Details
-
Bug
-
Status: In Progress (View Workflow)
-
Critical
-
Resolution: Unresolved
-
11.4
-
Related to performance
-
Q4/2025 Server Maintenance
Description
Every caller of buf_read_page() is expecting to gain access to the block. If that function returned a buffer-fixed block descriptor, we could avoid some busy work, such as a redundant lookup of buf_pool.page_hash.
This change was initially part of the 11.4 development version of MDEV-34431, but there are some issues with it, apparently some hangs due to not released buffer-fix under some special circumstances. In my own performance testing yesterday, I observed a 4% performance improvement in an I/O bound read-write workload where the active working set does not fit in the buffer pool.
Attachments
Issue Links
- blocks
-
MDEV-37115 Reverse index scan incurs unnecessary overhead
-
- Stalled
-
- relates to
-
MDEV-34431 More fine grained control of spin loops could be useful
-
- Stalled
-