Owned by:||simonmar|
|Keywords:||Cc:||simonmar, rrnewton@…, carter.schonwald@…|
|Type of failure:||Runtime performance bug||Test Case:|
|Related Tickets:||Differential Revisions:|
Currently, we take out the scheduler lock when a thread is allocated, in order to make sure we allocate a fresh thread ID and can attach it to the global thread list in the nursery. If we partition the thread ID space per originating capability and maintain thread lists per OS thread, we can remove this lock.
Marking low priority because no one seems to be running into this as their particular scaling problem (and there may be other reasons why high thread churn doesn't scale across cores).