Most efficient way to increment a Map value in Java

Question

I hope this question is not considered too basic for this forum, but we'll see. I'm wondering how to refactor some code for better performance that is getting run a bunch of times.

Say I'm creating a word frequency list, using a Map (probably a HashMap), where each key is a String with the word that's being counted and the value is an Integer that's incremented each time a token of the word is found.

In Perl, incrementing such a value would be trivially easy:

$map{$word}++;

But in Java, it's much more complicated. Here the way I'm currently doing it:

int count = map.containsKey(word) ? map.get(word) : 0;
map.put(word, count + 1);

Which of course relies on the autoboxing feature in the newer Java versions. I wonder if you can suggest a more efficient way of incrementing such a value. Are there even good performance reasons for eschewing the Collections framework and using a something else instead?

Update: I've done a test of several of the answers. See below.

Answer 1

Some test results

I've gotten a lot of good answers to this question--thanks folks--so I decided to run some tests and figure out which method is actually fastest. The five methods I tested are these:

• the "ContainsKey" method that I presented in the question
• the "TestForNull" method suggested by Aleksandar Dimitrov
• the "AtomicLong" method suggested by Hank Gay
• the "Trove" method suggested by jrudolph
• the "MutableInt" method suggested by phax.myopenid.com

Method

Here's what I did...

1. created five classes that were identical except for the differences shown below. Each class had to perform an operation typical of the scenario I presented: opening a 10MB file and reading it in, then performing a frequency count of all the word tokens in the file. Since this took an average of only 3 seconds, I had it perform the frequency count (not the I/O) 10 times.
2. timed the loop of 10 iterations but not the I/O operation and recorded the total time taken (in clock seconds) essentially using Ian Darwin's method in the Java Cookbook.
3. performed all five tests in series, and then did this another three times.
4. averaged the four results for each method.

Results

I'll present the results first and the code below for those who are interested.

The ContainsKey method was, as expected, the slowest, so I'll give the speed of each method in comparison to the speed of that method.

• ContainsKey: 30.654 seconds (baseline)
• AtomicLong: 29.780 seconds (1.03 times as fast)
• TestForNull: 28.804 seconds (1.06 times as fast)
• Trove: 26.313 seconds (1.16 times as fast)
• MutableInt: 25.747 seconds (1.19 times as fast)

Conclusions

It would appear that only the MutableInt method and the Trove method are significantly faster, in that only they give a performance boost of more than 10%. However, if threading is an issue, AtomicLong might be more attractive than the others (I'm not really sure). I also ran TestForNull with final variables, but the difference was negligible.

Note that I haven't profiled memory usage in the different scenarios. I'd be happy to hear from anybody who has good insights into how the MutableInt and Trove methods would be likely to affect memory usage.

Personally, I find the MutableInt method the most attractive, since it doesn't require loading any third-party classes. So unless I discover problems with it, that's the way I'm most likely to go.

The code

Here is the crucial code from each method.

ContainsKey

import java.util.HashMap;
import java.util.Map;
...
Map<String, Integer> freq = new HashMap<String, Integer>();
...
int count = freq.containsKey(word) ? freq.get(word) : 0;
freq.put(word, count + 1);

TestForNull

import java.util.HashMap;
import java.util.Map;
...
Map<String, Integer> freq = new HashMap<String, Integer>();
...
Integer count = freq.get(word);
if (count == null) {
    freq.put(word, 1);
}
else {
    freq.put(word, count + 1);
}

AtomicLong

import java.util.concurrent.ConcurrentHashMap;
import java.util.concurrent.ConcurrentMap;
import java.util.concurrent.atomic.AtomicLong;
...
final ConcurrentMap<String, AtomicLong> map = 
    new ConcurrentHashMap<String, AtomicLong>();
...
map.putIfAbsent(word, new AtomicLong(0));
map.get(word).incrementAndGet();

Trove

import gnu.trove.TObjectIntHashMap;
...
TObjectIntHashMap<String> freq = new TObjectIntHashMap<String>();
...
freq.adjustOrPutValue(word, 1, 1);

MutableInt

import java.util.HashMap;
import java.util.Map;
...
class MutableInt {
  int value = 1; // note that we start at 1 since we're counting
  public void increment () { ++value;      }
  public int  get ()       { return value; }
}
...
Map<String, MutableInt> freq = new HashMap<String, MutableInt>();
...
MutableInt count = freq.get(word);
if (count == null) {
    freq.put(word, new MutableInt());
}
else {
    count.increment();
}

Answer 2

OK, may be an old question, but there is a shorter way with Java 8 :

Map.merge(key, 1, Integer::sum)

What it does : if key do not exists, put 1 as value, otherwise sum 1 to the value linked to key. More information here

Answer 3

A little research in 2016: https://github.com/leventov/java-word-count, benchmark source code

Best results per method (smaller is better):

                 time, ms
kolobokeCompile  18.8
koloboke         19.8
trove            20.8
fastutil         22.7
mutableInt       24.3
atomicInteger    25.3
eclipse          26.9
hashMap          28.0
hppc             33.6
hppcRt           36.5

Time\space results:

Answer 4

Google Guava is your friend...

