![5
OutOfMemoryError
● Memory limit (-Xmx1G)
● Add entries to array in loop
● Print number of items
● JVM eventually crashes
try{
//add new entry in loop
while(true)
array[arrayPos++] = new HashMap();
}catch(OutOfMemoryError e){
memoryConsumption =
Runtime.getRuntime().maxMemory()
/ counter;
return;
}](https://image.slidesharecdn.com/save-java-memory-160108210536/85/Save-Java-memory-5-320.jpg)
![6
Profiler
● Shows memory consumed
by single class
– (byte[], Object[])
● VisualVM is easy to use
● Flight Recorder is best
– Free outside production env.](https://image.slidesharecdn.com/save-java-memory-160108210536/85/Save-Java-memory-6-320.jpg)
![10
java.lang.String overhead
● Two objects and two references
● One int field (String.hash)
● Two bytes per char (UTF16)
● Other problems:
– Weak hash (many collisions)
– Defensive char[] copies
– Substring create copies etc..
JOL output for java.lang.String
OFFSET SIZE TYPE DESC
0 4 (object header)
4 4 (object header)
8 4 (object header)
12 4 char[] String.value
16 4 int String.hash
20 4 (alignment loss)
total size: 24 bytes](https://image.slidesharecdn.com/save-java-memory-160108210536/85/Save-Java-memory-10-320.jpg)
![12
Represent String as char[]
● Field is char[] or byte[]
● Get/set takes string
● GC overhead, but short lived
objects are fast to collect
● Reduced memory with many
objects (large Maps etc)
public class Person{
private char[] name;
public String getName(){
return new String(name);
}
public void setName(String name){
this.name = name.toCharArray();
}
}](https://image.slidesharecdn.com/save-java-memory-160108210536/85/Save-Java-memory-12-320.jpg)
![13
char[] as key in Map
● char[] has no equals() and hashCode() methods
●
TreeMap key can use Comparator<char[]>
– com.google.common.primitives.Chars.lexicographicalComparator()
● java.util.HashMap can not change hash strategy
– Object2ObjectOpenCustomHashMap from FastUtil has Hash Strategies
– String.hashCode() is broken, many collisions
● Part of spec, can not be changed :-(
● ConcurrentHashMap Java8 has sorted lists to handle hash duplicates :-)](https://image.slidesharecdn.com/save-java-memory-160108210536/85/Save-Java-memory-13-320.jpg)
![14
char[] as value in Map
● Works (mostly)
● Concurrent Compare-And-Swap operations broken
– char[].equals() uses identity
– ConcurrentMap.replace(K key, V oldValue, V newValue);
– MapDB Maps allow custom equality](https://image.slidesharecdn.com/save-java-memory-160108210536/85/Save-Java-memory-14-320.jpg)
![16
Primitive key / value in collections
● HashMap<Long,Long> uses Object[] for keys and values
● ~40 bytes overhead for each key-value pair
● Use long[] or int[] instead
● Data locality → CPU cache friendly
– Binary search much faster
– Faster fetch of values
● Almost zero Garbage Collection
– entire HashMap uses two objects
– No objects allocated for new key or value](https://image.slidesharecdn.com/save-java-memory-160108210536/85/Save-Java-memory-16-320.jpg)
![19
Chronicle Map
● ConcurrentHashMap
● Keys and values are serialized
outside heap
● Supports any serializable data
type
● No objects or references on
heap, memory efficient
● http://chronicle.software/product
s/chronicle-map/
Map<Long, long[]> graph =
ChronicleMap
.of(Long.class, long[].class)
.entries(1_000_000_000L)
.averageValue(new long[150])
.create();](https://image.slidesharecdn.com/save-java-memory-160108210536/85/Save-Java-memory-19-320.jpg)
Save Java memory
- 1. 1 ● Jan Kotek – [email protected] – @JanKotek – MapDB author (database engine) – Java developer for past 12 years – Last 3 years independent consultant for MapDB ● Slides and code at – https://github.com/jankotek/talk-save-java-memory Intro
- 2. 2 Intro ● This talk is about: – Reducing memory footprint – Object overhead – Primitive variables and arrays – Alternative Java Collections ● Not about: – Garbage Collection – Off-heap – Performance – Memory leaks
- 3. 3 Measure memory consumption ● Runtime.getRuntime().freeMemory() difference ● Allocate until OutOfMemoryError ● Profiler ● Class or heap dump analyze such as Java Object Layout
- 4. 4 Runtime.getRuntime().freeMemory() ● Get allocated memory stats ● Create new objects ● Call GC many times ● Get allocated memory difference ● Not very precise, but easy ● May not always work, some JVMs ignore System.gc() NanoBench nanoBench = NanoBench.create(); nanoBench.memoryOnly().warmUps(2).measurements(10) .measure(String.format("Measure memory for %d keys", i), new Runnable() { @Override public void run() { Random setRandom = new Random(4532); ref = new HashSet<Integer>(cnt); w hile (ref.size() < cnt) { ref.add(setRandom.nextInt()); } } }); nanoBench.getMemoryBytes()
- 5. 5 OutOfMemoryError ● Memory limit (-Xmx1G) ● Add entries to array in loop ● Print number of items ● JVM eventually crashes try{ //add new entry in loop while(true) array[arrayPos++] = new HashMap(); }catch(OutOfMemoryError e){ memoryConsumption = Runtime.getRuntime().maxMemory() / counter; return; }
- 6. 6 Profiler ● Shows memory consumed by single class – (byte[], Object[]) ● VisualVM is easy to use ● Flight Recorder is best – Free outside production env.
