[
https://issues.apache.org/jira/browse/FLINK-14346?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16985001#comment-16985001
]
Roman Grebennikov commented on FLINK-14346:
-------------------------------------------
[~pnowojski], as for the concerns you've pointed to:
# I'm currently working on a flink-benchmarks, trying to find a visible
performance improvements there, as it's a bit tricky to run it over a custom
flink build. But these benchmarks will only cover the case with state-heavy
jobs, not the serialization heavy inter-node communication.
# As for CPU cache friendliness, as we are always doing sequential reads and
writes, CPU will happily prefetch all the needed data automatically. So
splitting buffer to multiple chunks most probably will complicate the algorithm
without significant performance gains. If you run deserializeImproved betchmark
with -prof perf jmh profiler, it already has 0.353 clocks per instruction, so
we already do ~3 operations in parallel, limited by, most probably, RAM -> CPU
throughtput.
> Performance issue with StringSerializer
> ---------------------------------------
>
> Key: FLINK-14346
> URL: https://issues.apache.org/jira/browse/FLINK-14346
> Project: Flink
> Issue Type: Improvement
> Components: API / Type Serialization System, Benchmarks
> Affects Versions: 1.9.0
> Environment: Tested on Flink 1.9.0, adoptopenjdk 8u222.
> Reporter: Roman Grebennikov
> Priority: Major
> Labels: performance, pull-request-available
> Time Spent: 10m
> Remaining Estimate: 0h
>
> While doing a performance profiling for our Flink state-heavy streaming job,
> we found that quite a significant amount of CPU time is spent inside
> StringSerializer writing data to the underlying byte buffer. The hottest part
> of the code is the StringValue.writeString function. And replacing the
> default StringSerializer with the custom one (to just play with a baseline),
> which is just calling DataOutput.writeUTF/readUTF surprisingly yielded to
> almost 2x speedup for string serialization.
> As writeUTF and writeString have incompatible wire formats, replacing latter
> with former is not a good idea in general as it may break
> checkpoint/savepoint compatibility.
> We also did an early performance analysis of the root cause of this
> performance issue, and the main reason of JDK's writeUTF being faster is that
> it's code is not writing directly to output stream byte-by-byte, but instead
> creating an underlying temporary byte buffer. This yields to a HotSpot almost
> perfectly unrolling the main loop, which results in much better data
> parallelism.
> I've tried to port the ideas from the JVM's implementation of writeUTF back
> to StringValue.writeString, and my current result is nice, having quite
> significant speedup compared to the current implementation:
> {{[info] Benchmark Mode Cnt Score Error Units}}
> {{[info] StringSerializerBenchmark.measureJDK avgt 30 82.871 ± 1.293 ns/op}}
> {{[info] StringSerializerBenchmark.measureNew avgt 30 94.004 ± 1.491 ns/op}}
> {{[info] StringSerializerBenchmark.measureOld avgt 30 156.905 ± 3.596 ns/op}}
>
> {{Where measureJDK is the JDK's writeUTF asa baseline, measureOld is the
> current upstream implementation in Flink, and the measureNew is the improved
> one. }}
>
> {{The code for the benchmark (and the improved version of the serializer) is
> here: [https://github.com/shuttie/flink-string-serializer]}}
>
> {{Next steps:}}
> # {{More benchmarks for non-ascii strings.}}
> # {{Benchmarks for long strings.}}
> # {{Benchmarks for deserialization.}}
> # {{Tests for old-new wire format compatibility.}}
> # {{PR to the Flink codebase.}}
> {{Is there an interest for this kind of performance improvement?}}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
