[SPARK-15985][SQL] Eliminate redundant cast from an array without null or a map without null by kiszk · Pull Request #13704 · apache/spark

kiszk · 2016-06-16T08:13:30Z

What changes were proposed in this pull request?

This PR eliminates redundant cast from an ArrayType with containsNull = false or a MapType with containsNull = false.

For example, in ArrayType case, current implementation leaves a cast cast(value#63 as array<double>).toDoubleArray. However, we can eliminate cast(value#63 as array<double>) if we know value#63 does not include null. This PR apply this elimination for ArrayType and MapType in SimplifyCasts at a plan optimization phase.

In summary, we got 1.2-1.3x performance improvements over the code before applying this PR.
Here are performance results of benchmark programs:

 test("Read array in Dataset") {
import sparkSession.implicits._
val iters = 5
val n = 1024 * 1024
val rows = 15
val benchmark = new Benchmark("Read primnitive array", n)
val rand = new Random(511)
val intDS = sparkSession.sparkContext.parallelize(0 until rows, 1)
.map(i => Array.tabulate(n)(i => i)).toDS()
intDS.count() // force to create ds
val lastElement = n - 1
val randElement = rand.nextInt(lastElement)
benchmark.addCase(s"Read int array in Dataset", numIters = iters)(iter => {
val idx0 = randElement
val idx1 = lastElement
intDS.map(a => a(0) + a(idx0) + a(idx1)).collect
})
val doubleDS = sparkSession.sparkContext.parallelize(0 until rows, 1)
.map(i => Array.tabulate(n)(i => i.toDouble)).toDS()
doubleDS.count() // force to create ds
benchmark.addCase(s"Read double array in Dataset", numIters = iters)(iter => {
val idx0 = randElement
val idx1 = lastElement
doubleDS.map(a => a(0) + a(idx0) + a(idx1)).collect
})
benchmark.run()
}
Java HotSpot(TM) 64-Bit Server VM 1.8.0_92-b14 on Mac OS X 10.10.4
Intel(R) Core(TM) i5-5257U CPU @ 2.70GHz
without this PR
Read primnitive array: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------
Read int array in Dataset 525 / 690 2.0 500.9 1.0X
Read double array in Dataset 947 / 1209 1.1 902.7 0.6X
with this PR
Read primnitive array: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
------------------------------------------------------------------------------------------------
Read int array in Dataset 400 / 492 2.6 381.5 1.0X
Read double array in Dataset 788 / 870 1.3 751.4 0.5X

An example program that originally caused this performance issue.

val ds = Seq(Array(1.0, 2.0, 3.0), Array(4.0, 5.0, 6.0)).toDS()
val ds2 = ds.map(p => {
var s = 0.0
for (i <- 0 to 2) { s += p(i) }
s
})
ds2.show
ds2.explain(true)

Plans before this PR

== Parsed Logical Plan ==
'SerializeFromObject [input[0, double, true] AS value#68]
+- 'MapElements <function1>, obj#67: double
+- 'DeserializeToObject unresolveddeserializer(upcast(getcolumnbyordinal(0, ArrayType(DoubleType,false)), ArrayType(DoubleType,false), - root class: "scala.Array").toDoubleArray), obj#66: [D
+- LocalRelation [value#63]
== Analyzed Logical Plan ==
value: double
SerializeFromObject [input[0, double, true] AS value#68]
+- MapElements <function1>, obj#67: double
+- DeserializeToObject cast(value#63 as array<double>).toDoubleArray, obj#66: [D
+- LocalRelation [value#63]
== Optimized Logical Plan ==
SerializeFromObject [input[0, double, true] AS value#68]
+- MapElements <function1>, obj#67: double
+- DeserializeToObject cast(value#63 as array<double>).toDoubleArray, obj#66: [D
+- LocalRelation [value#63]
== Physical Plan ==
*SerializeFromObject [input[0, double, true] AS value#68]
+- *MapElements <function1>, obj#67: double
+- *DeserializeToObject cast(value#63 as array<double>).toDoubleArray, obj#66: [D
+- LocalTableScan [value#63]

