Concurrency
Share by communicating
Concurrent programming is a large topic and there is space only for some Go-specific highlights here.
Concurrent programming in many environments is made difficult by the subtleties required to implement correct access to shared variables. Go encourages a different approach in which shared values are passed around on channels and, in fact, never actively shared by separate threads of execution. Only one goroutine has access to the value at any given time. Data races cannot occur, by design. To encourage this way of thinking we have reduced it to a slogan:
Do not communicate by sharing memory; instead, share memory by communicating.
This approach can be taken too far. Reference counts may be best done by putting a mutex around an integer variable, for instance. But as a high-level approach, using channels to control access makes it easier to write clear, correct programs.
One way to think about this model is to consider a typical single-threaded program running on one CPU. It has no need for synchronization primitives. Now run another such instance; it too needs no synchronization. Now let those two communicate; if the communication is the synchronizer, there’s still no need for other synchronization. Unix pipelines, for example, fit this model perfectly. Although Go’s approach to concurrency originates in Hoare’s Communicating Sequential Processes (CSP), it can also be seen as a type-safe generalization of Unix pipes.
Goroutines
They’re called goroutines because the existing terms—threads, coroutines, processes, and so on—convey inaccurate connotations. A goroutine has a simple model: it is a function executing concurrently with other goroutines in the same address space. It is lightweight, costing little more than the allocation of stack space. And the stacks start small, so they are cheap, and grow by allocating (and freeing) heap storage as required.
Goroutines are multiplexed onto multiple OS threads so if one should block, such as while waiting for I/O, others continue to run. Their design hides many of the complexities of thread creation and management.
Prefix a function or method call with the go keyword to run the call in a new goroutine. When the call completes, the goroutine exits, silently. (The effect is similar to the Unix shell’s & notation for running a command in the background.)
go list.Sort() // run list.Sort concurrently; don't wait for it.
A function literal can be handy in a goroutine invocation.
func Announce(message string, delay time.Duration) {
go func() {
time.Sleep(delay)
fmt.Println(message)
}() // Note the parentheses - must call the function.
}
In Go, function literals are closures: the implementation makes sure the variables referred to by the function survive as long as they are active.
These examples aren’t too practical because the functions have no way of signaling completion. For that, we need channels.
Channels
Like maps, channels are allocated with make, and the resulting value acts as a reference to an underlying data structure. If an optional integer parameter is provided, it sets the buffer size for the channel. The default is zero, for an unbuffered or synchronous channel.
ci := make(chan int) // unbuffered channel of integers
cj := make(chan int, 0) // unbuffered channel of integers
cs := make(chan *os.File, 100) // buffered channel of pointers to Files
Unbuffered channels combine communication—the exchange of a value—with synchronization—guaranteeing that two calculations (goroutines) are in a known state.
There are lots of nice idioms using channels. Here’s one to get us started. In the previous section we launched a sort in the background. A channel can allow the launching goroutine to wait for the sort to complete.
c := make(chan int) // Allocate a channel.
// Start the sort in a goroutine; when it completes, signal on the channel.
go func() {
list.Sort()
c <- 1 // Send a signal; value does not matter.
}()
doSomethingForAWhile()
<-c // Wait for sort to finish; discard sent value.
Receivers always block until there is data to receive. If the channel is unbuffered, the sender blocks until the receiver has received the value. If the channel has a buffer, the sender blocks only until the value has been copied to the buffer; if the buffer is full, this means waiting until some receiver has retrieved a value.
A buffered channel can be used like a semaphore, for instance to limit throughput. In this example, incoming requests are passed to handle, which sends a value into the channel, processes the request, and then receives a value from the channel to ready the “semaphore” for the next consumer. The capacity of the channel buffer limits the number of simultaneous calls to process.
var sem = make(chan int, MaxOutstanding)
func handle(r *Request) {
sem <- 1 // Wait for active queue to drain.
process(r) // May take a long time.
<-sem // Done; enable next request to run.
}
func Serve(queue chan *Request) {
for {
req := <-queue
go handle(req) // Don't wait for handle to finish.
}
}
Once MaxOutstanding handlers are executing process, any more will block trying to send into the filled channel buffer, until one of the existing handlers finishes and receives from the buffer.
This design has a problem, though: Serve creates a new goroutine for every incoming request, even though only MaxOutstanding of them can run at any moment. As a result, the program can consume unlimited resources if the requests come in too fast. We can address that deficiency by changing Serve to gate the creation of the goroutines. Here’s an obvious solution, but beware it has a bug we’ll fix subsequently:
func Serve(queue chan *Request) {
for req := range queue {
sem <- 1
go func() {
process(req) // Buggy; see explanation below.
<-sem
}()
}
}
The bug is that in a Go for loop, the loop variable is reused for each iteration, so the req variable is shared across all goroutines. That’s not what we want. We need to make sure that req is unique for each goroutine. Here’s one way to do that, passing the value of req as an argument to the closure in the goroutine:
func Serve(queue chan *Request) {
for req := range queue {
sem <- 1
go func(req *Request) {
process(req)
<-sem
}(req)
}
}
Compare this version with the previous to see the difference in how the closure is declared and run. Another solution is just to create a new variable with the same name, as in this example:
func Serve(queue chan *Request) {
for req := range queue {
req := req // Create new instance of req for the goroutine.
sem <- 1
go func() {
process(req)
<-sem
}()
}
}
It may seem odd to write
req := req
but it’s legal and idiomatic in Go to do this. You get a fresh version of the variable with the same name, deliberately shadowing the loop variable locally but unique to each goroutine.
Going back to the general problem of writing the server, another approach that manages resources well is to start a fixed number of handle goroutines all reading from the request channel. The number of goroutines limits the number of simultaneous calls to process. This Serve function also accepts a channel on which it will be told to exit; after launching the goroutines it blocks receiving from that channel.
func handle(queue chan *Request) {
for r := range queue {
process(r)
}
}
func Serve(clientRequests chan *Request, quit chan bool) {
// Start handlers
for i := 0; i < MaxOutstanding; i++ {
go handle(clientRequests)
}
<-quit // Wait to be told to exit.
}
Channels of channels
One of the most important properties of Go is that a channel is a first-class value that can be allocated and passed around like any other. A common use of this property is to implement safe, parallel demultiplexing.
In the example in the previous section, handle was an idealized handler for a request but we didn’t define the type it was handling. If that type includes a channel on which to reply, each client can provide its own path for the answer. Here’s a schematic definition of type Request.
type Request struct {
args []int
f func([]int) int
resultChan chan int
}
The client provides a function and its arguments, as well as a channel inside the request object on which to receive the answer.
func sum(a []int) (s int) {
for _, v := range a {
s += v
}
return
}
request := &Request{[]int{3, 4, 5}, sum, make(chan int)}
// Send request
clientRequests <- request
// Wait for response.
fmt.Printf("answer: %d\n", <-request.resultChan)
On the server side, the handler function is the only thing that changes.
func handle(queue chan *Request) {
for req := range queue {
req.resultChan <- req.f(req.args)
}
}
There’s clearly a lot more to do to make it realistic, but this code is a framework for a rate-limited, parallel, non-blocking RPC system, and there’s not a mutex in sight.