1. Skyway: Connecting Managed Heaps in Distributed Big Data Systems
Khanh Nguyen, Lu Fang, Christian Navasca,
Harry Xu, Brian Demsky
Shan Lu
University of Chicago
University of California, Irvine

2. BIG DATA

3. The managed runtime is costly
Data Shuffling
MR, Spark
Apps
Send & Receive
Objects
JVM
Skyway
less
The managed runtime is costly

4. deserialization serialization InputStream in =
Shuffler.GetInputStream(sender_id);
while (in.hasData()) {
Object o = in.readObject();
inDataset.store(o) }
inDataset
deserialization
outDataset
serialization
OutputStream out =
Shuffler.GetOutputStream(receiver_id);
for (Object o: outDataset) {
out.writeObject(o); }

5. TriangleCounting over LiveJournal on Spark 2.1.0 with 3 slaves
Data transfer costs
17%
14%
16%
18%
TriangleCounting over LiveJournal on Spark with 3 slaves

6. Data transfer Network Serialization Deserialization Sender Receiver
Object
Reflection.allocate
Serialization
Deserialization
Reflection.getField
Reflection.setField
Network
Binary

7. WANTED: a system-level solution Data transfer Network Serialization
Sender
Receiver
Network
Serialization
Deserialization
Object
Binary
WANTED:
a system-level
solution
Reflection.allocate
Reflection.getField
Reflection.setField

8. an analogy

9. Overlapping computation and data transfer
Our solution
Sender
Receiver
Skyway
Object
Reflection.
allocate
Reflection.
getField
Reflection.
setField
Overlapping computation and data transfer

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%

11. Challenges Type representation Pointer representation
Automated global type numbering
Pointer representation
Use relative offsets
Local JVM adaptation
Visible for garbage collection
Work pipelining
Buffering

12. Type registries

13. Output & Input buffer Input buffer Output buffer Segregated by senders
Multiple for each sender
In managed heap
Segregated by receivers
One for each receiver
In native, off-the-heap memory

14. Example: Serialization
writeObject()
Type Registry
Integer[]
0xbb
0xcc
0xaa 3
Integer 20 30 10
“java.lang.Integer” 6
“[java.lang.Integer” 7
TypeString ID
Output buffer
in native memory
0xbb
0xcc
0xaa 3 7 10 6 20 6 30 6 7 11 15
Offset 7 7 11 15

15. Example: Deserialization
Offset 7 11 15 11 15 7 3 10 20 30 6
readObject()
Input buffer
in heap
Integer[] 7
Integer 6
Integer 6
Integer 6 3
0xf7 7
0xfb 11
0xff 15 10 20 30
java.lang.Integer 6
java.lang.Integer[] 7
MetadataObject ID
Type Registry
java.lang.Integer
java.lang.Integer
java.lang.Integer
java.lang.Integer[]

16. In the paper Cyclic references Shared objects Support for threads
Interaction with GC
Integrating Skyway in real systems

17. Evaluations - Microbenchmark
Java Serializer Benchmark Set
Extensive performance evaluation with existing 90 serializers

18. SKYWAY
GOOGLE’s Protobuf
Kryo
(rec. by Spark)
1.8x
2.2x

19. Evaluations – Real Systems
Flink 1.3.2
5 query answering applications
TPC-H datasets
On average, reduces end-to-end time by 19%

20. Improvement Summary: Flink
Normalized Performance
to built-in serializer
-19%
-23%
-25%
Ser. Time
Deser. Time
Execution Time

21. Evaluations – Real Systems
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)

22. Improvement Summary: Spark
vs. Java
vs. Kryo
Normalized Performance
to Java Serializer
0.05%
-38%
-16%
-36%
-84%
-38%
Ser. Time
Deser. Time
Execution Time

23. Conclusion Goal: Reduce data transfer costs in Big Data systems
Solution: Skyway, the first JVM-based serializer
Efficiently transfer data
Easy to integrate

24. Thank You!