Skip to content

Initial Spark Dataset API spec #18

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Closed
wants to merge 2 commits into from
Closed

Initial Spark Dataset API spec #18

wants to merge 2 commits into from

Conversation

marmbrus
Copy link
Owner

No description provided.

marmbrus
marmbrus reviewed Oct 6, 2015
View reviewed changes
sql/core/src/main/scala/org/apache/spark/sql/Dataset.scala
* aggregation *
****************/

def reduce(f: (T, T) => T): T = ???
Copy link
Owner Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

reduce option?

@marmbrus
Copy link
Owner Author

@johnynek, thanks for the comments! You should look at the version in master as we've done a lot of implementation since this was first published.

@AmplabJenkins
Copy link

Merged build finished. Test FAILed.

@AmplabJenkins
Copy link

Test FAILed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/spark-streaming-df-test/8/
Test FAILed.

@marmbrus marmbrus closed this Mar 8, 2016
marmbrus pushed a commit that referenced this pull request May 3, 2017
@kiszk @cloud-fan
[SPARK-19008][SQL] Improve performance of Dataset.map by eliminating …
5949e6c 
…boxing/unboxing

## What changes were proposed in this pull request?

This PR improve performance of Dataset.map() for primitive types by removing boxing/unbox operations. This is based on [the discussion](apache#16391 (comment)) with cloud-fan.

Current Catalyst generates a method call to a `apply()` method of an anonymous function written in Scala. The types of an argument and return value are `java.lang.Object`. As a result, each method call for a primitive value involves a pair of unboxing and boxing for calling this `apply()` method and a pair of boxing and unboxing for returning from this `apply()` method.

This PR directly calls a specialized version of a `apply()` method without boxing and unboxing. For example, if types of an arguments ant return value is `int`, this PR generates a method call to `apply$mcII$sp`. This PR supports any combination of `Int`, `Long`, `Float`, and `Double`.

The following is a benchmark result using [this program](https://github.com/apache/spark/pull/16391/files) with 4.7x. Here is a Dataset part of this program.

Without this PR
```
OpenJDK 64-Bit Server VM 1.8.0_111-8u111-b14-2ubuntu0.16.04.2-b14 on Linux 4.4.0-47-generic
Intel(R) Xeon(R) CPU E5-2667 v3  3.20GHz
back-to-back map:                        Best/Avg Time(ms)    Rate(M/s)   Per Row(ns)   Relative
------------------------------------------------------------------------------------------------
RDD                                           1923 / 1952         52.0          19.2       1.0X
DataFrame                                      526 /  548        190.2           5.3       3.7X
Dataset                                       3094 / 3154         32.3          30.9       0.6X
```

With this PR
```
OpenJDK 64-Bit Server VM 1.8.0_111-8u111-b14-2ubuntu0.16.04.2-b14 on Linux 4.4.0-47-generic
Intel(R) Xeon(R) CPU E5-2667 v3  3.20GHz
back-to-back map:                        Best/Avg Time(ms)    Rate(M/s)   Per Row(ns)   Relative
------------------------------------------------------------------------------------------------
RDD                                           1883 / 1892         53.1          18.8       1.0X
DataFrame                                      502 /  642        199.1           5.0       3.7X
Dataset                                        657 /  784        152.2           6.6       2.9X
```

```java
  def backToBackMap(spark: SparkSession, numRows: Long, numChains: Int): Benchmark = {
    import spark.implicits._
    val rdd = spark.sparkContext.range(0, numRows)
    val ds = spark.range(0, numRows)
    val func = (l: Long) => l + 1
    val benchmark = new Benchmark("back-to-back map", numRows)
...
    benchmark.addCase("Dataset") { iter =>
      var res = ds.as[Long]
      var i = 0
      while (i < numChains) {
        res = res.map(func)
        i += 1
      }
      res.queryExecution.toRdd.foreach(_ => Unit)
    }
    benchmark
  }
```

