lockfree

Golang lock-free concurrent Hashmap

---

Table of Contents

• Overview
• Hashmap
• Queue
• Stack
• Benchmark

Overview

Golang's native data structures (such as map, List) are not designed to be thread-safe at first place. A common solution is to use mutex to synchronize access to data that are shared by multiple threads. This is however not very efficient due to multiple threads contending a single read/write mutex and in general will lead to performance loss in timing.

A more efficient alternative is to leverage the CAS technique to remove (or significantly reduce) the use of mutex. Several data structures are implemented using this idea.

Benchmark results show that the CAS solution is generally 2~3x times faster as compared to solution using mutex.

For technical details of the implementation, click here for an explanation of the concurrent hashmap.

Hashmap

• can be concurrently accessed
• allows different key types in the same map

package anyname

import "github.com/dustinxie/lockfree"

func main() {
    m := lockfree.NewHashMap()

    // set
    m.Set(1, "one")
    m.Set("one", 1)
    size := m.Len() // size = 2

    // get
    s, ok := m.Get(1)     // s.(string) = "one", ok = true
    i, ok := m.Get("one") // i.(int) = 1, ok = true

    // delete
    m.Del(1)
    m.Del("one")
    size = m.Len() // map is empty, size = 0

    // can have multiple threads/go-routines call m.Set/Get/Del
}

BucketSizeOption

You can specify a preferred average bucket size when creating the map, a smaller size gives faster access speed with more memory. While a larger size costs less memory but gives slower access speed. See the benchmark section below for details.

import (
    "github.com/dustinxie/lockfree"
    "github.com/dustinxie/lockfree/hashmap"
)

func main() {
    m := lockfree.NewHashMap(hashmap.BucketSizeOption(16))
}

The default bucket size is 24 if a BucketSizeOption is not set.

for k, v := range

Since this is a concurrent hashmap, you'll need to call Lock() before doing a range operation. And remember to call Unlock() afterwards.

func rangeMap(m HashMap) error {
    f := func(interface{}, interface{}) error {
        // define what to do for each entry in the map
        return
    }

    m.Lock()
    for k, v, ok := m.Next(); ok; k, v, ok = m.Next() {
        if err := f(k, v); err != nil {
            // unlock the map before return, otherwise it will deadlock
            m.Unlock()
            return err
        } 
    }
    m.Unlock()
    return nil
}

Queue

• FIFO list that can be concurrently accessed
• can put different data types into the queue

package anyname

import "github.com/dustinxie/lockfree"

func main() {
    q := lockfree.NewQueue()

    // add to the queue
    q.Enque(1)
    q.Enque("one")
    size := q.Len() // size = 2

    // remove from queue
    i := q.Deque() // i.(int) = 1
    s := q.Deque() // s.(string) = "one"

    size = q.Len()  // queue is empty, size = 0
    s = q.Deque() // queue is empty, s = nil

    // can have multiple threads/go-routines call q.Enque/Deque
}

Stack

• LIFO list that can be concurrently accessed
• can put different data types into the stack

package anyname

import "github.com/dustinxie/lockfree"

func main() {
    s := lockfree.NewStack()

    // add to the stack
    s.Push(1)
    s.Push("one")
    size := s.Len() // size = 2

    // peek the top of stack
    str := s.Pop() // str.(string) = "one"

    // remove from stack
    str = s.Pop() // str.(string) = "one"
    i := s.Pop()   // i.(int) = 1

    size = s.Len() // queue is empty, size = 0
    str = s.Pop()  // queue is empty, s = nil

    // can have multiple threads/go-routines call s.Push/Pop/Peek
}

Benchmark

The benchmark program starts 10 go-routines, each would perform a certain set of tasks concurrently. Tests are run on a machine with following config:

Hardware	Spec
CPU	Intel Core i7-4770HQ @ 2.20GHz
Core	4
L2 cache (per Core)	256kB
L3 cache	6MB
RAM	16GB 1600MHz DDR3 RAM
Storage	256GB SSD

Benchmark Hashmap

Task set for each concurrent thread is to Set() 10,000 keys, Get() these 10,000 keys, and finally Del() all these keys.

The total workload (10 threads) amounts to 100,000 keys, and 300,000 map operations.

Here's the timing and memory allocation data comparison between the lockfree map, sync.Map and golang's map + RWMutex:

BenchmarkLockfreeHashMap-8      27       42189873 ns/op      9935506 B/op     703336 allocs/op
BenchmarkMapAndRWMutex-8         9      115930416 ns/op      5752789 B/op       3916 allocs/op
BenchmarkSyncMap-8               8      127551298 ns/op     13686902 B/op     503735 allocs/op

with bucket size = 16 (faster time, more memory)

BenchmarkLockfreeHashMap-8      27       41867046 ns/op     10279584 B/op     707433 allocs/op

Couple of observations: 1. The lockfree hashmap is 3x as fast as sync.Map, and costs 37% less memory 2. The decrease in time (and increase in memory) of using bucket size 16 vs. 24 is very minimal, less than 4% 3. Golang's native map + RWMutex to synchronize access is even slightly faster than sync.Map, and costs least amount of memory

Benchmark Queue

Task set for each concurrent thread is to Enque() 10,000 items, then Deque() these 10,000 items.

The total workload (10 threads) amounts to 100,000 items, and 200,000 queue operations.

Here's the timing and memory allocation data comparison between the lockfree queue and golang's list + RWMutex:

BenchmarkLockfreeQueue-8          42      24980347 ns/op      3999350 B/op     299750 allocs/op
BenchmarkQueueAndRWMutex-8        27      43929601 ns/op      5599024 B/op     199750 allocs/op

The lockfree queue has better performance in both timing and memory.

Benchmark Stack

Task set for each concurrent thread is to Push() 10,000 items, then Pop() these 10,000 items.

The total workload (10 threads) amounts to 100,000 items, and 200,000 stack operations.

Here's the timing and memory allocation data comparison between the lockfree stack and golang's list + RWMutex:

BenchmarkLockfreeStack-8          34      31400698 ns/op      3999327 B/op     299751 allocs/op
BenchmarkStackAndRWMutex-8        28      43869583 ns/op      5598690 B/op     199750 allocs/op

The lockfree stack has better performance in both timing and memory.