Details
-
Task
-
Resolution: Fixed
-
Major
-
2.0.1
-
Security Level: Public
-
None
-
PCI Team - Sprint 3
Description
Whenever starting a perf test against cluster_run, such as lucky8-3d.conf for example, the parameter named 'items' in the test configuration, which refers to the number of items to load to the cluster before the index and access phases, is less than items, typically N / 4. For example, running the following test:
$ ./testrunner -i ../cluster_run-4.ini -c conf/perf/lucky8-3d.conf -p load_phase=1,index_phase=1,access_phase=1,stats=0
Where lucky8-3d.conf defines items as 8 000 000:
https://github.com/couchbase/testrunner/blob/master/conf/perf/lucky8-3d.conf#L18
cluster-run-4.ini file is:
[global]
username:root
password:couchbase
[membase]
rest_username:Administrator
rest_password:asdasd
[servers]
1:127.0.0.1_9000
2:127.0.0.1_9001
3:127.0.0.1_9002
4:127.0.0.1_9003
[127.0.0.1_9000]
ip:127.0.0.1
port:9000
[127.0.0.1_9001]
ip:127.0.0.1
port:9001
[clients]
1:127.0.0.1
The load phase ends up loading only 2M items, and after the test finishes, there's still about 2M items.
Same happens with lucky8.conf and single node clusters. Only N / 4 items end up being loaded to the cluster.
Some quick investigation, found out that removing the division by "num_clients" in the following line fixes the issue:
https://github.com/membase/testrunner/blob/master/pytests/performance/eperf.py#L448