[
https://issues.apache.org/jira/browse/FLINK-15171?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16993382#comment-16993382
]
Roman Grebennikov commented on FLINK-15171:
-------------------------------------------
[~pnowojski] [~roman_khachatryan] ,while running before-after benchmarks with
async-profiler, I noticed that the amount of GC produced by serializerTuple is
really different: the new code has much higher GC (about +10%) pressure due to
intermediate buffer allocations in StringValue.writeString. All the benchmarks
I've did were done on a Ryzen 7 2700 with 8 physical cores (16 with HT), but
the benchmarking Hetzner box with i7 7700 has only 4 (8 with HT).
Before PR: [^dec05.svg]
After PR: [^dec11.svg]
As GC threads are concurrent, on 8 core box they didn't interfere with the main
benchmark code (as most probably they were scheduled on the idle cores). But on
4 core box they started significantly interfere with the benchmark threads, as
there is much less CPU resources available.
I did a yet another round of improvements in the serialization code to avoid
additional allocations in StringValue.writeString by adding a static
ThreadLocal buffer for short strings, the same trick was done for
StringValue.readString (but it looks a bit ugly, as thread-locals are
dangerous). But anyway this seems to fix the issue with the too heavy GC
pressure. I will make a PR today when I will be able to finish all the
benchmarks to validate the results.
> Performance regression in serialisation benchmarks
> --------------------------------------------------
>
> Key: FLINK-15171
> URL: https://issues.apache.org/jira/browse/FLINK-15171
> Project: Flink
> Issue Type: Bug
> Components: API / Type Serialization System, Benchmarks
> Affects Versions: 1.10.0
> Reporter: Piotr Nowojski
> Assignee: Roman Khachatryan
> Priority: Blocker
> Fix For: 1.10.0
>
> Attachments: dec05.svg, dec11.svg
>
>
> There is quite significant performance regression in serialisation benchmarks
> in the commit range 2ecf7ca..9320f34 (which includes FLINK-14346).
> http://codespeed.dak8s.net:8000/timeline/?ben=serializerTuple&env=2
> http://codespeed.dak8s.net:8000/timeline/?ben=serializerRow&env=2
> http://codespeed.dak8s.net:8000/timeline/?ben=serializerPojo&env=2
> it coincides with the performance improvement for heavy strings
> http://codespeed.dak8s.net:8000/timeline/?ben=serializerHeavyString&env=2
> it might be caused by some accidental change in the benchmarking code
> (changing parallelism in one benchmarks is carried on to the next one?) or in
> the code itself.
> CC [~rgrebennikov] [~AHeise]
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
