Details
-
Bug
-
Resolution: Unresolved
-
Major
-
7.6.0
-
Untriaged
-
0
-
Yes
-
March-June 24
Description
Similar to MB-60986, we see throughput going down when running with multiple buckets. This seems to be a regression from 7.2.5 for multi-bucket tests, while single-bucket tests remain table
All runs below running on the ares cluster (4KV nodes with 40vCPU, 64GB). The test is: Max ops/sec, cbc-pillowfight, 2 nodes, 50/50 R/W, 512B JSON items, 1K batch size, s=1 c=1000. All runs with 7.6.0-2180 have cbcollect logs (which I'll link below):
- 1 bucket:
- 7.2.4-7069: https://perf.jenkins.couchbase.com/job/ares/39101/console. Throughput: 2,825,783 ops/sec.
- 7.6.0-2180: https://perf.jenkins.couchbase.com/job/ares/40116/console. Throughput: 2,860,522 ops/sec. Artifacts:
- https://s3-us-west-2.amazonaws.com/perf-artifacts/jenkins-ares-40116/172.23.133.11.zip
- https://s3-us-west-2.amazonaws.com/perf-artifacts/jenkins-ares-40116/172.23.133.12.zip
- https://s3-us-west-2.amazonaws.com/perf-artifacts/jenkins-ares-40116/172.23.133.13.zip
- https://s3-us-west-2.amazonaws.com/perf-artifacts/jenkins-ares-40116/172.23.133.14.zip
- 10 buckets:
- 7.2.5-7565: https://perf.jenkins.couchbase.com/job/ares/40036/console. Throughput: 2,118,579 ops/sec
- 7.6.0-2180: https://perf.jenkins.couchbase.com/job/ares/40115/console. Throughput: 1,731,008 ops/sec. Artifacts:
- https://s3-us-west-2.amazonaws.com/perf-artifacts/jenkins-ares-40115/172.23.133.11.zip
- https://s3-us-west-2.amazonaws.com/perf-artifacts/jenkins-ares-40115/172.23.133.12.zip
- https://s3-us-west-2.amazonaws.com/perf-artifacts/jenkins-ares-40115/172.23.133.13.zip
- https://s3-us-west-2.amazonaws.com/perf-artifacts/jenkins-ares-40115/172.23.133.14.zip
- 30 buckets:
- 7.2.5-7565: https://perf.jenkins.couchbase.com/job/ares/40038/console. Throughput: 2,090,508 ops/sec.
- 7.6.0-2180: https://perf.jenkins.couchbase.com/job/ares/40114/console. Throughput: 908,589 ops/sec. Artifacts:
It seems that performance degradation at a significantly higher rate in 7.6 (where throughput with 30 buckets is about half of that with 10 buckets) than in 7.2 (where difference between runs with 10 and 30 buckets is in a few percentage points).
