ahfuzhang / cowmap

Copy-On-Write Map (CowMap), Optimized for small data environments with very large reads and very small writes.

Copy-On-Write Map (CowMap)

en

cn

There is a scenario where a map with a small amount of data is in use with very many reads and very few writes. In order to optimize for this scenario, I implemented a copy-on-write map. The principle is that all reads can be concurrently read without locking, and once a write is needed, a copy of the original map is made, modified on the backup, and then the pointer is switched to the new object via an atomic operation. I compared the performance of CowMap (Copy-On-Write Map) and sync.Map , and ordinary map + read/write locks.

有这样一种场景：数据量不多的map，在使用中读极多写极少。为了在这种场景下做极致的优化，我实现了 copy-on-write 的map: 其实现原理为：所有的读都可以不加锁的并发读取，一旦需要写，则 copy 一份原来的map，在备份上修改，然后通过原子操作把指针切换到新的对象上。 我对比了CowMap(Copy-On-Write Map) 和 sync.Map , 以及普通map + 读写锁三种方式的性能。

Read-write rate	Type	ns/op
99.99% : 0.01%	CowMap	4.649
99.99% : 0.01%	map + sync.RWMutex	187.5
99.99% : 0.01%	sync.Map	15.06
99.90% : 0.10%	CowMap	32.70
99.90% : 0.10%	map + sync.RWMutex	159.9
99.90% : 0.10%	sync.Map	14.08
99.00% : 1.00%	CowMap	303.6
99.00% : 1.00%	map + sync.RWMutex	105.7
99.00% : 1.00%	sync.Map	14.08

因此，当读的比例超过 99.99%时，CowMap 是 sync.Map 的 3.24 倍。是 map+sync.RWMutex 的 40.33 倍。 Thus, when reading more than 99.99%, CowMap is 3.24 times as large as sync.Map. 40.33 times more than map+sync.RWMutex.

How to use

go get github.com/ahfuzhang/cowmap@latest

import (
    "github.com/ahfuzhang/cowmap"
)

func main(){
    var m cowmap.CowMap[uint64, uint64]
	m.Set(1, 2)
    m.Set(3, 4)
    value, has := m.Get(1)
    fmt.Println(value, has)
    m.Delete(3)
}