Akka-Stream implementation slower than single threaded implementation

Question

UPDATE FROM 2015-10-30

---

based on Roland Kuhn Awnser:

Akka Streams is using asynchronous message passing between Actors to implement stream processing stages. Passing data across an asynchronous boundary has an overhead that you are seeing here: your computation seems to take only about 160ns (derived from the single-threaded measurement) while the streaming solution takes roughly 1µs per element, which is dominated by the message passing.

Another misconception is that saying “stream” implies parallelism: in your code all computation runs sequentially in a single Actor (the map stage), so no benefit can be expected over the primitive single-threaded solution.

In order to benefit from the parallelism afforded by Akka Streams you need to have multiple processing stages that each perform tasks of

1µs per element, see also the docs.

I did some changes. My code now looks like:

object MultiThread {
  implicit val actorSystem = ActorSystem("Sys")
  implicit val materializer = ActorMaterializer()

  var counter = 0
  var oldProgess = 0

  //RunnableFlow: in -> flow -> sink
  val in = Source(() => Iterator.continually((1254785478l, "name", 48, 23.09f)))

  val flow = Flow[(Long, String, Int, Float)].map(p => SharedFunctions.transform2(SharedFunctions.transform(p)))

  val tupleToEvent = Flow[(Long, String, Int, Float)].map(SharedFunctions.transform)

  val eventToFactorial = Flow[Event].map(SharedFunctions.transform2)

  val eventChef: Flow[(Long, String, Int, Float), Int, Unit] = Flow() { implicit builder =>
    import FlowGraph.Implicits._

    val dispatchTuple = builder.add(Balance[(Long, String, Int, Float)](4))
    val mergeEvents = builder.add(Merge[Int](4))

    dispatchTuple.out(0) ~> tupleToEvent ~> eventToFactorial ~> mergeEvents.in(0)
    dispatchTuple.out(1) ~> tupleToEvent ~> eventToFactorial ~> mergeEvents.in(1)
    dispatchTuple.out(2) ~> tupleToEvent ~> eventToFactorial ~> mergeEvents.in(2)
    dispatchTuple.out(3) ~> tupleToEvent ~> eventToFactorial ~> mergeEvents.in(3)

    (dispatchTuple.in, mergeEvents.out)
  }

  val sink = Sink.foreach[Int]{
    v => counter += 1
    oldProgess = SharedFunctions.printProgress(oldProgess, SharedFunctions.maxEventCount, counter,
    DateTime.now.getMillis - SharedFunctions.startTime.getMillis)
    if(counter == SharedFunctions.maxEventCount) endAkka()
  }

  def endAkka() = {
    val duration = new Duration(SharedFunctions.startTime, DateTime.now)
    println("Time: " + duration.getMillis + " || Data: " + counter)
    actorSystem.shutdown
    actorSystem.awaitTermination
    System.exit(-1)
  }

  def main(args: Array[String]) {
    println("MultiThread started: " + SharedFunctions.startTime)
    in.via(flow).runWith(sink)
   // in.via(eventChef).runWith(sink)
  }

}

I not sure if I get something totally wrong, but still my implementation with akka-streams is much slower (now even slower as before) but what I found out is: If I increase the work for example by doing some division the implementation with akka-streams gets faster. So If I get it right (correct me otherwise) it seems there is too much overhead in my example. So you only get a benefit from akka-streams if the code has to do heavy work?

---

---

---

I'm relatively new in both scala & akka-stream. I wrote a little test project which creates some events until a counter has reached a specific number. For each event the factorial for one field of the event is being computed. I implemented this twice. One time with akka-stream and one time without akka-stream (single threaded) and compared the runtime.

I didn't expect that: When I create a single event the runtime of both programs are nearly the same. But if I create 70,000,000 events the implementation without akka-streams is much faster. Here are my results (the following data is based on 24 measurements):

---

• Single event without akka-streams: 403 (+- 2)ms

• Single event with akka-streams: 444 (+-13)ms

  ---

• 70Mio events without akka-streams: 11778 (+-70)ms

• 70Mio events with akka-steams: 75424(+-2959)ms

---

So my Question is: What is going on? Why is my implementation with akka-stream slower?

here my code:

