When should I use ConcurrentSkipListMap?

Question

In Java, ConcurrentHashMap is there for better multithreading solution. Then when should I use ConcurrentSkipListMap? Is it a redundancy?

Does multithreading aspects between these two are common?

Solution 1

These two classes vary in a few ways.

ConcurrentHashMap does not guarantee* the runtime of its operations as part of its contract. It also allows tuning for certain load factors (roughly, the number of threads concurrently modifying it).

ConcurrentSkipListMap, on the other hand, guarantees average O(log(n)) performance on a wide variety of operations. It also does not support tuning for concurrency's sake. ConcurrentSkipListMap also has a number of operations that ConcurrentHashMap doesn't: ceilingEntry/Key, floorEntry/Key, etc. It also maintains a sort order, which would otherwise have to be calculated (at notable expense) if you were using a ConcurrentHashMap.

Basically, different implementations are provided for different use cases. If you need quick single key/value pair addition and quick single key lookup, use the HashMap. If you need faster in-order traversal, and can afford the extra cost for insertion, use the SkipListMap.

_{*Though I expect the implementation is roughly in line with the general hash-map guarantees of O(1) insertion/lookup; ignoring re-hashing}

Solution 2

Sorted, navigable, and concurrent

See Skip List for a definition of the data structure.

A ConcurrentSkipListMap stores the Map in the natural order of its keys (or some other key order you define). So it'll have slower get/put/contains operations than a HashMap, but to offset this it supports the SortedMap, NavigableMap, and ConcurrentNavigableMap interfaces.

Solution 3

In terms of performance, skipList when is used as Map - appears to be 10-20 times slower. Here is the result of my tests (Java 1.8.0_102-b14, win x32)

Benchmark                    Mode  Cnt  Score    Error  Units
MyBenchmark.hasMap_get       avgt    5  0.015 ?  0.001   s/op
MyBenchmark.hashMap_put      avgt    5  0.029 ?  0.004   s/op
MyBenchmark.skipListMap_get  avgt    5  0.312 ?  0.014   s/op
MyBenchmark.skipList_put     avgt    5  0.351 ?  0.007   s/op

And additionally to that - use-case where comparing one-to-another really makes sense. Implementation of the cache of last-recently-used items using both of these collections. Now efficiency of skipList looks to be event more dubious.

MyBenchmark.hashMap_put1000_lru      avgt    5  0.032 ?  0.001   s/op
MyBenchmark.skipListMap_put1000_lru  avgt    5  3.332 ?  0.124   s/op

Here is the code for JMH (executed as java -jar target/benchmarks.jar -bm avgt -f 1 -wi 5 -i 5 -t 1)

static final int nCycles = 50000;
static final int nRep = 10;
static final int dataSize = nCycles / 4;
static final List<String> data = new ArrayList<>(nCycles);
static final Map<String,String> hmap4get = new ConcurrentHashMap<>(3000, 0.5f, 10);
static final Map<String,String> smap4get = new ConcurrentSkipListMap<>();

static {
    // prepare data
    List<String> values = new ArrayList<>(dataSize);
    for( int i = 0; i < dataSize; i++ ) {
        values.add(UUID.randomUUID().toString());
    }
    // rehash data for all cycles
    for( int i = 0; i < nCycles; i++ ) {
        data.add(values.get((int)(Math.random() * dataSize)));
    }
    // rehash data for all cycles
    for( int i = 0; i < dataSize; i++ ) {
        String value = data.get((int)(Math.random() * dataSize));
        hmap4get.put(value, value);
        smap4get.put(value, value);
    }
}

@Benchmark
public void skipList_put() {
    for( int n = 0; n < nRep; n++ ) {
        Map<String,String> map = new ConcurrentSkipListMap<>();

        for( int i = 0; i < nCycles; i++ ) {
            String key = data.get(i);
            map.put(key, key);
        }
    }
}

@Benchmark
public void skipListMap_get() {
    for( int n = 0; n < nRep; n++ ) {
        for( int i = 0; i < nCycles; i++ ) {
            String key = data.get(i);
            smap4get.get(key);
        }
    }
}

@Benchmark
public void hashMap_put() {
    for( int n = 0; n < nRep; n++ ) {
        Map<String,String> map = new ConcurrentHashMap<>(3000, 0.5f, 10);

        for( int i = 0; i < nCycles; i++ ) {
            String key = data.get(i);
            map.put(key, key);
        }
    }
}

@Benchmark
public void hasMap_get() {
    for( int n = 0; n < nRep; n++ ) {
        for( int i = 0; i < nCycles; i++ ) {
            String key = data.get(i);
            hmap4get.get(key);
        }
    }
}

