[
https://issues.apache.org/jira/browse/FLINK-18433?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17147180#comment-17147180
]
Zhijiang commented on FLINK-18433:
----------------------------------
Thanks for all the above analysis.
>From my side, there are three main network related changes in this release:
* Buffer management for reducing in-flight data for checkpoint: this
improvement is not fully merged yet, and the current merged code would only
effect on exactly-once mode, so i am convinced to dismiss this feature.
* Netty reuses flink's network buffer for reducing memory copy and overhead: it
should not bring any additional cost in theory. If it has any side effects, it
should be reflected by micro-benchmark network stacks as well. I am also
convinced to dismiss this feature.
* Unaligned checkpoint: there are also some additional code execution even if
UC is disabled as [~AHeise] mentioned above. I also checked the channel state
restore part during StreamTask#beforeInvoke to confirm no effects in practice.
But I have not checked every related changes yet to give a convinced answer now.
If possible, maybe we can execute and compare one of the above performance
testing cases with disabling checkpoint, then we might further narrow down the
scope whether there are any regressions in basic network stack to dismiss
checkpoint & state effects.
Another suspicion is for JVM options changes which might effect the behavior of
GC in practice.
> From the end-to-end performance test results, 1.11 has a regression
> -------------------------------------------------------------------
>
> Key: FLINK-18433
> URL: https://issues.apache.org/jira/browse/FLINK-18433
> Project: Flink
> Issue Type: Bug
> Components: API / Core, API / DataStream
> Affects Versions: 1.11.0
> Environment: 3 machines
> [|https://github.com/Li-Aihua/flink/blob/test_suite_for_basic_operations_1.11/flink-end-to-end-perf-tests/flink-basic-operations/src/main/java/org/apache/flink/basic/operations/PerformanceTestJob.java]
> Reporter: Aihua Li
> Priority: Major
> Attachments: flink_11.log.gz
>
>
>
> I ran end-to-end performance tests between the Release-1.10 and Release-1.11.
> the results were as follows:
> |scenarioName|release-1.10|release-1.11| |
> |OneInput_Broadcast_LazyFromSource_ExactlyOnce_10_rocksdb|46.175|43.81333333|-5.11%|
> |OneInput_Rescale_LazyFromSource_ExactlyOnce_100_heap|211.835|200.355|-5.42%|
> |OneInput_Rebalance_LazyFromSource_ExactlyOnce_1024_rocksdb|1721.041667|1618.323333|-5.97%|
> |OneInput_KeyBy_LazyFromSource_ExactlyOnce_10_heap|46|43.615|-5.18%|
> |OneInput_Broadcast_Eager_ExactlyOnce_100_rocksdb|212.105|199.6883333|-5.85%|
> |OneInput_Rescale_Eager_ExactlyOnce_1024_heap|1754.64|1600.123333|-8.81%|
> |OneInput_Rebalance_Eager_ExactlyOnce_10_rocksdb|45.91666667|43.09833333|-6.14%|
> |OneInput_KeyBy_Eager_ExactlyOnce_100_heap|212.0816667|200.7266667|-5.35%|
> |OneInput_Broadcast_LazyFromSource_AtLeastOnce_1024_rocksdb|1718.245|1614.381667|-6.04%|
> |OneInput_Rescale_LazyFromSource_AtLeastOnce_10_heap|46.12|43.55166667|-5.57%|
> |OneInput_Rebalance_LazyFromSource_AtLeastOnce_100_rocksdb|212.0383333|200.3883333|-5.49%|
> |OneInput_KeyBy_LazyFromSource_AtLeastOnce_1024_heap|1762.048333|1606.408333|-8.83%|
> |OneInput_Broadcast_Eager_AtLeastOnce_10_rocksdb|46.05833333|43.49666667|-5.56%|
> |OneInput_Rescale_Eager_AtLeastOnce_100_heap|212.2333333|201.1883333|-5.20%|
> |OneInput_Rebalance_Eager_AtLeastOnce_1024_rocksdb|1720.663333|1616.85|-6.03%|
> |OneInput_KeyBy_Eager_AtLeastOnce_10_heap|46.14|43.62333333|-5.45%|
> |TwoInputs_Broadcast_LazyFromSource_ExactlyOnce_100_rocksdb|156.9183333|152.9566667|-2.52%|
> |TwoInputs_Rescale_LazyFromSource_ExactlyOnce_1024_heap|1415.511667|1300.1|-8.15%|
> |TwoInputs_Rebalance_LazyFromSource_ExactlyOnce_10_rocksdb|34.29666667|34.16666667|-0.38%|
> |TwoInputs_KeyBy_LazyFromSource_ExactlyOnce_100_heap|158.3533333|151.8483333|-4.11%|
> |TwoInputs_Broadcast_Eager_ExactlyOnce_1024_rocksdb|1373.406667|1300.056667|-5.34%|
> |TwoInputs_Rescale_Eager_ExactlyOnce_10_heap|34.57166667|32.09666667|-7.16%|
> |TwoInputs_Rebalance_Eager_ExactlyOnce_100_rocksdb|158.655|147.44|-7.07%|
> |TwoInputs_KeyBy_Eager_ExactlyOnce_1024_heap|1356.611667|1292.386667|-4.73%|
> |TwoInputs_Broadcast_LazyFromSource_AtLeastOnce_10_rocksdb|34.01|33.205|-2.37%|
> |TwoInputs_Rescale_LazyFromSource_AtLeastOnce_100_heap|149.5883333|145.9966667|-2.40%|
> |TwoInputs_Rebalance_LazyFromSource_AtLeastOnce_1024_rocksdb|1359.74|1299.156667|-4.46%|
> |TwoInputs_KeyBy_LazyFromSource_AtLeastOnce_10_heap|34.025|29.68333333|-12.76%|
> |TwoInputs_Broadcast_Eager_AtLeastOnce_100_rocksdb|157.3033333|151.4616667|-3.71%|
> |TwoInputs_Rescale_Eager_AtLeastOnce_1024_heap|1368.74|1293.238333|-5.52%|
> |TwoInputs_Rebalance_Eager_AtLeastOnce_10_rocksdb|34.325|33.285|-3.03%|
> |TwoInputs_KeyBy_Eager_AtLeastOnce_100_heap|162.5116667|134.375|-17.31%|
> It can be seen that the performance of 1.11 has a regression, basically
> around 5%, and the maximum regression is 17%. This needs to be checked.
> the test code:
> flink-1.10.0:
> [https://github.com/Li-Aihua/flink/blob/test_suite_for_basic_operations/flink-end-to-end-perf-tests/flink-basic-operations/src/main/java/org/apache/flink/basic/operations/PerformanceTestJob.java]
> flink-1.11.0:
> [https://github.com/Li-Aihua/flink/blob/test_suite_for_basic_operations_1.11/flink-end-to-end-perf-tests/flink-basic-operations/src/main/java/org/apache/flink/basic/operations/PerformanceTestJob.java]
> commit cmd like tis：
> bin/flink run -d -m 192.168.39.246:8081 -c
> org.apache.flink.basic.operations.PerformanceTestJob
> /home/admin/flink-basic-operations_2.11-1.10-SNAPSHOT.jar --topologyName
> OneInput --LogicalAttributesofEdges Broadcast --ScheduleMode LazyFromSource
> --CheckpointMode ExactlyOnce --recordSize 10 --stateBackend rocksdb
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
