Skyway: Connecting Managed Heaps in Distributed Big Data ... · •Java Serializer Benchmark Set...

Skyway:Connecting Managed Heaps

in Distributed Big Data Systems

Khanh Nguyen, Lu Fang, Christian Navasca,Harry Xu, Brian Demsky

University of ChicagoUniversity of California, IrvineShan Lu

2

2

2

3

3

JVMMR,

Spark Apps

3

JVMMR,

Spark Apps

3

JVMMR,

Spark Apps

3

3

3

3

4

4

4

4

outDataset

4

OutputStream out = Shuffler.GetOutputStream(receiver_id);

for (Object o: outDataset) {out.writeObject(o);

}

serialization

outDataset

4

OutputStream out = Shuffler.GetOutputStream(receiver_id);

for (Object o: outDataset) {out.writeObject(o);

}

serialization

outDataset

4

InputStream in = Shuffler.GetInputStream(sender_id);

while (in.hasData()) {Object o = in.readObject();inDataset.store(o)

} deserialization

4

InputStream in = Shuffler.GetInputStream(sender_id);

while (in.hasData()) {Object o = in.readObject();inDataset.store(o)

} deserialization

inDataset

Tim

e (s

ec)

Serializers0

350

700

1050

1400

1750

Data transfer costs

5TriangleCounting over LiveJournal on Spark 2.1.0 with 3 slaves

Tim

e (s

ec)

Serializers0

350

700

1050

1400

1750

Data transfer costs

5TriangleCounting over LiveJournal on Spark 2.1.0 with 3 slaves

14%18%

17%

16%

Data transfer

6

Receiver

Object

Sender

Data transfer

6

Receiver

Serialization

Object

Sender

Data transfer

6

Receiver

Serialization

Object

Reflection.getField

Sender

Data transfer

6

Receiver

Serialization

Object

Reflection.getField

Sender

Data transfer

6

Receiver

Serialization

Object

Reflection.getField

Sender

Data transfer

6

Receiver

Serialization

Object

Binary

Reflection.getField

Sender

Network

Data transfer

6

Receiver

Serialization

Object

Binary

Reflection.getField

Sender

Network

Data transfer

6

Receiver

Object

Binary

Sender

Data transfer

6

Receiver

Deserialization

Object

Binary

Sender

Data transfer

6

Receiver

Deserialization

Object

Binary

Reflection.setField

Reflection.allocate

Sender

Data transfer

6

Receiver

Deserialization

Object

Binary

Reflection.setField

Reflection.allocate

Sender

Data transfer

6

Receiver

Deserialization

Object

Binary

Reflection.setField

Reflection.allocate

Sender

Data transfer

7

Receiver

Reflection.getField

Reflection.setField

Reflection.allocate

Sender

Data transfer

7

ReceiverSender

8

8

8

8

8

8

Our solution

9

Our solution

9

ReceiverSender

Our solution

9

Object

ReceiverSender

Our solution

9

Object

ReceiverSender

Skyway

Our solution

9

Object

Reflection.getField

Reflection.setField

Reflection.allocate

ReceiverSender

Skyway

Our solution

9

Object

Reflection.getField

Reflection.setField

Reflection.allocate

ReceiverSender

Skyway

Our solution

9

Object

Reflection.getField

Reflection.setField

Reflection.allocate

ReceiverSender

Skyway

Skyway Overview

10

Skyway Overview

– Implemented in OpenJDK 8• Modified the class loader, the object/heap layout,

the Parallel Scavenge GC– Efficiently handle data transfer:

• Outperforms 90 serializers• Improves Spark by 36% (Java) - 16% (Kryo)• Improves Flink by 19%

10

Challenges

11

Challenges

1. Type representation

11

Challenges

1. Type representation Automated global type numbering

11

Challenges


2. Pointer representation

11

Challenges


2. Pointer representation Use relative offsets

11

Challenges



3. Local JVM adaptation

11

Challenges



3. Local JVM adaptation Visible for garbage collection

11

Challenges




4. Work pipelining

11

Challenges




4. Work pipelining Buffering

11

Type registries

12

Metadata Object

Worker A

TypeString ID12

“java. lang.Object”“org.apache.spark.

rdd.RDD”“java. lang.String” 5

Type Registry A

klass for “java.lang.Object”

klass for “java.lang.String”

Worker B

TypeString ID15

“java. lang.Object”“java. lang.String”

120

Type Registry B

klass for “org.apache.spark.scheduler.Task”

klass for “java.lang.Object”

“org.apache.spark.scheduler.Task”

TypeString ID12

4

“java. lang.Object”“org.apache.spark.

rdd.RDD”“java.util.HashMap”

...

3“java.util.HashSet”“java. lang.String” 5

...“org.apache.spark.

scheduler.Task” 120

Type Registry

Cluster

klass for “org.apache.spark.rdd.RDD”

Metadata Object

klass for “java.lang.String”

Master

Output & Input buffer

13


13

Output buffer


13

Output buffer


13

Output buffer

•Segregated by receivers•One for each receiver


13

Output buffer

•Segregated by receivers•One for each receiver•In native, off-the-heap memory


13

Output bufferInput buffer



13




13



•Segregated by senders•Multiple for each sender


13



•Segregated by senders•Multiple for each sender•In managed heap

Example: Serialization

14

Integer[] 0xbb 0xcc0xaa3

Integer 20 Integer 30Integer 10


14



writeObject()


14



writeObject()


14



Output bufferin native memory

writeObject()


14



0Offset


writeObject()


14



0Offset


writeObject()


14



0Offset


“java.lang.Integer” 6“[java.lang.Integer” 7

TypeString ID

writeObject()