@Benchmark
public void skipListMap_put1000_lru() {
    int sizeLimit = 1000;

    for( int n = 0; n < nRep; n++ ) {
        ConcurrentSkipListMap<String,String> map = new ConcurrentSkipListMap<>();

        for( int i = 0; i < nCycles; i++ ) {
            String key = data.get(i);
            String oldValue = map.put(key, key);

            if( (oldValue == null) && map.size() > sizeLimit ) {
                // not real lru, but i care only about performance here
                map.remove(map.firstKey());
            }
        }
    }
}

@Benchmark
public void hashMap_put1000_lru() {
    int sizeLimit = 1000;
    Queue<String> lru = new ArrayBlockingQueue<>(sizeLimit + 50);

    for( int n = 0; n < nRep; n++ ) {
        Map<String,String> map = new ConcurrentHashMap<>(3000, 0.5f, 10);

        lru.clear();
        for( int i = 0; i < nCycles; i++ ) {
            String key = data.get(i);
            String oldValue = map.put(key, key);

            if( (oldValue == null) && lru.size() > sizeLimit ) {
                map.remove(lru.poll());
                lru.add(key);
            }
        }
    }
}

Solution 4

Then when should I use ConcurrentSkipListMap?

When you (a) need to keep keys sorted, and/or (b) need the first/last, head/tail, and submap features of a navigable map.

The ConcurrentHashMap class implements the ConcurrentMap interface, as does ConcurrentSkipListMap. But if you also want the behaviors of SortedMap and NavigableMap, use ConcurrentSkipListMap

ConcurrentHashMap

• ❌ Sorted
• ❌ Navigable
• ✅ Concurrent

ConcurrentSkipListMap

• ✅ Sorted
• ✅ Navigable
• ✅ Concurrent

Here is table guiding you through the major features of the various Map implementations bundled with Java 11. Click/tap to zoom.

Keep in mind that you can obtain other Map implementations, and similar such data structures, from other sources such as Google Guava.

Solution 5

Based on workloads ConcurrentSkipListMap could be slower than TreeMap with synchronized methods as in KAFKA-8802 if range queries are needed.

View more solutions

Comments

[DKSRathore]

Ok. Log(n) is fine but does ConcurrentSkipListMap is space efficient?

[Kevin Montrose]

Skip lists are necessarily larger than Hashtables, but the tunable nature of ConcurrentHashMap makes it possible to construct a Hashtable that is larger than the equivalent ConcurrentSkipListMap. In general, I'd expect the skip list to be larger but on the same order of magnitude.

[Pacerier]

"It also does not support tuning for concurrency's sake".. Why? What is the link?

[Kevin Montrose]

@Pacerier - I didn't mean it does support tuning because it's concurrent, I mean it doesn't allow you to tune parameters that influence it's concurrent performance (while ConcurrentHashMap does).

[Pacerier]

@KevinMontrose Ic, so you meant "It also does not support concurrency tuning."

[abbas]

I think ConcurrentSkipListMap should be compared to their non-concurrent counter part, the TreeMap.

[yusong]

Thanks for sharing. I am justing thinking to replace TreeMap with ConcurrentSkipListMap in one of my project, so it is good to know this! Do you have more context about why ConcurrentSkipListMap is slower, and more detail about performance comparison?

[Basil Bourque]

As abbas commented, comparing performance to ConcurrentHashMap seems silly. The purpose of ConcurrentSkipListMap is to (a) provide concurrency, and (b) maintain sort-order amongst the keys. ConcurrentHashMap does a, but not b. Comparing the 0-to-60 speed of a Tesla and a dump truck is senseless, as they serve different purposes.

[Xtra Coder]

While you don't know performance metrics you don't know which one is Tesla and which one is "dump truck" :) Also ... you do not know the price of "b" without metrics. Therefore - comparing performance is generally useful thing.

[simgineer]

Maybe add a comparison to tree map! :D

[AnV]

This picture is awesome. Do you have similar pictures for some or all of normal and concurrent collections?

[Basil Bourque]

@Anv Thanks, making that chart took quite a bit of work. It is part of my presentation for Java user groups: A map to Java Maps. And, No, I’ve made just one other class diagram, for String related classes and interface.

[Robert]

Skip lists are much more memory efficient. For me that's more important than a few milliseconds in lookups.

[Robert]

I find it odd that everyone likes to compare time complexity in the absence of required space. SkipLists are much more memory efficient (a point that I think is worth noting)

[Alec]

It feels a little bit weird to benchmark the performance of a datastructure meant for concurrent access only single threaded. This would be more interesting with more threads (even just -tc 1,2,4,8 to see performance with 1, 2, 4, 8 threads). I expect concurrent hashmap (even with the default config) would trounce concurrent skip list map, but the curve in vertical scalability could be interesting.