-
Type:
Task
-
Resolution: Fixed
-
Priority:
Major - P3
-
Affects Version/s: None
-
Component/s: DHandles
-
Security Level: Public (Available to anyone on the web)
-
Storage Engines, Storage Engines - Foundations
-
4,950.438
-
SE Foundations - 2025-11-21
-
3
Using Parca, I was able to determine why some of our disagg clusters trigger CPU alerts (and it's mostly kernel mode cpu; all the purple bars in the screenshot are kernel mode frames)
It seems that when we lock a dhandle, we spin instead of blocking waiting for it, and part of that spinning is an inefficient call into sleep(). I wonder if we could protect this call to sleep by including it in Windows builds only, to avoid jumping into kernel mode so often.
/* FIXME-WT-12037 Use a sleep to work around a Windows-specific scheduling issue. */
__wt_sleep(0, 1);
- causes
-
WT-16077 48.09% decrease in Average queries in Variant ubuntu2004-perf-tests for Task many-collection-test in Test many-collection-test
-
- Investigating
-
- is related to
-
-
- Backlog
-
- related to
-
-
- Backlog
-
-
-
- Closed
-
-
-
- Closed
-
-
-
- Closed
-
-
-
- Closed
-
-
WT-15577 Eviction server statistics should track every reason that a page isn't queued for eviction
-
- Closed
-
-
-
- Closed
-
-
-
- Closed
-
-
WT-15999
Add virtual destructor to kv_transaction_snapshot
-
- Closed
-
-
-
- Investigating
-