par - occam-style concurrency primitives

The par package provides functions that implement occam-style PAR and replicated-PAR control structures. These provide synchronization upon goroutine completion in the same way as idiomatic sync.WaitGroup usage.

The par.DO function calls some number of functions, concurrently, and waits for them to complete before it returns. The par.FOR function calls a single function a number of times defined by two integers. Each call occurs concurrently and as with par.DO, par.FOR only returns when all of its function calls complete.

par.DO mimics the occam PAR keyword and par.FOR the replicated-PAR construct (a concurrent for-loop). In Go the functions are implemented around sync.WaitGroup and hide the repetitive clutter of wait group manipulations.

An example

Imagine we have some functions that run loops to do some control operation. In our system we run these concurrently, perhaps they communicate but that's detail we can ignore for the time being. We run then concurrently and wait for them to finish.

par.DO(
ControlFuelRods,
MonitorCoolant,
MoveDials,
FlashLights,
ControlSirens,
func() {
par.FOR(0, 10, func(i int) {
MonitorDoor(i+1)
})
},
)

Hiding sync.WaitGroup

The par functions encapsulate the, now common, idiom of using a sync.WaitGroup to synchronize goroutine completion. The par functions offer no actual new functionality over what direct use of sync.WaitGroup affords, and actually provide less, but their use does make for cleaner code by hiding the implementation details of the synchronization. The functions eliminate clutter making the process structure more obvious and therefore more easily comprehended and maintained (i.e. not broken).

Abusing import

We can abuse Go's import . to let us use the package's functions without qualification. This makes them seem a little more like using a language construct.

Importing the package using,

import . "github.com/atrn/par"

lets us write,

DO(
ControlFuelRods,
MonitorCoolant,
MoveDials,
FlashLights,
ControlSirens,
func() {
FOR(0, 10, func(number int) {
MonitorDoor(number)
},
)

That looks okay, if you accept the namespace pollution, but DO() and FOR() are a little too generic and not that descriptive.

Synonyms, PAR and PAR_FOR

The package define synonyms for DO and FOR, PAR and PAR_FOR. Using these the code becomes,

PAR(
ControlFuelRods,
MonitorCoolant,
MoveDials,
FlashLights,
ControlSirens,
func() {
PAR_FOR(0, 10, func(number int) {
MonitorDoor(number)
},
)

Nested PARs

Each of the above examples shows nesting of PAR via the function literal calling par.FOR aka PAR_FOR. This pattern, a func() that just calls par.FOR is common, luckily Go lets us simplify it.

The package defines what it refers to as fn function (I never thought of a good name).

func FORfn(start, limit int, f func(int)) func()

The returned function calls par.FOR using the supplied arguments and is passed to par.DO as one of its functions to call concurrently. As with par.DO and par.FOR, par.FORfn has a synonym intended to be used via import . - PAR_FORfn.

Armed with PAR_FORfn we can write,

PAR(
ControlFuelRods,
MonitorCoolant,
MoveDials,
FlashLights,
ControlSirens,
PAR_FORfn(0, 10, func(number int) {
MonitorDoor(number)
}),
)

Classic fanout

jobs := make(chan Work)
results := make(chan Result)
par.DO(
func() {
for job := range Jobs() {
jobs <- job
}
close(work)
},
func() {
par.FOR(0, Nworkers, func(int) {
for job := range jobs {
results <- Process(job)
}
}
close(results)
},
func() {
for result := range results {
Consume(result)
}
},
)

Removing the explicit sync.WaitGroup use makes the process structure easier to comprehend (and may help stop the endless complaints about multiple channel closes).