Details
-
Task
-
Resolution: Fixed
-
Major
-
None
-
6.5.0
-
None
Description
The metric for test xdcr_init_5x5_unidir_1G_dgm_comp.test - "Avg. initial XDCR rate (items/sec), 5 -> 5 (2 source nozzles, 4 target nozzles), 1 bucket x 1G x 1KB, DGM" is significantly affected by the fragmentation of the source cluster.
As a result, changes unrelated to XDCR (e.g. source cluster disk write speeds / write pattens) can affect the performance metric of the test. This is non-obvious given the test implies it's just measuring XDCR throughput.
The reason for source fragmentation having an effect is that the DCP stream from KV-Engine -> XDCR requires a disk backfill to order the DCP messages (even if the actual items are resident in memory).
An experiment to compact the source cluster after the data load phase (but before the measured access phase) removes this variability and gives more consistent results - see http://review.couchbase.org/#/c/118262/ I would suggest a change along these lines is made to these tests going forward.
Attachments
For Gerrit Dashboard: MB-37050 | ||||||
---|---|---|---|---|---|---|
# | Subject | Branch | Project | Status | CR | V |
118262,3 | MB-37050: XdcrInitTest: Compact source bucket before measurement phase | master | perfrunner | Status: ABANDONED | +2 | +1 |