ZenQ

A low-latency thread-safe queue in golang implemented using a lock-free ringbuffer and runtime internals

Based on the LMAX Disruptor Pattern

Features

• Much faster than native channels in both SPSC (single-producer-single-consumer) and MPSC (multi-producer-single-consumer) modes in terms of time/op
• More resource efficient in terms of memory_allocation/op and num_allocations/op evident while benchmarking large batch size inputs
• Handles the case where NUM_WRITER_GOROUTINES > NUM_CPU_CORES much better than native channels
• Selection from multiple ZenQs just like golang's select{} ensuring fair selection and no starvation
• Closing a ZenQ

Benchmarks to support the above claims here

Installation

You need Golang 1.19.x or above

$ go get github.com/alphadose/zenq/v2

Usage

1. Simple Read/Write

package main

import (
	"fmt"

	"github.com/alphadose/zenq/v2"
)

type payload struct {
	alpha int
	beta  string
}

func main() {
	zq := zenq.New[payload](10)

	for j := 0; j < 5; j++ {
		go func() {
			for i := 0; i < 20; i++ {
				zq.Write(payload{
					alpha: i,
					beta:  fmt.Sprint(i),
				})
			}
		}()
	}

	for i := 0; i < 100; i++ {
		if data, queueOpen := zq.Read(); queueOpen {
			fmt.Printf("%+v\n", data)
		}
	}
}

2. Selection from multiple ZenQs just like golang's native select{}. The selection process is fair i.e no single ZenQ gets starved

package main

import (
	"fmt"

	"github.com/alphadose/zenq/v2"
)

type custom1 struct {
	alpha int
	beta  string
}

type custom2 struct {
	gamma int
}

const size = 100

var (
	zq1 = zenq.New[int](size)
	zq2 = zenq.New[string](size)
	zq3 = zenq.New[custom1](size)
	zq4 = zenq.New[*custom2](size)
)

func main() {
	go looper(intProducer)
	go looper(stringProducer)
	go looper(custom1Producer)
	go looper(custom2Producer)

	for i := 0; i < 40; i++ {

		// Selection occurs here
		if data := zenq.Select(zq1, zq2, zq3, zq4); data != nil {
			switch data.(type) {
			case int:
				fmt.Printf("Received int %d\n", data)
			case string:
				fmt.Printf("Received string %s\n", data)
			case custom1:
				fmt.Printf("Received custom data type number 1 %#v\n", data)
			case *custom2:
				fmt.Printf("Received pointer %#v\n", data)
			}
		}
	}
}

func intProducer(ctr int) { zq1.Write(ctr) }

func stringProducer(ctr int) { zq2.Write(fmt.Sprint(ctr * 10)) }

func custom1Producer(ctr int) { zq3.Write(custom1{alpha: ctr, beta: fmt.Sprint(ctr)}) }

func custom2Producer(ctr int) { zq4.Write(&custom2{gamma: 1 << ctr}) }

func looper(producer func(ctr int)) {
	for i := 0; i < 10; i++ {
		producer(i)
	}
}

Benchmarks

Benchmarking code available here

Note that if you run the benchmarks with --race flag then ZenQ will perform slower because the --race flag slows down the atomic operations in golang. Under normal circumstances, ZenQ will outperform golang native channels.

Hardware Specs

❯ neofetch
                    'c.          alphadose@ReiEki.local
                 ,xNMM.          ----------------------
               .OMMMMo           OS: macOS 12.3 21E230 arm64
               OMMM0,            Host: MacBookAir10,1
     .;loddo:' loolloddol;.      Kernel: 21.4.0
   cKMMMMMMMMMMNWMMMMMMMMMM0:    Uptime: 6 hours, 41 mins
 .KMMMMMMMMMMMMMMMMMMMMMMMWd.    Packages: 86 (brew)
 XMMMMMMMMMMMMMMMMMMMMMMMX.      Shell: zsh 5.8
;MMMMMMMMMMMMMMMMMMMMMMMM:       Resolution: 1440x900
:MMMMMMMMMMMMMMMMMMMMMMMM:       DE: Aqua
.MMMMMMMMMMMMMMMMMMMMMMMMX.      WM: Rectangle
 kMMMMMMMMMMMMMMMMMMMMMMMMWd.    Terminal: iTerm2
 .XMMMMMMMMMMMMMMMMMMMMMMMMMMk   Terminal Font: FiraCodeNerdFontComplete-Medium 16 (normal)
  .XMMMMMMMMMMMMMMMMMMMMMMMMK.   CPU: Apple M1
    kMMMMMMMMMMMMMMMMMMMMMMd     GPU: Apple M1
     ;KMMMMMMMWXXWMMMMMMMk.      Memory: 1370MiB / 8192MiB
       .cooc,.    .,coo:.

