Skip to content

mfktech/lockfree-library

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

26 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

LockFree Library

1. Introduction

LockFree library utilizes non-blocking Atomic classes to implement concurrent List, Queue and Map.

2. Jar Artifact

Group Id: com.mfk.lockfree
Artifact Id: lockfree-library
Latest Version: 1.0

3. Features

3.1 Linked-Array Data Structure

In a typical Linked List, the new objects are appended to the tail pointer, allowing it to increase in size dynamically in O(1). The problem is that the Linked List is not CPU Cache friendly because CPU needs to re-load the L1/L2/L3 caches when accessing individual entry in a Linked List.

Arrays, on the other hand, provide stride-1 performance (if accessed linearly), but they are fixed size and, therefore, cannot be increased in size dynamically. Java's ArrayList solves this problem by creating a new array and copying over all the elements from the old list to the new one, but this effort is not so efficient when dealing with large amount of elements, millions or billions.

The Linked-Array data structure gives the benefits of both worlds, Array and Linked List. This data structure is the combination of Linked List and Array data structures, referred to as Linked Array data structure. The objects are stored in fixed size arrays, referred to as Fragments, which are then chained together to form linked arrays. The fragment size is configurable; default value is 1000. All collections in this library uses this data structure internally.

3.2 LockFree (Atomic classes)

Atomic classes are used for concurrency, which are CAS based (if supported by underlying CPU). Please see JMH Benchmarking results to see the comparison between various other collections with the data structures offered in this library.

4. Queue

A non-blocking and concurrent Queue which provides thread-safe operations using Lock-Free algorithms (Atomic classes).

4.1. Add/Append

Add operation allows elements to be appended to the tail concurrently. Following is an example

    LockFreeQueue<String> queue = LockFreeQueue.newQueue();
    queue.add("element1");
    queue.add("element2");
    
    assertEquals(2, queue.size());

4.2. Poll

Poll operation will remove and return the head of the queue. If the queue is empty, then it will return Optional.empty object. Following is an example

    LockFreeQueue<String> queue = LockFreeQueue.newQueue();
    queue.add("element1");
    
    queue.poll().ifPresent(System.out::println);

5. Map

A non-blocking and concurrent Map which provides thread-safe operations using Lock-Free algorithms (Atomic classes). This map relies on the same rules for equals and hashCode as in Java to put and lookup elements. Please do internet search to learn more about hashCode/equals contract in Java.

5.1. Insertion/Put

Insert or Put operation is thread-safe which allows us to insert/put elements concurrently. It is highly recommended to provide efficient and consistent implementation of hashCode method.

One important difference from Java's Map is that the put operation will not overwrite the existing value if put method is called multiple times for the same key, instead the value objects are retained and can be retrieved by using getAll method.

    final LockFreeMap<String, String> map = LockFreeMap.newMap();
    map.put("key1", "value1");
    map.put("key1", "newValue1");

    final List<String> values = map.getAll("key1").collect(Collectors.toList());
    assertEquals(Arrays.asList("value1", "newValue1"), values);

5.2. Retrieval

Assuming equals method is implemented properly, the value objects can be retrieved by using the same key, as shown below

    final LockFreeMap<String, String> map = LockFreeMap.newMap();
    map.put("key1", "value1");
    assertEquals("value1", map.get("key1"));    

getAll method can be used to retrieve all previously put values for the same key, as shown below

    LockFreeMap<String, String> map = LockFreeMap.newMap();
    map.put("key1", "value1");
    map.put("key1", "newValue1");

    final List<String> values = map.getAll("key1").collect(Collectors.toList());
    
    // getAll method should return all previously set values for the key 'key1'
    assertEquals(Arrays.asList("value1", "newValue1"), values);

    // get method will return the last set value.  
    final Optional<String> value = map.get("key1");
    assertTrue(value.isPresent());
    assertEquals("newValue1", "newValue1", value.get());

If there exists multiple values for the same key, then get method will return the value which was put last. However, in highly concurrent environment, the "last" element would be in-deterministic because of the race condition. In that case getAll should be used to get all the values for that key.

5.3. Deletion

remove method can be used to remove all values for the given key, as shown below

    LockFreeMap<String, String> map = LockFreeMap.newMap();
    map.put("k1", "v1");
    map.put("k1", "v2");
    assertEquals(Arrays.asList("v1", "v2"), map.getAll("k1").collect(toList()));

    map.remove("k1");
    assertEquals(Collections.emptyList(), map.getAll("k1").collect(toList()));

5.4. LockFreeMap vs ConcurrentMap

JMH library is used to perform benchmarking between operations of LockFreeMap and ConcurrentHashMap.

All the benchmarking classes pertaining to LockFreeMap can be found in com.mfk.lockfree.benchmark.map package. If you are interested in running the benchmarking yourself, then execute the following command after mvn clean package:

java -jar jmh-benchmark/target/jmh-benchmark.jar "com.mfk.lockfree.benchmark.map.<class name>.*"

5.4.1. Put Benchmarking

8 threads put 1000 elements concurrently. Elements produced unique hash code, so there would be less number of collisions. This was repeated 100 times.

Following is the comparision of Average Response Time (avgt) using JMH benchmarking.

Benchmark Mode Cnt Score Error Units
PutBenchmark.measureJavaMap avgt 100 0.482 ±0.002 ms/op
PutBenchmark.measureLockFreeMap avgt 100 0.491 ±0.002 ms/op

5.4.2. Put Benchmarking With Collisions

8 threads put 1000 elements concurrently. The hashCode was poorly implemented to produce high rate of collisions. This was repeated 100 times.

Following is the comparision of Average Response Time (avgt) using JMH benchmarking.

Benchmark Mode Cnt Score Error Units
PutWithCollisionsBenchmark.measureJavaMap avgt 100 122.056 ±1.222 ms/op
PutWithCollisionsBenchmark.measureLockFreeMap avgt 100 2.282 ± 0.011 ms/op

5.4.3. Get Benchmarking With Collisions

8 threads attempted to fetch elements 1 millions times concurrently.

Following is the comparision of Average Response Time (avgt) using JMH benchmarking.

Benchmark Mode Cnt Score Error Units
GetBenchmark.measureJDKMap1 avgt 100 0.474 ±0.044 ms/op
GetBenchmark.measureLockFreeMap1 avgt 100 0.419 ±0.001 ms/op

6. Singleton Reference

Singleton Reference is a concurrent, non-blocking container of singleton which implements Singleton pattern using Atomic classes.

Assuming HeavyObject is a user-defined class

    SingleRef<HeavyObject> singleRef = new SingleRef<>(HeavyObject::new);
    HeavyObject singletonObj = singleRef.get();