[
https://issues.apache.org/jira/browse/SPARK-18591?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16160388#comment-16160388
]
Takeshi Yamamuro commented on SPARK-18591:
------------------------------------------
I just kindly give a head-up for the discussion on this thread; since we've
already have LogicalPlanVisitor, we might easily realise bottom-up
transformation in SparkStrategies like
https://github.com/apache/spark/compare/master...maropu:SPARK-18591. I'm not
sure this is the good timing now (cuz, probably, I think many committers and
qualified developers spending much time on Dataset API v2 reviews and others)
to change the transformation way though, I think it'd be better to modify this
in future because the bottom-up transformation makes catalyst easily select
better physical plans based on bottom sub-tree condition (costs and
partition/sort conditions), e.g., we could easily fix SPARK-12978 and this
ticket. cc: [~smilegator]
> Replace hash-based aggregates with sort-based ones if inputs already sorted
> ---------------------------------------------------------------------------
>
> Key: SPARK-18591
> URL: https://issues.apache.org/jira/browse/SPARK-18591
> Project: Spark
> Issue Type: Improvement
> Components: SQL
> Affects Versions: 2.0.2
> Reporter: Takeshi Yamamuro
>
> Spark currently uses sort-based aggregates only in limited condition; the
> cases where spark cannot use partial aggregates and hash-based ones.
> However, if input ordering has already satisfied the requirements of
> sort-based aggregates, it seems sort-based ones are faster than the other.
> {code}
> ./bin/spark-shell --conf spark.sql.shuffle.partitions=1
> val df = spark.range(10000000).selectExpr("id AS key", "id % 10 AS
> value").sort($"key").cache
> def timer[R](block: => R): R = {
> val t0 = System.nanoTime()
> val result = block
> val t1 = System.nanoTime()
> println("Elapsed time: " + ((t1 - t0 + 0.0) / 1000000000.0)+ "s")
> result
> }
> timer {
> df.groupBy("key").count().count
> }
> // codegen'd hash aggregate
> Elapsed time: 7.116962977s
> // non-codegen'd sort aggregarte
> Elapsed time: 3.088816662s
> {code}
> If codegen'd sort-based aggregates are supported in SPARK-16844, this seems
> to make the performance gap bigger;
> {code}
> - codegen'd sort aggregate
> Elapsed time: 1.645234684s
> {code}
> Therefore, it'd be better to use sort-based ones in this case.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
