Rui Fan created FLINK-30709:
-------------------------------

             Summary: NetworkInput#emitNext() should push records to DataOutput 
in a while loop
                 Key: FLINK-30709
                 URL: https://issues.apache.org/jira/browse/FLINK-30709
             Project: Flink
          Issue Type: Sub-task
          Components: Runtime / Task
            Reporter: Rui Fan
             Fix For: 1.17.0


It's similar to FLINK-30533, FLINK-30533 focus on source operator, this JIRA 
focus on Network input.

 

Currently, each invocation of AbstractStreamTaskNetworkInput#emitNext() push at 
most one record to the given DataOutput. This unnecessarily increases the 
average Java call stack depth needed to produce a record.

Take the following program as an example. For each element produced by this 
program, Flink runtime needs to include in the call stack these 3 function 
calls:
 * StreamTask#processInput()
 * StreamOneInputProcessor#processInput()
 * AbstractStreamTaskNetworkInput#emitNext()

This ticket proposes to update AbstractStreamTaskNetworkInput#emitNext() to 
push records to DataOutput in a while loop. It improves Flink performance by 
removing an average of 3 function from the call stack needed to produce a 
record.

Here are the benchmark results obtained by running the 
[InputBenchmark#mapSink|https://github.com/apache/flink-benchmarks/blob/0bafe0e85700c889894324aadb70302381f98e03/src/main/java/org/apache/flink/benchmark/InputBenchmark.java#L55]
 with env.disableOperatorChaining(). And I run it 4 times on My Mac.

 
{code:java}
Before the proposed change, the avg is 12429.0605 ops/ms, here is detailed 
results:

Benchmark                (sourceType)   Mode  Cnt      Score     Error   Units
InputBenchmark.mapSink  F27_UNBOUNDED  thrpt   30  12339.771 ± 414.649  ops/ms
InputBenchmark.mapSink  F27_UNBOUNDED  thrpt   30  12687.872 ± 320.084  ops/ms
InputBenchmark.mapSink  F27_UNBOUNDED  thrpt   30  12256.445 ± 512.219  ops/ms
InputBenchmark.mapSink  F27_UNBOUNDED  thrpt   30  12432.154 ± 405.083  ops/ms


After the proposed change, the avg is 13836.845 ops/ms, here is detailed 
results:

Benchmark                (sourceType)   Mode  Cnt      Score     Error   Units
InputBenchmark.mapSink  F27_UNBOUNDED  thrpt   30  13092.451 ± 490.886  ops/ms
InputBenchmark.mapSink  F27_UNBOUNDED  thrpt   30  13881.138 ± 370.249  ops/ms
InputBenchmark.mapSink  F27_UNBOUNDED  thrpt   30  13960.280 ± 389.505  ops/ms
InputBenchmark.mapSink  F27_UNBOUNDED  thrpt   30  14413.511 ± 727.844  
ops/ms{code}
 

The proposed change increases throughput by 11.3%.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

Reply via email to