Details
-
Bug
-
Status: In Progress (View Workflow)
-
Blocker
-
Resolution: Unresolved
-
10.6, 10.11, 11.4
Description
In an attempt to address MDEV-34178, I enabled spin loops on dict_index_t::lock. This helps in some workloads and causes performance degradation in others, especially when there are large numbers of concurrent connections.
It could make sense to enable spin loops on a case-by-case basis, or to introduce a configuration parameter for choosing when index tree latch acquisition should include spin loops.
Attachments
Issue Links
- relates to
-
MDEV-34178 Performance regression with IO-bound insert benchmark
- Closed