324 commits

Author	SHA1	Message	Date
Simon Cruanes	3956fb6566	fix ws_pool: no work stealing for pools of 1 worker ... there would be a loop because it'd try to find the index of another worker to steal from, but loop forever because there is no other worker.	2023-10-25 22:33:08 -04:00
Simon Cruanes	894851f6e8	comments in ws_deque	2023-10-25 22:32:42 -04:00
Simon Cruanes	530507d84e	fix fifo_pool: raise Shutdown, not Closed, in run	2023-10-25 22:11:08 -04:00
Simon Cruanes	3f720241b2	feat pool: use Shutdown when running a task after it's closed	2023-10-25 22:10:10 -04:00
Simon Cruanes	a89c0ce4f2	breaking: make Runner.t abstract	2023-10-25 21:59:22 -04:00
Simon Cruanes	6452ca89d1	rename Simple_pool into Fifo_pool, update doc	2023-10-25 21:55:29 -04:00
Simon Cruanes	db33bec13f	wip: better work stealing pool	2023-10-25 12:11:41 -04:00
Simon Cruanes	fdc188c291	wip: debug pool	2023-10-25 09:57:19 -04:00
Simon Cruanes	3d7e272d01	perf ws_deque: use bitmasks instead of modulo op	2023-10-25 09:57:05 -04:00
Simon Cruanes	ef05146e03	tweak retry thresholds in pool	2023-10-25 00:28:19 -04:00
Simon Cruanes	e67ab53f9f	feat pool: rewrite main pool to use work stealing ... there's a single blocking queue, and one WS_queue per worker. Scheduling into the pool from a worker (e.g. via fork_join or explicitly) will push into this WS queue; otherwise it goes into the main blocking queue. Workers will always try to empty their local queue first, then try to work steal, then block on the main queue.	2023-10-25 00:19:34 -04:00
Simon Cruanes	f2e9f99b36	perf ws_deque: implement shrinking and a push optim	2023-10-25 00:17:50 -04:00
Simon Cruanes	95b27b3a70	feat: add Ws_deque_	2023-10-25 00:09:28 -04:00
Simon Cruanes	0cec78eb30	refactor: move some common code to Suspend_	2023-10-25 00:09:09 -04:00
Simon Cruanes	4c4b720306	feat: add Simple_pool, with the naive single-queue implementation	2023-10-25 00:08:48 -04:00
Simon Cruanes	d15bfb07f2	fix pool: rework scheduler to use one condition	2023-10-24 13:47:53 -04:00
Simon Cruanes	69faea0bcb	wip: have only one condition in pool	2023-10-24 12:53:19 -04:00
Simon Cruanes	faeb95b49d	fix pool: on shutdown, finish reading from all queues	2023-10-24 09:53:18 -04:00
Simon Cruanes	b8a31b088f	fix @since tags	2023-10-19 11:08:28 -04:00
Simon Cruanes	d60e830e90	fixes for blocking queue	2023-10-12 22:24:28 -04:00
Simon Cruanes	3e8e734cf7	doc	2023-10-12 22:24:28 -04:00
Simon Cruanes	39d4e61331	some doc	2023-10-12 22:24:28 -04:00
Simon Cruanes	3834bf0796	add Bounded_queue	2023-10-12 22:24:28 -04:00
Simon Cruanes	4059903e09	moonpool: update interface of Blocking_queue	2023-09-14 22:32:58 -04:00
Simon Cruanes	ace43c0852	add Bb_queue.transfer	2023-09-14 22:30:44 -04:00
Simon Cruanes	f90773a99a	add Bb_queue.to_{iter,gen,seq}	2023-09-12 23:07:04 -04:00
Simon Cruanes	2068088255	some comments	2023-08-29 14:26:55 -04:00
Simon Cruanes	0345b65fb4	fix build on < 5.0	2023-08-29 14:22:14 -04:00
Simon Cruanes	c1b6312cad	Update src/d_pool_.ml	2023-08-29 14:22:13 -04:00
David Allsopp	c0db72b40c	Quickly hacked version joining a previous domain ... Ensure domains have been cleaned up before re-using a slot.	2023-08-29 14:22:13 -04:00
Simon Cruanes	25d42d5b8c	doc	2023-08-29 14:22:13 -04:00
Simon Cruanes	0f670c47d3	perf: in Bb_queue, only signal condition on push if queue was empty	2023-08-23 21:31:42 -04:00
Simon Cruanes	18d5bad2a9	domain pool: let domains live a bit longer ... this is an optimization to ensure we don't stop/start domains too often, which harms performance really badly.	2023-08-13 22:48:04 -04:00
Simon Cruanes	ed531e68e1	fix: race condition in shutdown, we need to wait for domain to quit ... risk is a tight loop of `Pool.with_`, where by not waiting for the pool to entirely shutdown (including the domains, potentially) we risk running out of domains in the next iterations.	2023-08-13 22:24:18 -04:00
Simon Cruanes	e38ee31b93	bugfix: forgot to dispose of the worker state on winding down	2023-08-12 22:07:08 -04:00
Simon Cruanes	9db0a9fe28	full lifecycle for worker domains ... domains can now stop when all worker threads running on them are done	2023-08-12 14:09:35 -04:00
Simon Cruanes	5680938a6c	fix: generalize type of create_arg	2023-08-01 12:34:13 -04:00
Simon Cruanes	aa0dea3e34	add Fut.{reify_error,bind_reify_error}	2023-07-26 16:22:42 -04:00
Simon Cruanes	8e9628ac81	prepare for 0.3	2023-07-16 23:36:34 -04:00
Simon Cruanes	20240a6190	fix: Fork_join.for_ 0 must not hang	2023-07-13 12:23:25 -04:00
Simon Cruanes	d995992917	remove Fork_join.map_reduce_commutative ... A bit too specialized, and not hard to do with `for_`. Let's see if it shows to be really useful.	2023-07-10 23:20:24 -04:00
Simon Cruanes	9ce94fd242	feat fork_join: add map_array and map_list	2023-07-10 23:18:26 -04:00
Simon Cruanes	1cb5342092	expose Runner	2023-07-10 22:48:04 -04:00
Simon Cruanes	6c73afbe5b	perf: optimize Fut.for_ a bit ... remove intermediate array.	2023-07-10 11:00:34 -04:00
Simon Cruanes	0f4e115bc0	doc	2023-07-10 01:31:12 -04:00
Simon Cruanes	f081c0f8ad	tweak: guess of chunk_size in Fork_join	2023-07-10 01:23:30 -04:00
Simon Cruanes	2326ae0078	modify Fork_join.for_ to avoid the allocator. ... The allocator means calling a closure at each step, which means local a reference will have to be heap allocated (and worse, that floats will be unboxed). Instead we give the function a pair of low,high bounds for a local for.	2023-07-10 01:19:16 -04:00
Simon Cruanes	55f831bc8b	add Fork_join.{for_,map_reduce_commutative}	2023-07-10 01:14:16 -04:00
Simon Cruanes	858755e812	feat: add Pool.with_	2023-07-10 01:14:07 -04:00
Simon Cruanes	427c462778	compat 4.x	2023-07-09 22:05:42 -04:00