Vindaar / forked

Syntax sugar for `cligen`'s procpool for easy multiprocessing

forked - Sugar for cligen/procpool

For quite a now while cligen has a ~procpool~ submodule. It is basically a module similar to Python’s multiproccesing module. Under the hood it uses fork() to produce child processes.

In many use cases the memory overhead and startup overhead compared to threads is completely negligible. In these cases using multiprocessing is incredibly nice, because one side-steps all memory safety and race condition problems. Until now, the API of procpool may have been a bit too low level, but with forked this changes.

forked is a for loop macro, which takes the iterator / collection of the loop and runs the loop body in parallel on a pool of processes. By default the number of logical CPU cores is used, but this can be overwritten either by an environment variable PP_JOBS or explicitly via the macro call (the latter takes precedent over the former).

Basic syntax overview

The basic syntax looks as follows:

var res = 0.0
for x in forked(iter):
  let y = doWork() # arbitrary loop body
  send y           # mandatory `send`
  join: res += y   # mandatory `join`

Here iter is any iterator or collection. Anything that would normally work as a for loop without forked. The loop body (in the example just let y = doWork()) can be any number of lines and can contain any Nim code. There are no restrictions and feel free to access any global variables and even mutate them. Just be aware that the advantage of not having to worry about race conditions and thread safety, also means mutated globals won’t persist the child processes, nor will any child see any other child’s mutation (because of course each process has its own memory space).

As you can see, there are 2 mandatory statements as part of the loop. A send statement, which receives a variable or expression. This is the data you wish to send back to the parent process.

Finally, the join statement is the code which describes how the main process consumes the data sent by each child. Here you will likely want to process the data you sent in some way.

See the following two examples for some further ideas:

# 1. Port of https://github.com/c-blake/cligen/blob/master/examples/piPar.nim
import forked
import flatBuffers # for `save/loadBuffer`, see below
proc term(k: int): float = (0.5 - float(k mod 2))*8'f/float(2*k + 1)

iterator batches(n, batch: int): Slice[int] =
  for start in countup(0, n - 1, batch):
    yield start ..< start+batch

proc piProcPool(n: int): float =
  for s in forked(batches(n, 5)):
    var res = 0.0
    for k in s:
      res += term(k)
    send res
    join: result += res

echo piProcPool(100) # run in parallel!

# 2. Example sending non trivial data (i.e. not mem-copyable) data from
#    child to parent. It just works 'magically' for you.
type
  Foo = object
    s: string
    x: float

proc calcFoo(x: int): Foo = Foo(s: $x, x: x.float)
proc testIt(): seq[Foo] =
  for x in forked(0 ..< 10):
    let f = calcFoo(x)
    send f
    join: result.add f

echo testIt() # run in parallel!

Iteration index

You can also explicitly name the iteration index variable by writing one of the following:

for i, x in forked(...)
for (i, x) in forked(...)  

This allows you to explicitly access the iteration index (and thus job index) in the loop body.

IMPORTANT: This also does allow you to access e.g. the i-th element of e.g. a sequence inside of the join portion!

For example the second example from above could also be:

type
  Foo = object
    s: string
    x: float

proc calcFoo(x: int): Foo = Foo(s: $x, x: x.float)
proc testIt(): seq[Foo] =
  result = newSeq[Foo](10)
  for i, x in forked(0 ..< 10):
    let f = calcFoo(x)
    send f
    join: result[i] = f

echo testIt()

This can be useful if you need the result in a sorted order.

Details about data transfer child ⇒ parent process

The biggest convenience feature of forked over using the raw procpool is ‘automagic’ transferring data from child to parent. For any data that can be mem-copied, we simply send the binary data through the file descriptor. However, if the data we want to send contains non mem-copyable data like strings, (nested) sequences or ref objects, this needs to be handled with manual care in procpool. While usually not an issue, it makes it less of a drop-in replacement than one sometimes might want.

The way we get around it is to save the data we send back to an in-memory file (using /dev/shm) and only sending the filename back to the parent, which loads the data back.

For this to work, the forked macro code utilizes a saveBuffer and loadBuffer procedure. In principle any ‘serialization’ library can be used. This works, because the forked macro only produces code, which emits the function calls without binding them.

The required API of these two functions is:

proc saveBuffer*[T](x: T, fname: string)
proc loadBuffer*[T](fname: string, deleteFile = false): T 

This is already implemented in flatBuffers, but feel free to e.g. use flatty and simply write the two procedures as wrappers around the equivalent flatty functionality (fromFlatty / toFlatty and likely calls to read / write to disk).

The deleteFile argument is used to remove the in-memory files after they have been read. If you write your own loadBuffer you should make sure to memory map the file instead of using readFile to avoid unnecessary overhead.

NOTE: If you always wish to use flatBuffers, you can compile with -d:UseFlatBuffers to import and reexport the required logic from it. If you put this into your ~/.config/nim/nim.cfg you won’t have to write it all the time. :)

Customizing behavior

There are 3 compile time variables, which can be changed to alter data transfer behavior.

These are:

## Decides if we send data from child processes back to the parent by producing (in memory)
## files using `/dev/shm` (or real files if `BasePath` is changed)
const WriteFiles* {.booldefine.} = true
## Decides if the binary data files are deleted upon being memory mapped by the parent
const DeleteFiles* {.booldefine.} = true
## Can be used to change the default path where binary files are stored
const BasePath* {.strdefine.} = "/dev/shm/pp_forked/"

(Bool and string define variables are adjusted by compiling with e.g. -d:WriteFiles=false or -d:BasePath=/tmp/)

In addition as mentioned above, the number of jobs can be adjusted at runtime by defining a PP_JOBS environment variable. For example to adjust the number of jobs to 8:

PP_JOBS=8 ./program_using_forked

(or define it beforehand using export PP_JOBS=8 of course)

Manually overwriting number of jobs directly

Alternatively you can also pass the number of jobs to forked directly:

let jobs = 8
for x in forked(iter, jobs):
  ...

to run the loop with only 8 jobs for example. This takes precedence over any PP_JOBS definition, so make sure you only define it if you want to fix it to some other constant or runtime value.

Note on code location

This module might very well become a submodule of cligen itself in the near future. For the time being it is separate until we have decided how to handle dependencies for data communication.