Implementation with Akka

object MultiThread {
  implicit val actorSystem = ActorSystem("Sys")
  implicit val materializer = ActorMaterializer()

  var counter = 0
  var oldProgess = 0

  //RunnableFlow: in -> flow -> sink
  val in = Source(() => Iterator.continually((1254785478l, "name", 48, 23.09f)))

  val flow = Flow[(Long, String, Int, Float)].map(p => SharedFunctions.transform2(SharedFunctions.transform(p)))

  val sink = Sink.foreach[Int]{
    v => counter += 1
    oldProgess = SharedFunctions.printProgress(oldProgess, SharedFunctions.maxEventCount, counter,
    DateTime.now.getMillis - SharedFunctions.startTime.getMillis)
    if(counter == SharedFunctions.maxEventCount) endAkka()
  }

  def endAkka() = {
    val duration = new Duration(SharedFunctions.startTime, DateTime.now)
    println("Time: " + duration.getMillis + " || Data: " + counter)
    actorSystem.shutdown
    actorSystem.awaitTermination
    System.exit(-1)
  }

  def main(args: Array[String]) {
    import scala.concurrent.ExecutionContext.Implicits.global
    println("MultiThread started: " + SharedFunctions.startTime)
    in.via(flow).runWith(sink).onComplete(_ => endAkka())
  }

}

Implementation without Akka

object SingleThread {

  def main(args: Array[String]) {
    println("SingleThread started at: " + SharedFunctions.startTime)
    println("0%")
    val i = createEvent(0)
    val duration = new Duration(SharedFunctions.startTime, DateTime.now());
    println("Time: " + duration.getMillis + " || Data: " + i)
  }

  def createEventWorker(oldProgress: Int, count: Int, randDate: Long, name: String, age: Int, myFloat: Float): Int = {
    if (count == SharedFunctions.maxEventCount) count
    else {
      val e = SharedFunctions.transform((randDate, name, age, myFloat))
      SharedFunctions.transform2(e)
      val p = SharedFunctions.printProgress(oldProgress, SharedFunctions.maxEventCount, count,
        DateTime.now.getMillis - SharedFunctions.startTime.getMillis)
      createEventWorker(p, count + 1, 1254785478l, "name", 48, 23.09f)
    }
  }

  def createEvent(count: Int): Int = {
    createEventWorker(0, count, 1254785478l, "name", 48, 23.09f)
  }
}

SharedFunctions

object SharedFunctions {
  val maxEventCount = 70000000
  val startTime = DateTime.now

  def transform(t : (Long, String, Int, Float)) : Event = new Event(t._1 ,t._2,t._3,t._4)
  def transform2(e : Event) : Int = factorial(e.getAgeYrs)

  def calculatePercentage(totalValue: Long, currentValue: Long) = Math.round((currentValue * 100) / totalValue)
  def printProgress(oldProgress : Int, fileSize: Long, currentSize: Int, t: Long) = {
    val cProgress = calculatePercentage(fileSize, currentSize)
    if (oldProgress != cProgress) println(s"$oldProgress% | $t ms")
    cProgress
  }

  private def factorialWorker(n1: Int, n2: Int): Int = {
    if (n1 == 0) n2
    else factorialWorker(n1 -1, n2*n1)
  }
  def factorial (n : Int): Int = {
    factorialWorker(n, 1)
  }
}

Implementation Event

/**
 * Autogenerated by Avro
 * 
 * DO NOT EDIT DIRECTLY
 */

@SuppressWarnings("all")
@org.apache.avro.specific.AvroGenerated
public class Event extends org.apache.avro.specific.SpecificRecordBase implements org.apache.avro.specific.SpecificRecord {
  public static final org.apache.avro.Schema SCHEMA$ = new org.apache.avro.Schema.Parser().parse("{\"type\":\"record\",\"name\":\"Event\",\"namespace\":\"week2P2\",\"fields\":[{\"name\":\"timestampMS\",\"type\":\"long\"},{\"name\":\"name\",\"type\":\"string\"},{\"name\":\"ageYrs\",\"type\":\"int\"},{\"name\":\"sizeCm\",\"type\":\"float\"}]}");
  public static org.apache.avro.Schema getClassSchema() { return SCHEMA$; }
  @Deprecated public long timestampMS;
  @Deprecated public CharSequence name;
  @Deprecated public int ageYrs;
  @Deprecated public float sizeCm;

  /**
   * Default constructor.  Note that this does not initialize fields
   * to their default values from the schema.  If that is desired then
   * one should use <code>newBuilder()</code>. 
   */
  public Event() {}