A motivating example
```java
Seq(1, 2, 3).toDS.map(i => i * 7).show
```

Generated code without this PR
```java
/* 005 */ final class GeneratedIterator extends org.apache.spark.sql.execution.BufferedRowIterator {
/* 006 */   private Object[] references;
/* 007 */   private scala.collection.Iterator[] inputs;
/* 008 */   private scala.collection.Iterator inputadapter_input;
/* 009 */   private UnsafeRow deserializetoobject_result;
/* 010 */   private org.apache.spark.sql.catalyst.expressions.codegen.BufferHolder deserializetoobject_holder;
/* 011 */   private org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter deserializetoobject_rowWriter;
/* 012 */   private int mapelements_argValue;
/* 013 */   private UnsafeRow mapelements_result;
/* 014 */   private org.apache.spark.sql.catalyst.expressions.codegen.BufferHolder mapelements_holder;
/* 015 */   private org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter mapelements_rowWriter;
/* 016 */   private UnsafeRow serializefromobject_result;
/* 017 */   private org.apache.spark.sql.catalyst.expressions.codegen.BufferHolder serializefromobject_holder;
/* 018 */   private org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter serializefromobject_rowWriter;
/* 019 */
/* 020 */   public GeneratedIterator(Object[] references) {
/* 021 */     this.references = references;
/* 022 */   }
/* 023 */
/* 024 */   public void init(int index, scala.collection.Iterator[] inputs) {
/* 025 */     partitionIndex = index;
/* 026 */     this.inputs = inputs;
/* 027 */     inputadapter_input = inputs[0];
/* 028 */     deserializetoobject_result = new UnsafeRow(1);
/* 029 */     this.deserializetoobject_holder = new org.apache.spark.sql.catalyst.expressions.codegen.BufferHolder(deserializetoobject_result, 0);
/* 030 */     this.deserializetoobject_rowWriter = new org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(deserializetoobject_holder, 1);
/* 031 */
/* 032 */     mapelements_result = new UnsafeRow(1);
/* 033 */     this.mapelements_holder = new org.apache.spark.sql.catalyst.expressions.codegen.BufferHolder(mapelements_result, 0);
/* 034 */     this.mapelements_rowWriter = new org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(mapelements_holder, 1);
/* 035 */     serializefromobject_result = new UnsafeRow(1);
/* 036 */     this.serializefromobject_holder = new org.apache.spark.sql.catalyst.expressions.codegen.BufferHolder(serializefromobject_result, 0);
/* 037 */     this.serializefromobject_rowWriter = new org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(serializefromobject_holder, 1);
/* 038 */
/* 039 */   }
/* 040 */
/* 041 */   protected void processNext() throws java.io.IOException {
/* 042 */     while (inputadapter_input.hasNext() && !stopEarly()) {
/* 043 */       InternalRow inputadapter_row = (InternalRow) inputadapter_input.next();
/* 044 */       int inputadapter_value = inputadapter_row.getInt(0);
/* 045 */
/* 046 */       boolean mapelements_isNull = true;
/* 047 */       int mapelements_value = -1;
/* 048 */       if (!false) {
/* 049 */         mapelements_argValue = inputadapter_value;
/* 050 */
/* 051 */         mapelements_isNull = false;
/* 052 */         if (!mapelements_isNull) {
/* 053 */           Object mapelements_funcResult = null;
/* 054 */           mapelements_funcResult = ((scala.Function1) references[0]).apply(mapelements_argValue);
/* 055 */           if (mapelements_funcResult == null) {
/* 056 */             mapelements_isNull = true;
/* 057 */           } else {
/* 058 */             mapelements_value = (Integer) mapelements_funcResult;
/* 059 */           }
/* 060 */
/* 061 */         }
/* 062 */
/* 063 */       }
/* 064 */
/* 065 */       serializefromobject_rowWriter.zeroOutNullBytes();
/* 066 */
/* 067 */       if (mapelements_isNull) {
/* 068 */         serializefromobject_rowWriter.setNullAt(0);
/* 069 */       } else {
/* 070 */         serializefromobject_rowWriter.write(0, mapelements_value);
/* 071 */       }
/* 072 */       append(serializefromobject_result);
/* 073 */       if (shouldStop()) return;
/* 074 */     }
/* 075 */   }
/* 076 */ }
```