Plans after this PR

== Parsed Logical Plan ==
'SerializeFromObject [input[0, double, true] AS value#6]
+- 'MapElements <function1>, obj#5: double
+- 'DeserializeToObject unresolveddeserializer(upcast(getcolumnbyordinal(0, ArrayType(DoubleType,false)), ArrayType(DoubleType,false), - root class: "scala.Array").toDoubleArray), obj#4: [D
+- LocalRelation [value#1]
== Analyzed Logical Plan ==
value: double
SerializeFromObject [input[0, double, true] AS value#6]
+- MapElements <function1>, obj#5: double
+- DeserializeToObject cast(value#1 as array<double>).toDoubleArray, obj#4: [D
+- LocalRelation [value#1]
== Optimized Logical Plan ==
SerializeFromObject [input[0, double, true] AS value#6]
+- MapElements <function1>, obj#5: double
+- DeserializeToObject value#1.toDoubleArray, obj#4: [D
+- LocalRelation [value#1]
== Physical Plan ==
*SerializeFromObject [input[0, double, true] AS value#6]
+- *MapElements <function1>, obj#5: double
+- *DeserializeToObject value#1.toDoubleArray, obj#4: [D
+- LocalTableScan [value#1]

How was this patch tested?

Tested by new test cases in SimplifyCastsSuite

SparkQA · 2016-06-16T09:57:19Z

Test build #60636 has finished for PR 13704 at commit 17f17d6.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

cloud-fan · 2016-06-27T09:10:48Z

where do we ensure it?

I assume that here is a part to generate code for Java primitive arrays regarding from and to. Since Java primitive array (e.g. int[]) cannot have null value unlike SQL, I said "ensure not null in input and output arrays."
What do you think?

cloud-fan · 2016-06-27T12:49:52Z

can you also put the plan tree of the example program in PR description? thanks!

kiszk · 2016-06-27T13:48:57Z

@cloud-fan I updated the PR description by adding plan trees.

cloud-fan · 2016-06-27T14:20:18Z

I'm wondering why we have this Cast in DeserializeToObject. Isn't the value#63 already a double array?

kiszk · 2016-06-27T15:02:00Z

@cloud-fan I think value#63 is UnsafeArrayData.
When I ran a DataFrame program, I got the following trees. Since operations for DataFrame access data in UnsafeArrayData, I think that LocalRelation and LocalTableScan keep an array as UnsafeArrayData in InternalRow.
What do you think?

val df = Seq(Array(1.0, 2.0, 3.0), Array(4.0, 5.0, 6.0)).toDF()
val df2 = df.selectExpr("value[0] + value[1] + value[2]")
df2.show
df2.explain(true)
== Analyzed Logical Plan ==
((value[0] + value[1]) + value[2]): double
Project [((value#63[0] + value#63[1]) + value#63[2]) AS ((value[0] + value[1]) + value[2])#67]
+- LocalRelation [value#63]
== Optimized Logical Plan ==
LocalRelation [((value[0] + value[1]) + value[2])#67]
== Physical Plan ==
LocalTableScan [((value[0] + value[1]) + value[2])#67]

cloud-fan · 2016-06-28T12:13:48Z

Sorry I should say it more explicitly: value#63 is already double array type. UnsafeArrayData is an internal data representation of array type, so it seems weird to have the Cast there.

kiszk · 2016-06-28T12:26:31Z

I agree that value#63 is double array type. value#63 is stored as UnsafeArrayData in UnsafeRow.
<function1> in MapElements <function1>, obj#67: double is represented by Java byte code instead of Expressions. We have to pass an Java primitive array double[] instead of double array in UnsafeArrayData. I think that Cast performs this conversion of double array from UnsafeArrayData to double[]. This is an issue only for Dataset.

What do you think?

cloud-fan · 2016-06-28T12:39:21Z

From the plan tree given by you, the Cast is casting value#63 to array type, not ObjectType(classOf[Array[Double]]), there should be some other reason that we need to find out(maybe it's because the nullability of array element is different).

kiszk · 2016-06-28T14:01:35Z

Regarding the plan tree printout (I removed my debug information), the Cast performs conversion from UnsafeArrayData to GenericArrayData since the target type of Cast is ArrayType<DoubleType>. Since Cast shows dataType. simpleType of the target type, it is shown as array<double>.

