[
https://issues.apache.org/jira/browse/SPARK-40303?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17599616#comment-17599616
]
Kris Mok commented on SPARK-40303:
----------------------------------
Nice findings [~LuciferYang]!
{quote}
After some experiments, I found when the number of parameters exceeds 50, the
performance of the case in the Jira description will significant deterioration.
{quote}
Sounds reasonable. Note that in a STD compilation, the only things that need to
be live at the method entry are the method parameters (both implicit ones like
{{this}}, and explicit ones); however, for an OSR compilation, it would be all
of the parameters/local variables that are live at the loop entry point, so in
this case both the {{doConsume}} parameters and the local variables contribute
to the problem.
Just FYI I have an old write on PrintCompilation and OSR here:
https://gist.github.com/rednaxelafx/1165804#file-notes-md
(Gee, just realized that was from 11 years ago.......)
{quote}
maybe try to make the input parameters of the `doConsume` method fixed length
will help, such as using a List or Array
{quote}
Welp, hoisting the parameters into an Arguments object is rather common in
"code splitting" in code generators. Since we're already doing codegen, it's
possible to generate tailor-made Arguments classes to retain the type
information. Using List/Array would require extra boxing for primitive types
and it's less ideal.
(An array-based box is already used in Spark SQL's codegen in the form of the
{{references}} array. Indeed the type info is lost on the interface level and
you'd have to do a cast when you get data out of it. It's still usable though.)
> The performance will be worse after codegen
> -------------------------------------------
>
> Key: SPARK-40303
> URL: https://issues.apache.org/jira/browse/SPARK-40303
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 3.4.0
> Reporter: Yuming Wang
> Priority: Major
>
> {code:scala}
> import org.apache.spark.benchmark.Benchmark
> val dir = "/tmp/spark/benchmark"
> val N = 2000000
> val columns = Range(0, 100).map(i => s"id % $i AS id$i")
> spark.range(N).selectExpr(columns: _*).write.mode("Overwrite").parquet(dir)
> // Seq(1, 2, 5, 10, 15, 25, 40, 60, 100)
> Seq(60).foreach{ cnt =>
> val selectExps = columns.take(cnt).map(_.split(" ").last).map(c =>
> s"count(distinct $c)")
> val benchmark = new Benchmark("Benchmark count distinct", N, minNumIters =
> 1)
> benchmark.addCase(s"$cnt count distinct with codegen") { _ =>
> withSQLConf(
> "spark.sql.codegen.wholeStage" -> "true",
> "spark.sql.codegen.factoryMode" -> "FALLBACK") {
> spark.read.parquet(dir).selectExpr(selectExps:
> _*).write.format("noop").mode("Overwrite").save()
> }
> }
> benchmark.addCase(s"$cnt count distinct without codegen") { _ =>
> withSQLConf(
> "spark.sql.codegen.wholeStage" -> "false",
> "spark.sql.codegen.factoryMode" -> "NO_CODEGEN") {
> spark.read.parquet(dir).selectExpr(selectExps:
> _*).write.format("noop").mode("Overwrite").save()
> }
> }
> benchmark.run()
> }
> {code}
> {noformat}
> Java HotSpot(TM) 64-Bit Server VM 1.8.0_281-b09 on Mac OS X 10.15.7
> Intel(R) Core(TM) i9-9980HK CPU @ 2.40GHz
> Benchmark count distinct: Best Time(ms) Avg Time(ms)
> Stdev(ms) Rate(M/s) Per Row(ns) Relative
> ------------------------------------------------------------------------------------------------------------------------
> 60 count distinct with codegen 628146 628146
> 0 0.0 314072.8 1.0X
> 60 count distinct without codegen 147635 147635
> 0 0.0 73817.5 4.3X
> {noformat}
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
