A long time ago in a galaxy far, far away... Java open source developers managed to the see the previously secret plans to the Empire's ultimate weapon, the JAVA™ COLLECTIONS FRAMEWORK. Evading the dreaded Imperial Starfleet, a group of freedom fighters investigate common developer errors and bugs to help protect their vital software. In addition, they investigate the performance of the Empire’s most popular weapon: HashMap. With this new found knowledge they strike back! Pursued by the Empire's sinister agents, JDuchess races home aboard her JVM, investigating proposed future changes to the Java Collections and other options such as Immutable Persistent Collections which could save her people and restore freedom to the galaxy....

1. Java Collections
The Force Awakens
Darth @RaoulUK
Darth @RichardWarburto
#javaforceawakens

2. Evolution can be interesting ...
Java 1.2 Java 10?

3. Collection API Improvements
Persistent & Immutable Collections
Performance Improvements

4. Collection bugs
1. Element access (Off-by-one error, ArrayOutOfBound)
2. Concurrent modification
3. Check-then-Act

5. Scenario 1
List<String> jedis = new ArrayList<>(asList("Luke", "yoda"));
for (String jedi: jedis) {
if (Character.isLowerCase(jedi.charAt(0))) {
jedis.remove(jedi);
}
}

6. Scenario 2
Map<String, BigDecimal> movieViews = new HashMap<>();
BigDecimal views = movieViews.get(MOVIE);
if(views != null) {
movieViews.put(MOVIE, views.add(BigDecimal.ONE));
}
views != nullmoviesViews.get movieViews.put
Then
Check Act

7. Reducing scope for bugs
● ~280 bugs in 28 projects including Cassandra, Lucene
● ~80% check-then-act bugs discovered are put-if-absent
● Library designers can help by updating APIs as new idioms emerge
● Different data structures can provide alternatives by restricting reads &
updates to reduce scope for bugs
CHECK-THEN-ACT Misuse of Java Concurrent Collections
http://dig.cs.illinois.edu/papers/checkThenAct.pdf

8. Java 9 API updates
Collection factory methods
● Non-goal to provide persistent immutable collections
● http://openjdk.java.net/jeps/269
Live Demo using jShell
http://iteratrlearning.com/java9/2016/11/09/java9-collection-factory-methods

9. Collection API Improvements
Persistent & Immutable Collections
Performance Improvements

10. Categorising Collections
Mutable
Immutable
Non-Persistent Persistent
Unsynchronized Concurrent
Unmodifiable View
Available in
Core Library

11. Mutable
● Popular friends include ArrayList, HashMap, TreeSet
● Memory-efficient modification operations
● State can be accidentally modified
● Can be thread-safe, but requires careful design

12. Unmodifiable
List<String> jedis = new ArrayList<>();
jedis.add("Luke Skywalker");
List<String> cantChangeMe = Collections.unmodifiableList(jedis);
// java.lang.UnsupportedOperationException
//cantChangeMe.add("Darth Vader");
System.out.println(cantChangeMe); // [Luke Skywalker]
jedis.add("Darth Vader");
System.out.println(cantChangeMe); // [Luke Skywalker, Darth Vader]

14. Immutable & Non-persistent
● No updates
● Flexibility to convert source in a more efficient representation
● No locking in context of concurrency
● Satisfies co-variant subtyping requirements
● Can be copied with modifications to create a new version (can be
expensive)

15. Immutable vs. Mutable hierarchy
ImmutableList MutableList
+ ImmutableList<T> toImmutable()
java.util.List
+ MutableList<T> toList()
Eclipse Collections (formaly GSCollections) https://projects.eclipse.org/projects/technology.collections/
ListIterable

16. Immutable and Persistent
● Changing source produces a new (version) of the collection
● Resulting collections shares structure with source to avoid full copying
on updates

17. LISP anyone?

18. Persistent List (aka Cons)
public final class Cons<T> implements ConsList<T> {
private final T head;
private final ConsList<T> tail;
public Cons(T head, ConsList<T> tail) {
this.head = head; this.tail = tail;
}
@Override
public ConsList<T> add(T e) {
return new Cons(e, this);
}
}

19. Updating Persistent List
A B C X Y Z
Before

20. Updating Persistent List
A B C X Y Z
Before
A B D
After
Blue nodes indicate new copies
Purple nodes indicates nodes we wish to update

21. Concatenating Two Persistent Lists
A B C
X Y Z
Before

22. Concatenating Two Persistent Lists
- Poor locality due to pointer chasing
- Copying of nodes
A B C
X Y Z
Before
A B C
After

23. Persistent List
● Structural sharing: no need to copy full structure
● Poor locality due to pointer chasing
● Copying becomes more expensive with larger lists
● Poor Random Access and thus Data Decomposition

24. Updating Persistent Binary Tree
Before

25. Updating Persistent Binary Tree
After

26. Persistent Array
How do we get the immutability benefits with performance of mutable
variants?

27. Trie
root
10 4520
3. Picking the right branch is done by using
parts of the key as a lookup
1. Branch factor
not limited to
binary
2. Leaf nodes
contain actual
values
a
a e
b
c
b c f

