While working with the repro for
WT-6175 I noticed that there were extended stalls during the insert phase.
- The stalls seem to end with the start of the next checkpoint
- With checkpoints disabled the stalls lasted as long as 10 minutes
- During the stalls the log reports operations that took the entire duration of the stall to complete
- They appear to have something to do with page splits.
FTDC, logs, and repro script attached. The repro creates two collections of 5 GB each with a 5 GB cache, using 50 client threads on a machine with 24 cpus.