  /**
   * All-args constructor.
   */
  public Event(Long timestampMS, CharSequence name, Integer ageYrs, Float sizeCm) {
    this.timestampMS = timestampMS;
    this.name = name;
    this.ageYrs = ageYrs;
    this.sizeCm = sizeCm;
  }

  public org.apache.avro.Schema getSchema() { return SCHEMA$; }
  // Used by DatumWriter.  Applications should not call. 
  public Object get(int field$) {
    switch (field$) {
    case 0: return timestampMS;
    case 1: return name;
    case 2: return ageYrs;
    case 3: return sizeCm;
    default: throw new org.apache.avro.AvroRuntimeException("Bad index");
    }
  }
  // Used by DatumReader.  Applications should not call. 
  @SuppressWarnings(value="unchecked")
  public void put(int field$, Object value$) {
    switch (field$) {
    case 0: timestampMS = (Long)value$; break;
    case 1: name = (CharSequence)value$; break;
    case 2: ageYrs = (Integer)value$; break;
    case 3: sizeCm = (Float)value$; break;
    default: throw new org.apache.avro.AvroRuntimeException("Bad index");
    }
  }

  /**
   * Gets the value of the 'timestampMS' field.
   */
  public Long getTimestampMS() {
    return timestampMS;
  }

  /**
   * Sets the value of the 'timestampMS' field.
   * @param value the value to set.
   */
  public void setTimestampMS(Long value) {
    this.timestampMS = value;
  }

  /**
   * Gets the value of the 'name' field.
   */
  public CharSequence getName() {
    return name;
  }

  /**
   * Sets the value of the 'name' field.
   * @param value the value to set.
   */
  public void setName(CharSequence value) {
    this.name = value;
  }

  /**
   * Gets the value of the 'ageYrs' field.
   */
  public Integer getAgeYrs() {
    return ageYrs;
  }

  /**
   * Sets the value of the 'ageYrs' field.
   * @param value the value to set.
   */
  public void setAgeYrs(Integer value) {
    this.ageYrs = value;
  }

  /**
   * Gets the value of the 'sizeCm' field.
   */
  public Float getSizeCm() {
    return sizeCm;
  }

  /**
   * Sets the value of the 'sizeCm' field.
   * @param value the value to set.
   */
  public void setSizeCm(Float value) {
    this.sizeCm = value;
  }

  /** Creates a new Event RecordBuilder */
  public static Event.Builder newBuilder() {
    return new Event.Builder();
  }

  /** Creates a new Event RecordBuilder by copying an existing Builder */
  public static Event.Builder newBuilder(Event.Builder other) {
    return new Event.Builder(other);
  }

  /** Creates a new Event RecordBuilder by copying an existing Event instance */
  public static Event.Builder newBuilder(Event other) {
    return new Event.Builder(other);
  }

  /**
   * RecordBuilder for Event instances.
   */
  public static class Builder extends org.apache.avro.specific.SpecificRecordBuilderBase<Event>
    implements org.apache.avro.data.RecordBuilder<Event> {

    private long timestampMS;
    private CharSequence name;
    private int ageYrs;
    private float sizeCm;

    /** Creates a new Builder */
    private Builder() {
      super(Event.SCHEMA$);
    }

    /** Creates a Builder by copying an existing Builder */
    private Builder(Event.Builder other) {
      super(other);
      if (isValidValue(fields()[0], other.timestampMS)) {
        this.timestampMS = data().deepCopy(fields()[0].schema(), other.timestampMS);
        fieldSetFlags()[0] = true;
      }
      if (isValidValue(fields()[1], other.name)) {
        this.name = data().deepCopy(fields()[1].schema(), other.name);
        fieldSetFlags()[1] = true;
      }
      if (isValidValue(fields()[2], other.ageYrs)) {
        this.ageYrs = data().deepCopy(fields()[2].schema(), other.ageYrs);
        fieldSetFlags()[2] = true;
      }
      if (isValidValue(fields()[3], other.sizeCm)) {
        this.sizeCm = data().deepCopy(fields()[3].schema(), other.sizeCm);
        fieldSetFlags()[3] = true;
      }
    }

