-
Type: Task
-
Resolution: Unresolved
-
Priority: Minor - P4
-
None
-
Affects Version/s: None
-
Component/s: None
-
Labels:None
The changes to WT-3636 appear to have had an unexpected but significant (positive) impact on performance of the multi-btree-zipfian-workload.wtperf read throughput. It is a consistent 33% improvement.
I have attached a screen shot of the plot. The last build at the older/slower rate is #3276 and its final changeset number is b7e2b30. Build #3277, the first showing the gain, is changeset 3b079fb. The changes in #3277 include 3 things, two of which negate each other:
WT-3640 Fix bytes-read and bytes-written statistics. (#3706) (commit: 0b3903a) (detail / githubweb) Revert "WT-3640 Fix bytes-read and bytes-written statistics. (#3706)" (commit: e68caee) (detail / githubweb) WT-3636 Account for page image sizes in cache consistently (#3703) (commit: 3b079fb) (detail / githubweb)
This leads me to believe only the changes for WT-3636 are relevant for this improvement. keith.bostic says that those changes should not have any operational impact, only statistic accuracy.
We should examine where changes are coming from. Did the cache behavior change in some way we weren't expecting? What can account for this improvement?