Details
-
Bug
-
Resolution: Fixed
-
Critical
-
None
-
None
-
None
-
1
Description
After we upgraded java client from 3.0.4 to 3.2.0, we saw regression in YCSB throughput performance. All of the runs were running on the same couchbase build (7.1.0-2021).
Avg Throughput (ops/sec), Workload A, 3 nodes, 1 bucket x 1.5B x 1KB, 1% Resident Ratio, Magma
| Java Client Version | Throughput | Job |
|---|---|---|
| 3.0.4 | 437,899 | http://perf.jenkins.couchbase.com/job/magma-nvme-ycsb/1826/ |
| 3.2.0 | 347,698 | http://perf.jenkins.couchbase.com/job/magma-nvme-ycsb/1881/ |
Avg Throughput (ops/sec), Workload B, 3 nodes, 1 bucket x 1.5B x 1KB, 1% Resident Ratio, Magma
| Java Client Version | Throughput | Job |
|---|---|---|
| 3.0.4 | 512,825 | http://perf.jenkins.couchbase.com/job/magma-nvme-ycsb/1827/ |
| 3.2.0 | 408,068 | http://perf.jenkins.couchbase.com/job/magma-nvme-ycsb/1880/ |
Avg Throughput (ops/sec), Workload C, 3 nodes, 1 bucket x 1.5B x 1KB, 1% Resident Ratio, Magma
| Java Client Version | Throughput | Job |
|---|---|---|
| 3.0.4 | 693,577 | http://perf.jenkins.couchbase.com/job/magma-nvme-ycsb/1840/ |
| 3.2.0 | 533,512 | http://perf.jenkins.couchbase.com/job/magma-nvme-ycsb/1848/ |
I compared the runs. Compared to 3.0.4, the run with 3.2.0 has fewer connections.
On the client side, the run with 3.2.0 has lower cpu utilization and lower resident set size from Java.