Regarding the generated code, we seems to be on the same page. What you said is not done in Cast now, and what I did. IIUC, current Castcode generation for array seems to be conservative. Current goal of cast code generation is to create GenericArrayData object. This code generation always creates Object[] and assign values into each Object[] element. Then, the generated code passes Object[] to the constructor of GenericArrayData. As you pointed out, if code generation takes care of nullability, it can avoid to create Object[]. Unfortunately, it is not done in [the current code](code generation for array (Lines 828-862).
The code generated by this PR can also use specialized GenericArrayData implemented by #13758.

cloud-fan · 2016-06-29T02:05:07Z

the Cast performs conversion from UnsafeArrayData to GenericArrayData.

Cast is used to cast one type to another, not cast one kind of data representation to another, so I don't quite understand why the Cast is there.

This code generation always creates Object[] and assign values into each Object[] element.

This is reasonable, as it needs to take care of null elements. And we do have a chance to optimize it: if the target array type's element type is primitive and the input array type's element nullability is false, we can avoid using Object[].

kiszk · 2016-06-29T03:50:47Z

Let me check which code portion inserts Cast in this tree.

And we do have a chance to optimize it: if the target array type's element type is primitive and the input array type's element nullability is false, we can avoid using Object[].

I agree. This PR generates optimized code without using Object[] by check the above conditions.

cloud-fan · 2016-06-29T05:00:31Z

we need to make sure the input array's element nullability is false, but primitive type array doesn't guarantee it. e.g. we can have ArrayType(ByteType, true)

Yes, you are right. The latest code also checks ArrayType.containsNull of from and to.

kiszk · 2016-07-01T18:28:49Z

@cloud-fan, I checked the following

Let me check which code portion inserts Cast in this tree.

IIUC, this code inserts the corresponding Cast in Analyzed Logical Plan to upcast in Parsed Logical Plan.

SparkQA · 2016-07-01T19:12:18Z

Test build #61629 has finished for PR 13704 at commit 2d3d34a.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-07-01T19:18:01Z

Test build #61630 has finished for PR 13704 at commit 256c861.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-07-03T07:26:12Z

Test build #61677 has finished for PR 13704 at commit 77859cf.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

kiszk · 2016-07-03T07:51:37Z

@cloud-fan regarding a cast, I think that you are right. The issue was that the original cast was ArrayType(Double, false) to ArrayType(Double, true).
I changed to pass information on ArrayType.containsNull of destination of Cast now if the source of ArrayType.containsNull is false . In the example in this PR description, the condition is satisfied. Now, the cast is ArrayType(Double, false) to ArrayType(Double, false). As a result, after the logical optimizations, we got a plan tree without Cast. I updated the description of PR.

== Analyzed Logical Plan ==
value: double
SerializeFromObject [input[0, double, true] AS value#6]
+- MapElements <function1>, obj#5: double
+- DeserializeToObject cast(value#1 as array<double>).toDoubleArray, obj#4: [D
+- LocalRelation [value#1]
== Optimized Logical Plan ==
SerializeFromObject [input[0, double, true] AS value#6]
+- MapElements <function1>, obj#5: double
+- DeserializeToObject value#1.toDoubleArray, obj#4: [D
+- LocalRelation [value#1]

SparkQA · 2016-07-10T09:48:05Z

Test build #62050 has finished for PR 13704 at commit 43ced15.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds the following public classes (experimental):
- class PrimitiveArrayBenchmark extends BenchmarkBase

SparkQA · 2016-07-10T16:00:55Z

Test build #62057 has finished for PR 13704 at commit 677d81e.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

kiszk · 2016-07-10T16:09:55Z

@cloud-fan could you please review this? As you pointed, I also changed code related to cast. I added benchmark results, too.

cloud-fan · 2016-07-11T01:22:39Z

how about we move this into another PR? I think the main purpose of this PR is to eliminate the unnecessary Cast

Sure, I will create another PR and update the description.

SparkQA · 2016-07-11T05:08:25Z

Test build #62071 has finished for PR 13704 at commit 66800fa.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

cloud-fan · 2016-07-11T10:03:42Z

how about
case c @ Cast(e, dataType) => (e.dataType, dataType) match { case (ArrayType(from, false), ArrayType(to, true)) if from == to => e case (MapType(fromKey, fromValue, false), ArrayType(toKey, toValue, true)) if fromKey == toKey && fromValue == toValue => e case _ => c }

thanks, I like this simple one

SparkQA · 2016-07-11T12:41:43Z

Test build #62084 has finished for PR 13704 at commit c31729f.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-07-11T12:48:23Z

Test build #62087 has finished for PR 13704 at commit 1bbe859.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-07-11T16:38:56Z

Test build #62097 has finished for PR 13704 at commit b7477de.