...at least in some cases. They have this nice AtomicLongMap. Especially nice because you are dealing with long as value in your map.

E.g.

AtomicLongMap<String> map = AtomicLongMap.create();
[...]
map.getAndIncrement(word);

Also possible to add more then 1 to the value:

map.getAndAdd(word, 112L); 

Answer 5

@Hank Gay

As a follow-up to my own (rather useless) comment: Trove looks like the way to go. If, for whatever reason, you wanted to stick with the standard JDK, ConcurrentMap and AtomicLong can make the code a tiny bit nicer, though YMMV.

    final ConcurrentMap<String, AtomicLong> map = new ConcurrentHashMap<String, AtomicLong>();
    map.putIfAbsent("foo", new AtomicLong(0));
    map.get("foo").incrementAndGet();

will leave 1 as the value in the map for foo. Realistically, increased friendliness to threading is all that this approach has to recommend it.

Answer 6

It's always a good idea to look at the Google Collections Library for this kind of thing. In this case a Multiset will do the trick:

Multiset bag = Multisets.newHashMultiset();
String word = "foo";
bag.add(word);
bag.add(word);
System.out.println(bag.count(word)); // Prints 2

There are Map-like methods for iterating over keys/entries, etc. Internally the implementation currently uses a HashMap<E, AtomicInteger>, so you will not incur boxing costs.

Answer 7

You should be aware of the fact that your original attempt

int count = map.containsKey(word) ? map.get(word) : 0;

contains two potentially expensive operations on a map, namely containsKey and get. The former performs an operation potentially pretty similar to the latter, so you're doing the same work twice!

If you look at the API for Map, get operations usually return null when the map does not contain the requested element.

Note that this will make a solution like

map.put( key, map.get(key) + 1 );

dangerous, since it might yield NullPointerExceptions. You should check for a null first.

Also note, and this is very important, that HashMaps can contain nulls by definition. So not every returned null says "there is no such element". In this respect, containsKey behaves differently from get in actually telling you whether there is such an element. Refer to the API for details.

For your case, however, you might not want to distinguish between a stored null and "noSuchElement". If you don't want to permit nulls you might prefer a Hashtable. Using a wrapper library as was already proposed in other answers might be a better solution to manual treatment, depending on the complexity of your application.

To complete the answer (and I forgot to put that in at first, thanks to the edit function!), the best way of doing it natively, is to get into a final variable, check for null and put it back in with a 1. The variable should be final because it's immutable anyway. The compiler might not need this hint, but its clearer that way.

final HashMap map = generateRandomHashMap();
final Object key = fetchSomeKey();
final Integer i = map.get(key);
if (i != null) {
    map.put(i + 1);
} else {
    // do something
}

If you do not want to rely on autoboxing, you should say something like map.put(new Integer(1 + i.getValue())); instead.

Answer 8

Map<String, Integer> map = new HashMap<>();
String key = "a random key";
int count = map.getOrDefault(key, 0);
map.put(key, count + 1);

And that's how you increment a value with simple code.

Benefit:

• Not creating another class for mutable int
• Short code
• Easy to understand
• No null pointer exception

Another way is to use merge method, but this is too much for just incrementing a value.

map.merge(key, 1, (a,b) -> a+b);

Suggestion: you should care about code readability more than little performance gain in most of the time.

Answer 9

Another way would be creating a mutable integer:

class MutableInt {
  int value = 0;
  public void inc () { ++value; }
  public int get () { return value; }
}
...
Map<String,MutableInt> map = new HashMap<String,MutableInt> ();
MutableInt value = map.get (key);
if (value == null) {
  value = new MutableInt ();
  map.put (key, value);
} else {
  value.inc ();
}

of course this implies creating an additional object but the overhead in comparison to creating an Integer (even with Integer.valueOf) should not be so much.

Answer 10

You can make use of computeIfAbsent method in Map interface provided in Java 8.

final Map<String,AtomicLong> map = new ConcurrentHashMap<>();
map.computeIfAbsent("A", k->new AtomicLong(0)).incrementAndGet();
map.computeIfAbsent("B", k->new AtomicLong(0)).incrementAndGet();
map.computeIfAbsent("A", k->new AtomicLong(0)).incrementAndGet(); //[A=2, B=1]

The method computeIfAbsent checks if the specified key is already associated with a value or not? If no associated value then it attempts to compute its value using the given mapping function. In any case it returns the current (existing or computed) value associated with the specified key, or null if the computed value is null.