14



0xbb 0xcc0xaa37

0Offset



TypeString ID

writeObject()


14



0xbb 0xcc0xaa37

0Offset



TypeString ID

writeObject()


14



0xbb 0xcc0xaa37

0Offset



TypeString ID

7

writeObject()


14



0xbb 0xcc0xaa37 106

0Offset



TypeString ID

7

writeObject()


14



0xbb 0xcc0xaa37 106

0Offset


7


TypeString ID

7

writeObject()


14



0xbb 0xcc0xaa37 106 206

0 11Offset


7 11


TypeString ID

7

writeObject()


14



0xbb 0xcc0xaa37 106 206 306

0 11 15Offset


7 11 15


TypeString ID

7

writeObject()

Example: Deserialization

15

0 7 11 15

0xfb 0xff 20 300xf7 103

Offset

7 6 6 611 157

11 15737 10 20 306 6 6


15

0 7 11 15

0xfb 0xff 20 300xf7 103

Offset

7 6 6 611 157

11 15737 10 20 306 6 6

readObject()


15

0 7 11 15

0xfb 0xff 20 300xf7 103

Offset

7 6 6 611 157

Input buffer in heap

readObject()


15

0 7 11 15

0xfb 0xff 20 300xf7 103

Offset

7 6 6 611 157


java.lang.Integer6java.lang.Integer[]7

MetadataObjectID

readObject()


15

0 7 11 15

0xfb 0xff 20 300xf7 10Integer[] 3

Offset

6 6 611 157



MetadataObjectID

readObject()


15

0 7 11 15


Offset

6 6 611 15



MetadataObjectID

readObject()


15

0 7 11 15


Offset

6 6 615



MetadataObjectID

readObject()


15

0 7 11 15


Offset

6 6 6



MetadataObjectID

readObject()


15

0 7 11 15

0xfb 0xff 20 300xf7 10Integer[] 3 Integer

Offset

6 6



MetadataObjectID

readObject()


15

0 7 11 15

0xfb 0xff 20 300xf7 10Integer[] 3 IntegerInteger

Offset

6



MetadataObjectID

readObject()


15

0 7 11 15

0xfb 0xff 20 300xf7 10Integer[] 3 IntegerIntegerInteger

Offset



MetadataObjectID

readObject()

In the paper

• Cyclic references• Shared objects• Support for threads• Interaction with GC• Integrating Skyway in real systems

16

Evaluations - Microbenchmark

17

• Java Serializer Benchmark Set– Extensive performance evaluation with

existing 90 serializers

18

SKYWAY

18

1.8x

SKYWAY

GOOGLE’s Protobuf

18

1.8x

2.2x

SKYWAY

GOOGLE’s Protobuf

Kryo(rec. by Spark)

Evaluations – Real Systems

19

• Flink 1.3.2– 5 query answering applications– TPC-H datasets


19



19


On average, reduces end-to-end time by 19%

0

0.2

0.4

0.6

0.8

1

1.2

Improvement Summary: Flink

20Execution TimeSer. Time Deser. Time

Norm

alize

d Pe

rform

ance

to b

uilt-

in se

rializ

er

0

0.2

0.4

0.6

0.8

1

1.2



Norm

alize

d Pe

rform

ance

to b

uilt-

in se

rializ

er

0

0.2

0.4

0.6

0.8

1

1.2



Norm

alize

d Pe

rform

ance

to b

uilt-

in se

rializ

er

0

0.2

0.4

0.6

0.8

1

1.2



Norm

alize

d Pe

rform

ance

to b

uilt-

in se

rializ

er


21

• Spark 2.1.0– 4 applications: WordCount, PageRank,

ConnectedComponents, and TriangleCounting

– 4 datasets:LiveJournal, Orkut, UK-2005, and Twitter


21

• Spark 2.1.0– 4 applications: WordCount, PageRank,

ConnectedComponents, and TriangleCounting

– 4 datasets:LiveJournal, Orkut, UK-2005, and Twitter

On average, reduces end-to-end time by 16% (w.r.t. Kryo)

by 36% (w.r.t. Java serializer)

Kryo Skyway0

0.2

0.4

0.6

0.8

1

1.2

1.4

Kryo Skyway Kryo Skyway

Improvement Summary: Spark

22

Norm

alize

d Pe

rform

ance

to

Java

Ser

ialize

r

Execution TimeSer. Time Deser. Time

Kryo Skyway0

0.2

0.4

0.6

0.8

1

1.2

1.4



22

Norm

alize

d Pe

rform

ance

to

Java

Ser

ialize

r


Kryo Skyway0

0.2

0.4

0.6

0.8

1

1.2

1.4



22

Norm

alize

d Pe

rform

ance

to

Java

Ser

ialize

r


Kryo Skyway0

0.2

0.4

0.6

0.8

1

1.2

1.4



22

Norm

alize

d Pe

rform

ance

to

Java

Ser

ialize

r


Kryo Skyway0

0.2

0.4

0.6

0.8

1

1.2

1.4



22

Norm

alize

d Pe

rform

ance

to

Java

Ser

ialize

r


Kryo Skyway0

0.2

0.4

0.6

0.8

1

1.2

1.4



22

Norm

alize

d Pe

rform

ance

to

Java

Ser

ialize

r


Conclusion• Goal: Reduce data transfer costs in

Big Data systems

• Solution: Skyway, the first JVM-based serializer– Efficiently transfer data– Easy to integrate

23

Thank You!

24

Date post:	20-Sep-2020
Category:	Documents
Upload:	others
View:	0 times
Download:	0 times

Skyway: Connecting Managed Heaps in Distributed Big Data ... · •Java Serializer Benchmark Set...

Documents