- 7. 7 Java Object Layout ● Command line tool to analyze memory layout (and space usage) ● Most advanced ● http://openjdk.java.net/projects/code-tools/jol/ ● Written by Aleksey Shipilev ● Example: $ java -jar jol-cli/target/jol-cli.jar internals java.util.HashMap Instance size: 48 bytes (reported by Instrumentation API) Space losses: 0 bytes internal + 4 bytes external = 4 bytes total
- 8. 8 Instance overhead in Java ● Pointer overhead – 4 to 8 bytes – depends on JVM and amount of memory – 64bit JVM 8 uses compressed pointer when possible ● Class info and object header – 12 bytes ● Object alligment padding – Between 0 and 15 35 bytes worst case scenario
- 9. 9 java.lang.Integer overhead ● One object + one reference ● Total overhead 16 bytes + reference (4 to 8 bytes) JOL output for java.lang.Integer OFFSET SIZE TYPE DESC 0 12 (object header) 12 4 int Integer.value total size: 16 bytes
- 10. 10 java.lang.String overhead ● Two objects and two references ● One int field (String.hash) ● Two bytes per char (UTF16) ● Other problems: – Weak hash (many collisions) – Defensive char[] copies – Substring create copies etc.. JOL output for java.lang.String OFFSET SIZE TYPE DESC 0 4 (object header) 4 4 (object header) 8 4 (object header) 12 4 char[] String.value 16 4 int String.hash 20 4 (alignment loss) total size: 24 bytes
- 11. 11 String.intern() ● Reduces number of instances ● Each string creates new weak ref (huge GC overhead) ● Long.valueOf(11) is similar name.toLowerCase().intern() == name.toLowerCase().intern()
- 12. 12 Represent String as char[] ● Field is char[] or byte[] ● Get/set takes string ● GC overhead, but short lived objects are fast to collect ● Reduced memory with many objects (large Maps etc) public class Person{ private char[] name; public String getName(){ return new String(name); } public void setName(String name){ this.name = name.toCharArray(); } }
- 13. 13 char[] as key in Map ● char[] has no equals() and hashCode() methods ● TreeMap key can use Comparator<char[]> – com.google.common.primitives.Chars.lexicographicalComparator() ● java.util.HashMap can not change hash strategy – Object2ObjectOpenCustomHashMap from FastUtil has Hash Strategies – String.hashCode() is broken, many collisions ● Part of spec, can not be changed :-( ● ConcurrentHashMap Java8 has sorted lists to handle hash duplicates :-)
- 14. 14 char[] as value in Map ● Works (mostly) ● Concurrent Compare-And-Swap operations broken – char[].equals() uses identity – ConcurrentMap.replace(K key, V oldValue, V newValue); – MapDB Maps allow custom equality
- 15. 15 Bitwise class Address{ String street; int houseNumber; } ● Replace street with number from dictionary ● Combine streetId and houseNumber into single long number – 5 bytes for streetId, 3 bytes for house number and flags ● Adress become primitive number, no objects or references
- 16. 16 Primitive key / value in collections ● HashMap<Long,Long> uses Object[] for keys and values ● ~40 bytes overhead for each key-value pair ● Use long[] or int[] instead ● Data locality → CPU cache friendly – Binary search much faster – Faster fetch of values ● Almost zero Garbage Collection – entire HashMap uses two objects – No objects allocated for new key or value
- 17. 17 FastUtil ● http://fastutil.di.unimi.it/ ● Autogenerated classes with primitive keys and values ● Int2IntMap, Int2BooleanMap, Int2ObjectMap<V> …. ● Many classes, 17MB jar
- 18. 18 FastUtil // create new map Int2ObjectMap<String> map = new Int2ObjectOpenHashMap(); // put/get has primitive arg, no unboxing! map.put(1, "aa"); map.get(1); // also implements full mapdb interface Map<Integer, String> map2 = map; ● Parallel interface hierarchy ● Get / set has primitive arg version, no unboxing! ● Implements Map interface
- 19. 19 Chronicle Map ● ConcurrentHashMap ● Keys and values are serialized outside heap ● Supports any serializable data type ● No objects or references on heap, memory efficient ● http://chronicle.software/product s/chronicle-map/ Map<Long, long[]> graph = ChronicleMap .of(Long.class, long[].class) .entries(1_000_000_000L) .averageValue(new long[150]) .create();
- 20. 20 Conclusion ● Avoid too many object instances in long lived objects ● Short lived objects are fine, when traded for smaller memory ● Email: [email protected] ● Slides and code at – https://github.com/jankotek/talk-save-java-memory