Reworked missed work avoidance mechanism #417
Conversation
pavelkumbrasev
requested review from
aleksei-fedotov,
alex-katranov and
anton-potapov
|
@aleksei-fedotov @anton-potapov @alexey-katranov, could you please review this patch? |
Yep. Working on it! |
| //! Index of the element following the last ready task in the deque. | ||
| /** Modified by the owner thread. **/ | ||
| std::atomic<std::size_t> tail; | ||
| std::atomic<std::size_t> shadow_tail; |
There was a problem hiding this comment.
Please add a comment what shadow_tail is for and how different it is from the tail.
| //! Index of the first ready task in the deque. | ||
| /** Modified by thieves, and by the owner during compaction/reallocation **/ | ||
| std::atomic<std::size_t> head; | ||
| std::atomic<std::size_t> shadow_head; |
There was a problem hiding this comment.
Please add a comment what shadow_head is for and how different it is from the head.
Signed-off-by: pavelkumbrasev <[email protected]>
Signed-off-by: pkumbras <[email protected]>
akukanov
force-pushed
the
dev/pavelkumbrasev/reworked_missed_wake_up_avoidance
branch
from
e3f2165 to
40b4e23
Compare
|
Rebased on top of the current code base. However, I think the patch needs to be revisited first. Two tests failing on ARM likely mean that "shadowing" the head and the tail may require some extra synchronization.
Update: apparently, the patch does not even achieve its goal, because |
|
Reimplemented in a different way in #1779; closing this PR. |
akukanov
deleted the
dev/pavelkumbrasev/reworked_missed_wake_up_avoidance
branch
4 participants
When threads execute
get_taskorsteal_taskthey can omit unsuitable tasks (different isolation or affinity). The omitted tasks are virtually extracted from task pool, that makes them temporary invisible for other threads. It might lead to situation when another thread marks the arenaEMPTY(as a result ofis_out_of_work()), while these tasks are still in the arena. If tasks were omitted we need to notify the arena like new work was added (callingadvertise_new_work()).The proposed idea is that omit tasks in such way that other threads can not observe it. It will prevent unnecessary arena state transition and premature threads leaving.
Therefore, the patch introduce shadow tail and head which move after real task extracted(do not participate in tasks omitting mechanism).
Signed-off-by: pavelkumbrasev [email protected]