On a side note if you have a situation where multiple threads update a common sum you can have a look at LongAdder class.Under high contention, expected throughput of this class is significantly higher than AtomicLong, at the expense of higher space consumption.

Answer 11

Memory rotation may be an issue here, since every boxing of an int larger than or equal to 128 causes an object allocation (see Integer.valueOf(int)). Although the garbage collector very efficiently deals with short-lived objects, performance will suffer to some degree.

If you know that the number of increments made will largely outnumber the number of keys (=words in this case), consider using an int holder instead. Phax already presented code for this. Here it is again, with two changes (holder class made static and initial value set to 1):

static class MutableInt {
  int value = 1;
  void inc() { ++value; }
  int get() { return value; }
}
...
Map<String,MutableInt> map = new HashMap<String,MutableInt>();
MutableInt value = map.get(key);
if (value == null) {
  value = new MutableInt();
  map.put(key, value);
} else {
  value.inc();
}

If you need extreme performance, look for a Map implementation which is directly tailored towards primitive value types. jrudolph mentioned GNU Trove.

By the way, a good search term for this subject is "histogram".

Answer 12

Instead of calling containsKey() it is faster just to call map.get and check if the returned value is null or not.

    Integer count = map.get(word);
    if(count == null){
        count = 0;
    }
    map.put(word, count + 1);

Answer 13

Are you sure that this is a bottleneck? Have you done any performance analysis?

Try using the NetBeans profiler (its free and built into NB 6.1) to look at hotspots.

Finally, a JVM upgrade (say from 1.5->1.6) is often a cheap performance booster. Even an upgrade in build number can provide good performance boosts. If you are running on Windows and this is a server class application, use -server on the command line to use the Server Hotspot JVM. On Linux and Solaris machines this is autodetected.

Answer 14

There are a couple of approaches:

1. Use a Bag alorithm like the sets contained in Google Collections.

2. Create mutable container which you can use in the Map:

    class My{
        String word;
        int count;
    }

And use put("word", new My("Word") ); Then you can check if it exists and increment when adding.

Avoid rolling your own solution using lists, because if you get innerloop searching and sorting, your performance will stink. The first HashMap solution is actually quite fast, but a proper like that found in Google Collections is probably better.

Counting words using Google Collections, looks something like this:

    HashMultiset s = new HashMultiset();
    s.add("word");
    s.add("word");
    System.out.println(""+s.count("word") );

Using the HashMultiset is quite elegent, because a bag-algorithm is just what you need when counting words.

Answer 15

I think your solution would be the standard way, but - as you noted yourself - it is probably not the fastest way possible.

You may look at GNU Trove. That is a library which contains all sorts of fast primitive Collections. Your example would use a TObjectIntHashMap which has a method adjustOrPutValue which does exactly what you want.

Answer 16

A variation on the MutableInt approach that might be even faster, if a bit of a hack, is to use a single-element int array:

Map<String,int[]> map = new HashMap<String,int[]>();
...
int[] value = map.get(key);
if (value == null) 
  map.put(key, new int[]{1} );
else
  ++value[0];

It would be interesting if you could rerun your performance tests with this variation. It might be the fastest.

---

Edit: The above pattern worked fine for me, but eventually I changed to use Trove's collections to reduce memory size in some very large maps I was creating -- and as a bonus it was also faster.

One really nice feature is that the TObjectIntHashMap class has a single adjustOrPutValue call that, depending on whether there is already a value at that key, will either put an initial value or increment the existing value. This is perfect for incrementing:

TObjectIntHashMap<String> map = new TObjectIntHashMap<String>();
...
map.adjustOrPutValue(key, 1, 1);

Answer 17

Google Collections HashMultiset :
- quite elegant to use
- but consume CPU and memory

Best would be to have a method like : Entry<K,V> getOrPut(K); (elegant, and low cost)

Such a method will compute hash and index only once, and then we could do what we want with the entry (either replace or update the value).

More elegant:
- take a HashSet<Entry>
- extend it so that get(K) put a new Entry if needed
- Entry could be your own object.
--> (new MyHashSet()).get(k).increment();

Answer 18

"put" need "get" (to ensure no duplicate key).
So directly do a "put",
and if there was a previous value, then do an addition:

Map map = new HashMap ();

MutableInt newValue = new MutableInt (1); // default = inc
MutableInt oldValue = map.put (key, newValue);
if (oldValue != null) {
  newValue.add(oldValue); // old + inc
}

If count starts at 0, then add 1: (or any others values...)

