[
https://issues.apache.org/jira/browse/FLINK-30531?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Dong Lin updated FLINK-30531:
-----------------------------
Description:
Benchmark results show that Flink time to execute simple programs is more than
3X slower than Spark. For example, if we run the following program with object
re-use enabled and with parallelism=1, it takes roughtly 120 sec on a macbook,
whereas it takes Spark less than 40 sec to run the same logic on the same
machine.
{code:java}
DataStream<Long> stream = env.fromSequence(1, 1000000000L)
.map(x -> x)
.map(x -> x)
.map(x -> x)
.map(x -> x)
.map(x -> x).addSink(new DiscardingSink<>());
{code}
It turns out that the operator chain overhead introduced by Flink is
surprisingly high. For the above example program, Flink runtime goes through a
call stack of 24 functions to produce 1 element. And each extra map(...)
operation introduces 3 extra functions in the call stack.
Here are the 24 functions in the call stack:
{code:bash}
StreamTask#processInput
StreamOneInputProcessor#processInput
StreamTaskSourceInput#emitNext
SourceOperator#emitNext
IteratorSourceReaderBase#pollNext
SourceOutputWithWatermarks#collect
AsyncDataOutputToOutput#emitRecord
ChainingOutput#collect
StreamMap#processElement
CountingOutput#collect
ChainingOutput#collect
StreamMap#processElement
CountingOutput#collect
ChainingOutput#collect
StreamMap#processElement
CountingOutput#collect
ChainingOutput#collect
StreamMap#processElement
CountingOutput#collect
ChainingOutput#collect
StreamMap#processElement
CountingOutput#collect
ChainingOutput#collect
StreamSink#processElement
{code}
Given the evidence described above, we find the following explanations for why
Flink is slow for programs with low computation overhead:
* Operator chain currently uses pull-based loop, which has worse branch
prediction than push-based loop.
* Java's maximum inline level is less than 18 [2]. It is easy for operator
chain call stack to exceeds this limit and prevent Java from inlining function
calls, which further increases the function call overhead.
* For function calls that are not inlined, it requires looking up a virtual
table since most functions are virtual functions.
Given the understanding of the root cause described above, here are the ideas
to reduce Flink's runtime overhead:
* Update operator chain to use push-based loop to reduce the average call
stack depth for each element.
* Fuse some functions (e.g. ChainingOutput, StreamMap, CountingOutput) to
reduce the call stack depth required for each extra operation (e.g. map(...)).
[1] [https://arxiv.org/pdf/1610.09166.pdf]
[2] [https://bugs.openjdk.org/browse/JDK-8234863]
was:
Benchmark results show that Flink time to execute simple programs is more than
3X slower than Spark. For example, if we run the following program with object
re-use enabled and with parallelism=1, it takes roughtly 120 sec on a macbook,
whereas it takes Spark less than 40 sec to run the same logic on the same
machine.
{code:java}
DataStream<Long> stream = env.fromSequence(1, 1000000000L)
.map(x -> x)
.map(x -> x)
.map(x -> x)
.map(x -> x)
.map(x -> x).addSink(new DiscardingSink<>());
{code}
It turns out that the operator chain overhead introduced by Flink is
surprisingly high. For the above example program, Flink runtime goes through a
call stack of 24 functions to produce 1 element. And each extra map(...)
operation introduces 4 extra functions in the call stack.
Here are the 24 functions in the call stack:
{code:bash}
StreamTask#processInput
StreamOneInputProcessor#processInput
StreamTaskSourceInput#emitNext
SourceOperator#emitNext
IteratorSourceReaderBase#pollNext
SourceOutputWithWatermarks#collect
AsyncDataOutputToOutput#emitRecord
ChainingOutput#collect
StreamMap#processElement
CountingOutput#collect
ChainingOutput#collect
StreamMap#processElement
CountingOutput#collect
ChainingOutput#collect
StreamMap#processElement
CountingOutput#collect
ChainingOutput#collect
StreamMap#processElement
CountingOutput#collect
ChainingOutput#collect
StreamMap#processElement
CountingOutput#collect
ChainingOutput#collect
StreamSink#processElement
{code}
Given the evidence described above, we find the following explanations for why
Flink is slow for programs with low computation overhead:
* Operator chain currently uses pull-based loop, which has worse branch
prediction than push-based loop.
