Software Transactional Memory in Go

#go #stm #concurrency

I have recently written decillion/go-stm, which is a software transactional memory (STM) implementation for Go. It provides a new way of synchronization to Go programmers, while it is still experimental one.

The package enables one to atomically execute a sequence of operations on shared variables without explicit locking. I present a simple example below. Please see the repository or a blog post (in Japanese) for further information.


// Initialize two transactional variables.
x := New(0)
y := New(0)

// Define a transaction that transfer 100 from x to y.
transfer := func(rec *TRec) interface{} {
    currX := rec.Load(x).(int) // Read the value of x.
    currY := rec.Load(y).(int) // Read the value of y.
    rec.Store(x, currX-100)    // Write currX-100 to x.
    rec.Store(y, currY+100)    // Write currY-100 to y.
    return nil
}

// Execute the transaction atomically.
Atomically(transfer)

If you want to know more about the STM itself, I recommend you to read the article, Beautiful concurrency, written by Simon Peyton Jones.

There has been another STM package written by lukechampine. It provides a richer STM interface but is less performant than the present package. Here is the result of a very simple benchmark, in which two transactional variables are atomically incremented or read. The benchmark is taken at Ubuntu 17.10 on a DigitalOcean's High CPU Droplet with 32 cores.

Benchmark_Read90Write10_decillion-4         10000000           156 ns/op
Benchmark_Read90Write10_decillion-8         10000000           144 ns/op
Benchmark_Read90Write10_decillion-16        10000000           214 ns/op
Benchmark_Read90Write10_decillion-32         5000000           289 ns/op

Benchmark_Read90Write10_lukechampine-4       2000000           761 ns/op
Benchmark_Read90Write10_lukechampine-8       2000000           822 ns/op
Benchmark_Read90Write10_lukechampine-16      2000000           912 ns/op
Benchmark_Read90Write10_lukechampine-32      2000000           966 ns/op