Details
-
Task
-
Resolution: Done
-
Major
-
None
-
6.6.0
-
None
Description
As per http://showfast.sc.couchbase.com/#/timeline/Linux/kv/max_ops/all performance of SyncWrite based persist writes is 1/6th of the old Observe based persist writes (see attached screenshot).
- SyncWrites: http://perf.jenkins.couchbase.com/job/ares/16398/console
- Observe: http://perf.jenkins.couchbase.com/job/ares/16408/console
Side-by-side comparison of 6.6.0-7909 builds: http://cbmonitor.sc.couchbase.com/reports/html/?snapshot=ares_660-7909_access_706c&snapshot=ares_660-7909_access_b2f5
(Note: Orange=Observe, Green=SyncWrites)
The original performance itself was non-optimal, so this causes even more concern.
- It can be a test problem. In which case, the perf test should be redefined.
- There may be a genuine perf issue lurking here, and should be investigated.
- Out-of-order returns are designed to help with this sort of test. Should be run with OoO and verified.
Attachments
Issue Links
- is duplicated by
-
MB-41683 Durability throughput with persist writes is very low
- Closed