This patch fails Scala style tests.
This patch merges cleanly.
This patch adds the following public classes (experimental):
- class SimplifyCastsSuite extends PlanTest

SparkQA · 2016-08-29T21:24:37Z

Test build #64594 has finished for PR 13704 at commit c8f87a1.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

kiszk · 2016-08-30T01:47:45Z

@liancheng Could you please review this since I resolved conflict?

cloud-fan · 2016-08-30T02:14:17Z

+ comparePlans(optimized, expected)
+ }
+
+ test("non-nullable to nullable array cast") {


non-nullable element array to nullable element array cast

cloud-fan · 2016-08-30T02:16:24Z

left some comment, let's go ahead and merge it after that :)

SparkQA · 2016-08-30T15:27:02Z

Test build #64658 has finished for PR 13704 at commit 40ac2bc.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

cloud-fan · 2016-08-31T04:41:20Z

thanks, merging to master!

kiszk mentioned this pull request Jun 25, 2016
[SPARK-16043][SQL] Prepare GenericArrayData implementation specialized for a primitive array #13758
Closed

cloud-fan reviewed Jun 27, 2016
View reviewed changes

cloud-fan reviewed Jun 29, 2016
View reviewed changes

kiszk force-pushed the SPARK-15985 branch from 17f17d6 to 2d3d34aCompare July 1, 2016 18:32

cloud-fan reviewed Jul 11, 2016
View reviewed changes

kiszk added 18 commits August 30, 2016 04:09

remove debug print
6026a21

fix test failures by precisely passing nullable information to cast
f3df6c5

add benchmark program
f4c8448

fix test failures
f8fe24e

revert changes for improvement of cast code generation
7622502

fix typo
90b6542

eliminate unnecessary cast with non-nullable at SimplifyCasts
7615342

simplify case statements
4a198b2

add test cases
6c7904e

fix scala style error
3fbe5e9

drop benchmark program
ee29c1b

add more test cases
abdb529

update test cases
52e8bfd

update test cases
ab0bed1

update test cases
0207204

update test cases
d936cf0

update test cases
bff189a

update test cases
c8f87a1

kiszk force-pushed the SPARK-15985 branch from e4cd571 to c8f87a1Compare August 29, 2016 19:19

cloud-fan reviewed Aug 30, 2016
View reviewed changes

addressed comment
40ac2bc

asfgit closed this in d92cd22Aug 31, 2016

Uh oh!

Conversation

kiszk commented Jun 16, 2016• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What changes were proposed in this pull request?

How was this patch tested?

Uh oh!

SparkQA commented Jun 16, 2016

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

cloud-fan commented Jun 27, 2016• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

kiszk commented Jun 27, 2016

Uh oh!

cloud-fan commented Jun 27, 2016

Uh oh!

kiszk commented Jun 27, 2016• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cloud-fan commented Jun 28, 2016

Uh oh!

kiszk commented Jun 28, 2016• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cloud-fan commented Jun 28, 2016

Uh oh!

kiszk commented Jun 28, 2016

Uh oh!

cloud-fan commented Jun 29, 2016

Uh oh!

kiszk commented Jun 29, 2016• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kiszk commented Jul 1, 2016

Uh oh!

SparkQA commented Jul 1, 2016

Uh oh!

SparkQA commented Jul 1, 2016

Uh oh!

SparkQA commented Jul 3, 2016

Uh oh!

kiszk commented Jul 3, 2016• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

SparkQA commented Jul 10, 2016

Uh oh!

SparkQA commented Jul 10, 2016

Uh oh!

kiszk commented Jul 10, 2016

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SparkQA commented Jul 11, 2016

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SparkQA commented Jul 11, 2016

Uh oh!

SparkQA commented Jul 11, 2016

Uh oh!

SparkQA commented Jul 11, 2016

Uh oh!

SparkQA commented Aug 29, 2016

Uh oh!

kiszk commented Aug 30, 2016

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kiszk commented Jun 16, 2016•
edited
Loading

cloud-fan commented Jun 27, 2016•
edited
Loading

kiszk commented Jun 27, 2016•
edited
Loading

kiszk commented Jun 28, 2016•
edited
Loading

kiszk commented Jun 29, 2016•
edited
Loading

kiszk commented Jul 3, 2016•
edited
Loading