* Java's maximum inline level is less than 18 [2]. It is easy for operator
chain call stack to exceeds this limit and prevent Java from inlining function
calls, which further increases the function call overhead.
* For function calls that are not inlined, it requires looking up a virtual
table since most functions are virtual functions.
Given the understanding of the root cause described above, here are the ideas
to reduce Flink's runtime overhead:
* Update operator chain to use push-based loop to reduce the average call
stack depth for each element.
* Fuse some functions (e.g. ChainingOutput, StreamMap, CountingOutput) to
reduce the call stack depth required for each extra operation (e.g. map(...)).
[1] [https://arxiv.org/pdf/1610.09166.pdf]
[2] [https://bugs.openjdk.org/browse/JDK-8234863]
> Reduce operator chain call stack depth
> --------------------------------------
>
> Key: FLINK-30531
> URL: https://issues.apache.org/jira/browse/FLINK-30531
> Project: Flink
> Issue Type: Improvement
> Components: Runtime / Task
> Reporter: Dong Lin
> Assignee: Dong Lin
> Priority: Major
>
> Benchmark results show that Flink time to execute simple programs is more
> than 3X slower than Spark. For example, if we run the following program with
> object re-use enabled and with parallelism=1, it takes roughtly 120 sec on a
> macbook, whereas it takes Spark less than 40 sec to run the same logic on the
> same machine.
> {code:java}
> DataStream<Long> stream = env.fromSequence(1, 1000000000L)
> .map(x -> x)
> .map(x -> x)
> .map(x -> x)
> .map(x -> x)
> .map(x -> x).addSink(new DiscardingSink<>());
> {code}
>
> It turns out that the operator chain overhead introduced by Flink is
> surprisingly high. For the above example program, Flink runtime goes through
> a call stack of 24 functions to produce 1 element. And each extra map(...)
> operation introduces 3 extra functions in the call stack.
> Here are the 24 functions in the call stack:
> {code:bash}
> StreamTask#processInput
> StreamOneInputProcessor#processInput
> StreamTaskSourceInput#emitNext
> SourceOperator#emitNext
> IteratorSourceReaderBase#pollNext
> SourceOutputWithWatermarks#collect
> AsyncDataOutputToOutput#emitRecord
> ChainingOutput#collect
> StreamMap#processElement
> CountingOutput#collect
> ChainingOutput#collect
> StreamMap#processElement
> CountingOutput#collect
> ChainingOutput#collect
> StreamMap#processElement
> CountingOutput#collect
> ChainingOutput#collect
> StreamMap#processElement
> CountingOutput#collect
> ChainingOutput#collect
> StreamMap#processElement
> CountingOutput#collect
> ChainingOutput#collect
> StreamSink#processElement
> {code}
>
> Given the evidence described above, we find the following explanations for
> why Flink is slow for programs with low computation overhead:
> * Operator chain currently uses pull-based loop, which has worse branch
> prediction than push-based loop.
> * Java's maximum inline level is less than 18 [2]. It is easy for operator
> chain call stack to exceeds this limit and prevent Java from inlining
> function calls, which further increases the function call overhead.
> * For function calls that are not inlined, it requires looking up a virtual
> table since most functions are virtual functions.
> Given the understanding of the root cause described above, here are the ideas
> to reduce Flink's runtime overhead:
> * Update operator chain to use push-based loop to reduce the average call
> stack depth for each element.
> * Fuse some functions (e.g. ChainingOutput, StreamMap, CountingOutput) to
> reduce the call stack depth required for each extra operation (e.g. map(...)).
> [1] [https://arxiv.org/pdf/1610.09166.pdf]
> [2] [https://bugs.openjdk.org/browse/JDK-8234863]
>
>
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