Terminology

• NUM_WRITERS -> The number of goroutines concurrently writing to ZenQ/Channel
• INPUT_SIZE -> The number of input payloads to be passed through ZenQ/Channel from producers to consumer

Computed from benchstat of 30 benchmarks each via go test -benchmem -bench=. benchmarks/simple/*.go

name                                     time/op
_Chan_NumWriters1_InputSize600-8          23.2µs ± 1%
_ZenQ_NumWriters1_InputSize600-8          17.9µs ± 1%
_Chan_NumWriters3_InputSize60000-8        5.27ms ± 3%
_ZenQ_NumWriters3_InputSize60000-8        2.36ms ± 2%
_Chan_NumWriters8_InputSize6000000-8       671ms ± 2%
_ZenQ_NumWriters8_InputSize6000000-8       234ms ± 6%
_Chan_NumWriters100_InputSize6000000-8     1.59s ± 4%
_ZenQ_NumWriters100_InputSize6000000-8     309ms ± 2%
_Chan_NumWriters1000_InputSize7000000-8    1.97s ± 0%
_ZenQ_NumWriters1000_InputSize7000000-8    389ms ± 4%
_Chan_Million_Blocking_Writers-8           10.4s ± 2%
_ZenQ_Million_Blocking_Writers-8           2.32s ±21%

name                                     alloc/op
_Chan_NumWriters1_InputSize600-8           0.00B
_ZenQ_NumWriters1_InputSize600-8           0.00B
_Chan_NumWriters3_InputSize60000-8          109B ±68%
_ZenQ_NumWriters3_InputSize60000-8        24.6B ±107%
_Chan_NumWriters8_InputSize6000000-8       802B ±241%
_ZenQ_NumWriters8_InputSize6000000-8     1.18kB ±100%
_Chan_NumWriters100_InputSize6000000-8    44.2kB ±41%
_ZenQ_NumWriters100_InputSize6000000-8    10.7kB ±38%
_Chan_NumWriters1000_InputSize7000000-8    476kB ± 8%
_ZenQ_NumWriters1000_InputSize7000000-8   90.6kB ±10%
_Chan_Million_Blocking_Writers-8           553MB ± 0%
_ZenQ_Million_Blocking_Writers-8           122MB ± 3%

name                                     allocs/op
_Chan_NumWriters1_InputSize600-8            0.00
_ZenQ_NumWriters1_InputSize600-8            0.00
_Chan_NumWriters3_InputSize60000-8          0.00
_ZenQ_NumWriters3_InputSize60000-8          0.00
_Chan_NumWriters8_InputSize6000000-8       2.76 ±190%
_ZenQ_NumWriters8_InputSize6000000-8        5.47 ±83%
_Chan_NumWriters100_InputSize6000000-8       159 ±26%
_ZenQ_NumWriters100_InputSize6000000-8      25.1 ±39%
_Chan_NumWriters1000_InputSize7000000-8    1.76k ± 6%
_ZenQ_NumWriters1000_InputSize7000000-8     47.3 ±31%
_Chan_Million_Blocking_Writers-8           2.00M ± 0%
_ZenQ_Million_Blocking_Writers-8           1.00M ± 0%

The above results show that ZenQ is more efficient than channels in all 3 metrics i.e time/op, mem_alloc/op and num_allocs/op for the following tested cases:-

1. SPSC
2. MPSC with NUM_WRITER_GOROUTINES < NUM_CPU_CORES
3. MPSC with NUM_WRITER_GOROUTINES > NUM_CPU_CORES

Cherry on the Cake

In SPSC mode ZenQ is faster than channels by 92 seconds in case of input size of 6 * 10⁸ elements

❯ go run benchmarks/simple/main.go

With Input Batch Size: 60 and Num Concurrent Writers: 1

Native Channel Runner completed transfer in: 26.916µs
ZenQ Runner completed transfer in: 20.292µs
====================================================================

With Input Batch Size: 600 and Num Concurrent Writers: 1

Native Channel Runner completed transfer in: 135.75µs
ZenQ Runner completed transfer in: 105.792µs
====================================================================

With Input Batch Size: 6000 and Num Concurrent Writers: 1

Native Channel Runner completed transfer in: 2.100209ms
ZenQ Runner completed transfer in: 510.792µs
====================================================================

With Input Batch Size: 6000000 and Num Concurrent Writers: 1

Native Channel Runner completed transfer in: 1.241481917s
ZenQ Runner completed transfer in: 226.068209ms
====================================================================

With Input Batch Size: 600000000 and Num Concurrent Writers: 1

Native Channel Runner completed transfer in: 1m55.074638875s
ZenQ Runner completed transfer in: 22.582667917s
====================================================================