tl;dr: Multiprocessing in Python is crippled by pickles poor function serialization. The more robust serialization package dill improves the situation. Dill-based solutions for both multiprocessing and IPython.parallel make distributed computing simple again.

To leverage the cores found in modern processors we need to communicate functions between different processes. I.e. if we have some function in one process

def do_expensive_computations(data):
    ...
    return fancy result

then we need to communicate that functionality and all functionality on which it depends to our other worker processes.

To communicate this function we translate it down into a blob of text, ship that text over a wire, and then retranslate that text back into a fully operational function. This process, called serialization, is like the teleporters in Star Trek; it takes an important thing (function or crew member) translates it into something manageable (text or bits) moves it quickly to some other location, and then reassembles it correctly (we hope!) Just as accidents happen in Star Trek it’s easy for function serialization to go awry.

Pickle

The standard serialization package in Python is pickle. The pickle package can serialize and deserialize most Python objects, not just functions.

In [1]: import pickle

In [2]: pickle.dumps({'Alice': 100})
Out[2]: "(dp0\nS'Alice'\np1\nI100\ns."

In [3]: pickle.loads("(dp0\nS'Alice'\np1\nI100\ns.")
Out[3]: {'Alice': 100}

How does Pickle go about serializing functions?

In [4]: from math import sin

In [5]: pickle.dumps(sin)
Out[5]: 'cmath\nsin\np0\n.'

Pickle specifies a function using its module name (see math on the left) and its function name (see sin in the middle). Sadly this approach fails for many cases. In particular pickle fails to serialize the following

  • Methods
  • Lambdas
  • Closures
  • Some functions defined interactively
In [2]: pickle.dumps(str.split)
TypeError: can't pickle method_descriptor objects

In [4]: pickle.dumps(lambda x: x**2)
PicklingError: Can't pickle <function <lambda> at 0x1172410>: it's not found as
__main__.<lambda>

Most large projects use at least one (often all) of these features. This makes multiprocessing a pain.

Multiprocessing

We care about function serialization because we want to send one function to many processes in order to leverage parallelism. The standard way to do this is with the multiprocessing module. One simple approach is with the Pool abstraction

In [12]: import multiprocessing as mp

In [13]: p = mp.Pool(4)  # Processing Pool with four processors

In [14]: p.map(sin, range(10))
Out[14]:
[0.0,
 0.8414709848078965,
 0.9092974268256817,
 0.1411200080598672,
-0.7568024953079282,
-0.9589242746631385,
-0.27941549819892586,
 0.6569865987187891,
 0.9893582466233818,
 0.4121184852417566]

But multiprocessing uses pickle and so inherits its limitations. Here it fails to serialize and broadcast a lambda square function.

In [15]: p.map(lambda x: x**2, range(10))
PicklingError: Can't pickle <type 'function'>: attribute lookup
__builtin__.function failed

I rarely see multiprocessing in the wild. I suspect that this is because poor function serialization makes it a pain for any but the most trivial of projects.

dill replaces pickle

The dill library is a drop-in alternative to pickle that can robustly handle function serialization.

In [8]: import dill

In [4]: dill.dumps(str.split)
Out[4]:
'cdill.dill\n_getattr\np0\n(cdill.dill\n_load_type\np1\n(S\'StringType\'\np2\ntp3\nRp4\nS\'split\'\np5\nS"<method\'split\' of \'str\' objects>"\np6\ntp7\nRp8\n.'

In [9]: dill.dumps(lambda x: x)
Out[9]:
'\x80\x02cdill.dill\n_load_type\nq\x00U\x0cFunctionTypeq\x01\x85q\x02Rq\x03(cdill.dill\n_unmarshal\nq\x04Usc\x01\x00\x00\x00\x01\x00\x00\x00\x01\x00\x00\x00C\x00\x00\x00s\x04\x00\x00\x00|\x00\x00S(\x01\x00\x00\x00N(\x00\x00\x00\x00(\x01\x00\x00\x00t\x01\x00\x00\x00x(\x00\x00\x00\x00(\x00\x00\x00\x00s\x1e\x00\x00\x00<ipython-input-9-70b342a16b4d>t\x08\x00\x00\x00<lambda>\x01\x00\x00\x00s\x00\x00\x00\x00q\x05\x85q\x06Rq\x07c__builtin__\n__main__\nU\x08<lambda>q\x08NNtq\tRq\n.'

As a result most of the speed-bumps of using multiprocessing should disappear.

Dill and Multiprocessing

The makers of dill apparently know this and so have developed their own fork of multiprocessing that uses dill. This resides in the pathos library

In [1]: import pathos.multiprocessing as mp

In [2]: p = mp.Pool(4)  # Processing Pool with four processors

In [3]: p.map(lambda x: x**2, range(10))
Out[3]: [0, 1, 4, 9, 16, 25, 36, 49, 64, 81]

Dill and IPython Parallel

You should know about IPython parallel.

The IPython notebook has gotten a lot of press recently. The notebook became possible after the project was restructured to separate computation and interaction. One important result is that we can now perform computation in a process while interacting in a web browser, giving rise to the ever-popular notebook.

This same computation-is-separate-from-interaction concept supports other innovations. In particular IPython parallel uses this to create a simple platform for both multiprocessing and distributed computing.

mrocklin@notebook:~$ ipcluster start --n=4
mrocklin@notebook:~$ ipython
In [1]: from IPython.parallel import Client

In [2]: p = Client()[:]

In [3]: p.map_sync(lambda x: x**2, range(10))
Out[3]: [0, 1, 4, 9, 16, 25, 36, 49, 64, 81]

Note that this system handles the lambda without failing. IPython performs some custom serializations on top of pickle. Unfortunately these customizations still don’t cover all use cases. Fortunately IPython provides hooks to specify your preferred serialization technique. Thanks to a recent change, IPython views now provide a convenient .use_dill method.

In [4]: p.use_dill()

In [5]: p.map_sync(str.split, ['Hello world!', 'foo bar'])
[['Hello', 'world!'], ['foo', 'bar']]

A more explicit treatment of switching IPython’s serializers to dill can be found in this notebook.

Acknowledgements

My interest into multiprocessing and serialization was originally spurred by a talk by Ian Langmore.

The dill project is developed by Mike Mckerns. Several people have pointed it out to me. These include @asmeurer, @themodernscientist, @twiecki, and @lidavidm.

Issues

Thanks to @mmckerns and @minrk for their recent interactions to resolve issues related to this topic.



blog comments powered by Disqus