Generated code with this PR (lines 48-56 are changed)
```java
/* 005 */ final class GeneratedIterator extends org.apache.spark.sql.execution.BufferedRowIterator {
/* 006 */   private Object[] references;
/* 007 */   private scala.collection.Iterator[] inputs;
/* 008 */   private scala.collection.Iterator inputadapter_input;
/* 009 */   private UnsafeRow deserializetoobject_result;
/* 010 */   private org.apache.spark.sql.catalyst.expressions.codegen.BufferHolder deserializetoobject_holder;
/* 011 */   private org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter deserializetoobject_rowWriter;
/* 012 */   private int mapelements_argValue;
/* 013 */   private UnsafeRow mapelements_result;
/* 014 */   private org.apache.spark.sql.catalyst.expressions.codegen.BufferHolder mapelements_holder;
/* 015 */   private org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter mapelements_rowWriter;
/* 016 */   private UnsafeRow serializefromobject_result;
/* 017 */   private org.apache.spark.sql.catalyst.expressions.codegen.BufferHolder serializefromobject_holder;
/* 018 */   private org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter serializefromobject_rowWriter;
/* 019 */
/* 020 */   public GeneratedIterator(Object[] references) {
/* 021 */     this.references = references;
/* 022 */   }
/* 023 */
/* 024 */   public void init(int index, scala.collection.Iterator[] inputs) {
/* 025 */     partitionIndex = index;
/* 026 */     this.inputs = inputs;
/* 027 */     inputadapter_input = inputs[0];
/* 028 */     deserializetoobject_result = new UnsafeRow(1);
/* 029 */     this.deserializetoobject_holder = new org.apache.spark.sql.catalyst.expressions.codegen.BufferHolder(deserializetoobject_result, 0);
/* 030 */     this.deserializetoobject_rowWriter = new org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(deserializetoobject_holder, 1);
/* 031 */
/* 032 */     mapelements_result = new UnsafeRow(1);
/* 033 */     this.mapelements_holder = new org.apache.spark.sql.catalyst.expressions.codegen.BufferHolder(mapelements_result, 0);
/* 034 */     this.mapelements_rowWriter = new org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(mapelements_holder, 1);
/* 035 */     serializefromobject_result = new UnsafeRow(1);
/* 036 */     this.serializefromobject_holder = new org.apache.spark.sql.catalyst.expressions.codegen.BufferHolder(serializefromobject_result, 0);
/* 037 */     this.serializefromobject_rowWriter = new org.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(serializefromobject_holder, 1);
/* 038 */
/* 039 */   }
/* 040 */
/* 041 */   protected void processNext() throws java.io.IOException {
/* 042 */     while (inputadapter_input.hasNext() && !stopEarly()) {
/* 043 */       InternalRow inputadapter_row = (InternalRow) inputadapter_input.next();
/* 044 */       int inputadapter_value = inputadapter_row.getInt(0);
/* 045 */
/* 046 */       boolean mapelements_isNull = true;
/* 047 */       int mapelements_value = -1;
/* 048 */       if (!false) {
/* 049 */         mapelements_argValue = inputadapter_value;
/* 050 */
/* 051 */         mapelements_isNull = false;
/* 052 */         if (!mapelements_isNull) {
/* 053 */           mapelements_value = ((scala.Function1) references[0]).apply$mcII$sp(mapelements_argValue);
/* 054 */         }
/* 055 */
/* 056 */       }
/* 057 */
/* 058 */       serializefromobject_rowWriter.zeroOutNullBytes();
/* 059 */
/* 060 */       if (mapelements_isNull) {
/* 061 */         serializefromobject_rowWriter.setNullAt(0);
/* 062 */       } else {
/* 063 */         serializefromobject_rowWriter.write(0, mapelements_value);
/* 064 */       }
/* 065 */       append(serializefromobject_result);
/* 066 */       if (shouldStop()) return;
/* 067 */     }
/* 068 */   }
/* 069 */ }
```

Java bytecode for methods for `i => i * 7`
```java
$ javap -c Test\$\$anonfun\$5\$\$anonfun\$apply\$mcV\$sp\$1.class
Compiled from "Test.scala"
public final class org.apache.spark.sql.Test$$anonfun$5$$anonfun$apply$mcV$sp$1 extends scala.runtime.AbstractFunction1$mcII$sp implements scala.Serializable {
  public static final long serialVersionUID;

  public final int apply(int);
    Code:
       0: aload_0
       1: iload_1
       2: invokevirtual #18                 // Method apply$mcII$sp:(I)I
       5: ireturn

  public int apply$mcII$sp(int);
    Code:
       0: iload_1
       1: bipush        7
       3: imul
       4: ireturn

  public final java.lang.Object apply(java.lang.Object);
    Code:
       0: aload_0
       1: aload_1
       2: invokestatic  #29                 // Method scala/runtime/BoxesRunTime.unboxToInt:(Ljava/lang/Object;)I
       5: invokevirtual #31                 // Method apply:(I)I
       8: invokestatic  apache#35                 // Method scala/runtime/BoxesRunTime.boxToInteger:(I)Ljava/lang/Integer;
      11: areturn

  public org.apache.spark.sql.Test$$anonfun$5$$anonfun$apply$mcV$sp$1(org.apache.spark.sql.Test$$anonfun$5);
    Code:
       0: aload_0
       1: invokespecial apache#42                 // Method scala/runtime/AbstractFunction1$mcII$sp."<init>":()V
       4: return
}
```
## How was this patch tested?

Added new test suites to `DatasetPrimitiveSuite`.