Map map = new HashMap ();

MutableInt newValue = new MutableInt (0); // default
MutableInt oldValue = map.put (key, newValue);
if (oldValue != null) {
  newValue.setValue(oldValue + 1); // old + inc
}

Notice : This code is not thread safe. Use it to build then use the map, not to concurrently update it.

Optimization : In a loop, keep old value to become the new value of next loop.

Map map = new HashMap ();
final int defaut = 0;
final int inc = 1;

MutableInt oldValue = new MutableInt (default);
while(true) {
  MutableInt newValue = oldValue;

  oldValue = map.put (key, newValue); // insert or...
  if (oldValue != null) {
    newValue.setValue(oldValue + inc); // ...update

    oldValue.setValue(default); // reuse
  } else
    oldValue = new MutableInt (default); // renew
  }
}

Answer 19

The various primitive wrappers, e.g., Integer are immutable so there's really not a more concise way to do what you're asking unless you can do it with something like AtomicLong. I can give that a go in a minute and update. BTW, Hashtable is a part of the Collections Framework.

Answer 20

I'd use Apache Collections Lazy Map (to initialize values to 0) and use MutableIntegers from Apache Lang as values in that map.

Biggest cost is having to serach the map twice in your method. In mine you have to do it just once. Just get the value (it will get initialized if absent) and increment it.

Answer 21

The Functional Java library's TreeMap datastructure has an update method in the latest trunk head:

public TreeMap<K, V> update(final K k, final F<V, V> f)

Example usage:

import static fj.data.TreeMap.empty;
import static fj.function.Integers.add;
import static fj.pre.Ord.stringOrd;
import fj.data.TreeMap;

public class TreeMap_Update
  {public static void main(String[] a)
    {TreeMap<String, Integer> map = empty(stringOrd);
     map = map.set("foo", 1);
     map = map.update("foo", add.f(1));
     System.out.println(map.get("foo").some());}}

This program prints "2".

Answer 22

@Vilmantas Baranauskas: Regarding this answer, I would comment if I had the rep points, but I don't. I wanted to note that the Counter class defined there is NOT thread-safe as it is not sufficient to just synchronize inc() without synchronizing value(). Other threads calling value() are not guaranteed to see the the value unless a happens-before relationship has been established with the update.

Answer 23

I don't know how efficient it is but the below code works as well.You need to define a BiFunction at the beginning. Plus, you can make more than just increment with this method.

public static Map<String, Integer> strInt = new HashMap<String, Integer>();

public static void main(String[] args) {
    BiFunction<Integer, Integer, Integer> bi = (x,y) -> {
        if(x == null)
            return y;
        return x+y;
    };
    strInt.put("abc", 0);


    strInt.merge("abc", 1, bi);
    strInt.merge("abc", 1, bi);
    strInt.merge("abc", 1, bi);
    strInt.merge("abcd", 1, bi);

    System.out.println(strInt.get("abc"));
    System.out.println(strInt.get("abcd"));
}

output is

3
1

Answer 24

If you're using Eclipse Collections, you can use a HashBag. It will be the most efficient approach in terms of memory usage and it will also perform well in terms of execution speed.

HashBag is backed by a MutableObjectIntMap which stores primitive ints instead of Counter objects. This reduces memory overhead and improves execution speed.

HashBag provides the API you'd need since it's a Collection that also allows you to query for the number of occurrences of an item.

Here's an example from the Eclipse Collections Kata.

MutableBag<String> bag =
  HashBag.newBagWith("one", "two", "two", "three", "three", "three");

Assert.assertEquals(3, bag.occurrencesOf("three"));

bag.add("one");
Assert.assertEquals(2, bag.occurrencesOf("one"));

bag.addOccurrences("one", 4);
Assert.assertEquals(6, bag.occurrencesOf("one"));

Note: I am a committer for Eclipse Collections.

Answer 25

Quite simple, just use the built-in function in Map.java as followed

map.put(key, map.getOrDefault(key, 0) + 1);

Answer 26

Since a lot of people search Java topics for Groovy answers, here's how you can do it in Groovy:

dev map = new HashMap<String, Integer>()
map.put("key1", 3)

map.merge("key1", 1) {a, b -> a + b}
map.merge("key2", 1) {a, b -> a + b}

Answer 27

Hope I'm understanding your question correctly, I'm coming to Java from Python so I can empathize with your struggle.

if you have

map.put(key, 1)

you would do

map.put(key, map.get(key) + 1)

Hope this helps!