[
https://issues.apache.org/jira/browse/FLINK-12818?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16894906#comment-16894906
]
Haibo Sun edited comment on FLINK-12818 at 7/29/19 7:19 AM:
------------------------------------------------------------
Hi [~pnowojski],
The benchmark on
`TwoInputSelectableStreamTask`/`StreamTwoInputSelectableProcessor` was also
unstable, and the original expectation of stabilizing was broken.
I made some other attempts, including upgrading `JDK 1.8` to the latest version
"1.8.0_212", closing the hyper-threading of CPUs, and disabling checkpointing,
but the benchmark is still unstable. After using VTune for analysis, it was
found that the slow JVM-fork was more time-consuming than the fast one, mainly
in the `RecordWriter#emit()` method (the stack information is shown in the
following figure). I suspect this is related to the cache miss of CPU. After
disabling checkpointing and adjusting the settings by the following code, the
benchmark becomes stable on my local machine, but it becomes unstable once
checkpointing is enabled.
*Code of Class FlinkEnvironmentContext :*
{code:java}
public class FlinkEnvironmentContext {
public StreamExecutionEnvironment env;
private final int parallelism = 1;
private final boolean objectReuse = true;
@Setup
public void setUp() throws IOException {
Configuration configuration = new Configuration();
configuration.setInteger(NettyShuffleEnvironmentOptions.NETWORK_BUFFERS_PER_CHANNEL,
2);
configuration.setInteger(NettyShuffleEnvironmentOptions.NETWORK_EXTRA_BUFFERS_PER_GATE,
0);
configuration.setString(NettyShuffleEnvironmentOptions.NETWORK_BUFFERS_MEMORY_MIN,
"1mb");
env = StreamExecutionEnvironment.createLocalEnvironment(parallelism,
configuration);
// set up the execution environment
env.setParallelism(parallelism);
env.getConfig().disableSysoutLogging();
if (objectReuse) {
env.getConfig().enableObjectReuse();
}
env.setStateBackend(new MemoryStateBackend());
}
public void execute() throws Exception {
env.execute();
}
}{code}
*Call Stack of RecordWriter#emit() :*
!RecordWriter-emit.png!
was (Author: sunhaibotb):
Hi [~pnowojski],
The benchmark on
`TwoInputSelectableStreamTask`/`StreamTwoInputSelectableProcessor` was also
unstable, and the original expectation of stabilizing was broken.
I made some other attempts, including upgrading `JDK 1.8` to the latest version
"1.8.0_212", closing the hyper-threading of CPUs, and disabling checkpointing,
but the benchmark is still unstable. After using VTune for analysis, it was
found that the slow JVM-fork was more time-consuming than the fast one, mainly
in the `RecordWriter#emit()` method (the stack information is shown in the
following figure). I suspect this is related to the cache miss of CPU. After
disabling checkpointing and adjusting the settings by the following code, the
benchmark becomes stable, but it becomes unstable once checkpointing is enabled.
*Code of Class FlinkEnvironmentContext :*
{code:java}
public class FlinkEnvironmentContext {
public StreamExecutionEnvironment env;
private final int parallelism = 1;
private final boolean objectReuse = true;
@Setup
public void setUp() throws IOException {
Configuration configuration = new Configuration();
configuration.setInteger(NettyShuffleEnvironmentOptions.NETWORK_BUFFERS_PER_CHANNEL,
2);
configuration.setInteger(NettyShuffleEnvironmentOptions.NETWORK_EXTRA_BUFFERS_PER_GATE,
0);
configuration.setString(NettyShuffleEnvironmentOptions.NETWORK_BUFFERS_MEMORY_MIN,
"1mb");
env = StreamExecutionEnvironment.createLocalEnvironment(parallelism,
configuration);
// set up the execution environment
env.setParallelism(parallelism);
env.getConfig().disableSysoutLogging();
if (objectReuse) {
env.getConfig().enableObjectReuse();
}
env.setStateBackend(new MemoryStateBackend());
}
public void execute() throws Exception {
env.execute();
}
}{code}
*Call Stack of RecordWriter#emit() :*
!RecordWriter-emit.png!
> Improve stability of twoInputMapSink benchmark
> ----------------------------------------------
>
> Key: FLINK-12818
> URL: https://issues.apache.org/jira/browse/FLINK-12818
> Project: Flink
> Issue Type: Sub-task
> Components: Benchmarks
> Reporter: Piotr Nowojski
> Priority: Critical
> Attachments: RecordWriter-emit.png
>
>
> The {{twoInputMapSink}} benchmark is very unstable over time:
> http://codespeed.dak8s.net:8000/timeline/#/?exe=1&ben=twoInputMapSink&env=2&revs=200&equid=off&quarts=on&extr=on
> It should be fixed, otherwise the benchmark can not be used.
--
This message was sent by Atlassian JIRA
(v7.6.14#76016)