28. Persistent Array (Bitmapped Vector Trie)
... ...
... ...
... ...
... ...
.
.
.
.
.
.
1 31
0 1 31
Level 1 (root)
Level 2
Leaf nodes

29. Trade-offs
● Large branching factor facilitates iteration but hinders updates
● Small branching factor facilitates updates but hinders traversal

30. Java Persistent Collections
- Not available as part of Java Core Library
- Existing projects includes
- PCollections: https://github.com/hrldcpr/pcollections
- Port of Clojure DS: https://github.com/krukow/clj-ds
- Port of Scala DS: https://github.com/andrewoma/dexx
- Now also in Javaslang: http://javaslang.io

31. Memory usage survey
10,000,000 elements, heap < 32GB
int[] : 40MB
Integer[]: 160MB
ArrayList<Integer>: 215MB
PersistentVector<Integer>: 214MB (Clojure-DS)
Vector<Integer>: 206MB (Dexx, port of Scala-DS)
Data collected using Java Object Layout:
http://openjdk.java.net/projects/code-tools/jol/

32. Takeaways
● Immutable collections reduce the scope for bugs
● Always a compromise between programming safety and performance
● Performance of persistent data structure is improving

33. Collection API Improvements
Persistent & Immutable Collections
Performance Improvements

35. O(N)
O(1)
O(HYPERSPACE)

36. Primitive specialised collections
● Collections often hold boxed representations of primitive values
● Java 8 introduced IntStream, LongStream, DoubleStream and
primitive specialised functional interfaces
● Other libraries, eg: Agrona, Koloboke and Eclipse-Collections provide
primitive specialised collections today.
● Valhalla investigates primitive specialised generics

37. Java 8 Lazy Collection Initialization
Many allocated HashMaps and ArrayLists never written to, eg Null object
pattern
Java 8 adds Lazy Initialization for the default initialization case
Typically 1-2% reduction in memory consumption
http://www.javamagazine.mozaicreader.com/MarApr2016/Twitter#&pageS
et=28&page=0

39. HashMaps Basics
...
Han Solo
hash = 72309
Chewbacca
hash = 72309

40. Chaining Probing
HashMaps
a separate data
structure for
collision lookups
Store inline and
have a probing
sequence

41. Aliases: Palpatine vs Darth Sidious

42. Chaining Probing
HashMaps
aka Closed
Addressing
aka Open Hashing
aka Open
Addressing
aka Closed
Hashing

43. Chaining Probing
HashMaps
Linked List Based Tree Based

44. java.util.HashMap
Chaining Based HashMap
Historically maintained a LinkedList in the case of a collision
Problem: with high collision rates that the HashMap approaches O(N)
lookup

45. java.util.HashMap in Java 8
Starts by using a List to store colliding values.
Trees used when there are over 8 elements
Tree based nodes use about twice the memory
Make heavy collision lookup case O(log(N)) rather than O(N)
Relies on keys being Comparable
https://github.com/RichardWarburton/map-visualiser

46. So which HashMap is best?

47. Example Jar-Jar Benchmark
call get() on a single value for a map
of size 1
No model of the different factors that
affect things!

48. Tree Optimization - 60% Collisions

49. Tree Optimization - 10% Collisions

50. Probing vs Chaining
Probing Maps usually have lower memory consumption
Small Maps: Probing never has long clusters, can be up to 91% faster.
In large maps with high collision rates, probing scales poorly and can be
significantly slower.

51. Takeaways
There’s no clearcut “winner”.
JDK Implementations try to minimise worst case.
Linear Probing requires a good hashCode() distribution, Often hashmaps
“precondition” their hashes.
IdentityHashMap has low memory consumption and is fast, use it!
3rd Party libraries offer probing HashMaps, eg Koloboke & Eclipse-Collections.

52. Conclusions

54. Any Questions?
www.iteratrlearning.com
● Modern Development with Java 8
● Reactive and Asynchronous Java
● Java Software Development Bootcamp
#javaforceawakens

55. Further reading
Fast Functional Lists, Hash-Lists, Deques and Variable Length Arrays
https://infoscience.epfl.ch/record/64410/files/techlists.pdf
Smaller Footprint for Java Collections
http://www.lirmm.fr/~ducour/Doc-objets/ECOOP2012/ECOOP/ecoop/356.pdf
Optimizing Hash-Array Mapped Tries for Fast and Lean Immutable JVM Collections
http://michael.steindorfer.name/publications/oopsla15.pdf
RRB-Trees: Efficient Immutable Vectors
https://infoscience.epfl.ch/record/169879/files/RMTrees.pdf

56. Further reading
Doug Lea’s Analysis of the HashMap implementation tradeoffs
http://www.mail-archive.com/core-libs-dev@openjdk.java.net/msg02147.html
Java Specialists HashMap article
http://www.javaspecialists.eu/archive/Issue235.html
Sample and Benchmark Code
https://github.com/RichardWarburton/Java-Collections-The-Force-Awakens