    /** Creates a Builder by copying an existing Event instance */
    private Builder(Event other) {
            super(Event.SCHEMA$);
      if (isValidValue(fields()[0], other.timestampMS)) {
        this.timestampMS = data().deepCopy(fields()[0].schema(), other.timestampMS);
        fieldSetFlags()[0] = true;
      }
      if (isValidValue(fields()[1], other.name)) {
        this.name = data().deepCopy(fields()[1].schema(), other.name);
        fieldSetFlags()[1] = true;
      }
      if (isValidValue(fields()[2], other.ageYrs)) {
        this.ageYrs = data().deepCopy(fields()[2].schema(), other.ageYrs);
        fieldSetFlags()[2] = true;
      }
      if (isValidValue(fields()[3], other.sizeCm)) {
        this.sizeCm = data().deepCopy(fields()[3].schema(), other.sizeCm);
        fieldSetFlags()[3] = true;
      }
    }

    /** Gets the value of the 'timestampMS' field */
    public Long getTimestampMS() {
      return timestampMS;
    }

    /** Sets the value of the 'timestampMS' field */
    public Event.Builder setTimestampMS(long value) {
      validate(fields()[0], value);
      this.timestampMS = value;
      fieldSetFlags()[0] = true;
      return this; 
    }

    /** Checks whether the 'timestampMS' field has been set */
    public boolean hasTimestampMS() {
      return fieldSetFlags()[0];
    }

    /** Clears the value of the 'timestampMS' field */
    public Event.Builder clearTimestampMS() {
      fieldSetFlags()[0] = false;
      return this;
    }

    /** Gets the value of the 'name' field */
    public CharSequence getName() {
      return name;
    }

    /** Sets the value of the 'name' field */
    public Event.Builder setName(CharSequence value) {
      validate(fields()[1], value);
      this.name = value;
      fieldSetFlags()[1] = true;
      return this; 
    }

    /** Checks whether the 'name' field has been set */
    public boolean hasName() {
      return fieldSetFlags()[1];
    }

    /** Clears the value of the 'name' field */
    public Event.Builder clearName() {
      name = null;
      fieldSetFlags()[1] = false;
      return this;
    }

    /** Gets the value of the 'ageYrs' field */
    public Integer getAgeYrs() {
      return ageYrs;
    }

    /** Sets the value of the 'ageYrs' field */
    public Event.Builder setAgeYrs(int value) {
      validate(fields()[2], value);
      this.ageYrs = value;
      fieldSetFlags()[2] = true;
      return this; 
    }

    /** Checks whether the 'ageYrs' field has been set */
    public boolean hasAgeYrs() {
      return fieldSetFlags()[2];
    }

    /** Clears the value of the 'ageYrs' field */
    public Event.Builder clearAgeYrs() {
      fieldSetFlags()[2] = false;
      return this;
    }

    /** Gets the value of the 'sizeCm' field */
    public Float getSizeCm() {
      return sizeCm;
    }

    /** Sets the value of the 'sizeCm' field */
    public Event.Builder setSizeCm(float value) {
      validate(fields()[3], value);
      this.sizeCm = value;
      fieldSetFlags()[3] = true;
      return this; 
    }

    /** Checks whether the 'sizeCm' field has been set */
    public boolean hasSizeCm() {
      return fieldSetFlags()[3];
    }

    /** Clears the value of the 'sizeCm' field */
    public Event.Builder clearSizeCm() {
      fieldSetFlags()[3] = false;
      return this;
    }

    @Override
    public Event build() {
      try {
        Event record = new Event();
        record.timestampMS = fieldSetFlags()[0] ? this.timestampMS : (Long) defaultValue(fields()[0]);
        record.name = fieldSetFlags()[1] ? this.name : (CharSequence) defaultValue(fields()[1]);
        record.ageYrs = fieldSetFlags()[2] ? this.ageYrs : (Integer) defaultValue(fields()[2]);
        record.sizeCm = fieldSetFlags()[3] ? this.sizeCm : (Float) defaultValue(fields()[3]);
        return record;
      } catch (Exception e) {
        throw new org.apache.avro.AvroRuntimeException(e);
      }
    }
  }
}

Answer 1

Akka Streams is using asynchronous message passing between Actors to implement stream processing stages. Passing data across an asynchronous boundary has an overhead that you are seeing here: your computation seems to take only about 160ns (derived from the single-threaded measurement) while the streaming solution takes roughly 1µs per element, which is dominated by the message passing.

Another misconception is that saying “stream” implies parallelism: in your code all computation runs sequentially in a single Actor (the map stage), so no benefit can be expected over the primitive single-threaded solution.

In order to benefit from the parallelism afforded by Akka Streams you need to have multiple processing stages that each perform tasks of >1µs per element, see also the docs.