Author: Kazuaki Ishizaki <ishizaki@jp.ibm.com>

Closes apache#17172 from kiszk/SPARK-19008.
marmbrus pushed a commit that referenced this pull request May 3, 2017
@zsxwing
[SPARK-20131][CORE] Don't use this lock in StandaloneSchedulerBacke…
c5f1cc3 
…nd.stop

## What changes were proposed in this pull request?

`o.a.s.streaming.StreamingContextSuite.SPARK-18560 Receiver data should be deserialized properly` is flaky is because there is a potential dead-lock in StandaloneSchedulerBackend which causes `await` timeout. Here is the related stack trace:
```
"Thread-31" apache#211 daemon prio=5 os_prio=31 tid=0x00007fedd4808000 nid=0x16403 waiting on condition [0x00007000239b7000]
   java.lang.Thread.State: TIMED_WAITING (parking)
	at sun.misc.Unsafe.park(Native Method)
	- parking to wait for  <0x000000079b49ca10> (a scala.concurrent.impl.Promise$CompletionLatch)
	at java.util.concurrent.locks.LockSupport.parkNanos(LockSupport.java:215)
	at java.util.concurrent.locks.AbstractQueuedSynchronizer.doAcquireSharedNanos(AbstractQueuedSynchronizer.java:1037)
	at java.util.concurrent.locks.AbstractQueuedSynchronizer.tryAcquireSharedNanos(AbstractQueuedSynchronizer.java:1328)
	at scala.concurrent.impl.Promise$DefaultPromise.tryAwait(Promise.scala:208)
	at scala.concurrent.impl.Promise$DefaultPromise.ready(Promise.scala:218)
	at scala.concurrent.impl.Promise$DefaultPromise.result(Promise.scala:223)
	at org.apache.spark.util.ThreadUtils$.awaitResult(ThreadUtils.scala:201)
	at org.apache.spark.rpc.RpcTimeout.awaitResult(RpcTimeout.scala:75)
	at org.apache.spark.rpc.RpcEndpointRef.askSync(RpcEndpointRef.scala:92)
	at org.apache.spark.rpc.RpcEndpointRef.askSync(RpcEndpointRef.scala:76)
	at org.apache.spark.scheduler.cluster.CoarseGrainedSchedulerBackend.stop(CoarseGrainedSchedulerBackend.scala:402)
	at org.apache.spark.scheduler.cluster.StandaloneSchedulerBackend.org$apache$spark$scheduler$cluster$StandaloneSchedulerBackend$$stop(StandaloneSchedulerBackend.scala:213)
	- locked <0x00000007066fca38> (a org.apache.spark.scheduler.cluster.StandaloneSchedulerBackend)
	at org.apache.spark.scheduler.cluster.StandaloneSchedulerBackend.stop(StandaloneSchedulerBackend.scala:116)
	- locked <0x00000007066fca38> (a org.apache.spark.scheduler.cluster.StandaloneSchedulerBackend)
	at org.apache.spark.scheduler.TaskSchedulerImpl.stop(TaskSchedulerImpl.scala:517)
	at org.apache.spark.scheduler.DAGScheduler.stop(DAGScheduler.scala:1657)
	at org.apache.spark.SparkContext$$anonfun$stop$8.apply$mcV$sp(SparkContext.scala:1921)
	at org.apache.spark.util.Utils$.tryLogNonFatalError(Utils.scala:1302)
	at org.apache.spark.SparkContext.stop(SparkContext.scala:1920)
	at org.apache.spark.streaming.StreamingContext.stop(StreamingContext.scala:708)
	at org.apache.spark.streaming.StreamingContextSuite$$anonfun$43$$anonfun$apply$mcV$sp$66$$anon$3.run(StreamingContextSuite.scala:827)

"dispatcher-event-loop-3" #18 daemon prio=5 os_prio=31 tid=0x00007fedd603a000 nid=0x6203 waiting for monitor entry [0x0000700003be4000]
   java.lang.Thread.State: BLOCKED (on object monitor)
	at org.apache.spark.scheduler.cluster.CoarseGrainedSchedulerBackend$DriverEndpoint.org$apache$spark$scheduler$cluster$CoarseGrainedSchedulerBackend$DriverEndpoint$$makeOffers(CoarseGrainedSchedulerBackend.scala:253)
	- waiting to lock <0x00000007066fca38> (a org.apache.spark.scheduler.cluster.StandaloneSchedulerBackend)
	at org.apache.spark.scheduler.cluster.CoarseGrainedSchedulerBackend$DriverEndpoint$$anonfun$receive$1.applyOrElse(CoarseGrainedSchedulerBackend.scala:124)
	at org.apache.spark.rpc.netty.Inbox$$anonfun$process$1.apply$mcV$sp(Inbox.scala:117)
	at org.apache.spark.rpc.netty.Inbox.safelyCall(Inbox.scala:205)
	at org.apache.spark.rpc.netty.Inbox.process(Inbox.scala:101)
	at org.apache.spark.rpc.netty.Dispatcher$MessageLoop.run(Dispatcher.scala:213)
	at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142)
	at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617)
	at java.lang.Thread.run(Thread.java:745)
```

This PR removes `synchronized` and changes `stopping` to AtomicBoolean to ensure idempotent to fix the dead-lock.

## How was this patch tested?

Jenkins

Author: Shixiong Zhu <shixiong@databricks.com>

Closes apache#17610 from zsxwing/SPARK-20131.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
3 participants
@marmbrus @AmplabJenkins @johnynek