deciding among subprocess, multiprocessing, and thread in Python?

Question

I'd like to parallelize my Python program so that it can make use of multiple processors on the machine that it runs on. My parallelization is very simple, in that all the parallel "threads" of the program are independent and write their output to separate files. I don't need the threads to exchange information but it is imperative that I know when the threads finish since some steps of my pipeline depend on their output.

Portability is important, in that I'd like this to run on any Python version on Mac, Linux, and Windows. Given these constraints, which is the most appropriate Python module for implementing this? I am trying to decide between thread, subprocess, and multiprocessing, which all seem to provide related functionality.

Any thoughts on this? I'd like the simplest solution that's portable.

Answer 1

For me this is actually pretty simple:

The subprocess option:

subprocess is for running other executables --- it's basically a wrapper around os.fork() and os.execve() with some support for optional plumbing (setting up PIPEs to and from the subprocesses. Obviously you could other inter-process communications (IPC) mechanisms, such as sockets, or Posix or SysV shared memory. But you're going to be limited to whatever interfaces and IPC channels are supported by the programs you're calling.

Commonly, one uses any subprocess synchronously --- simply calling some external utility and reading back its output or awaiting its completion (perhaps reading its results from a temporary file, or after it's posted them to some database).

However one can spawn hundreds of subprocesses and poll them. My own personal favorite utility classh does exactly that. The biggest disadvantage of the subprocess module is that I/O support is generally blocking. There is a draft PEP-3145 to fix that in some future version of Python 3.x and an alternative asyncproc (Warning that leads right to the download, not to any sort of documentation nor README). I've also found that it's relatively easy to just import fcntl and manipulate your Popen PIPE file descriptors directly --- though I don't know if this is portable to non-UNIX platforms.

(Update: 7 August 2019: Python 3 support for ayncio subprocesses: asyncio Subprocessses)

subprocess has almost no event handling support ... though you can use the signal module and plain old-school UNIX/Linux signals --- killing your processes softly, as it were.

The multiprocessing option:

multiprocessing is for running functions within your existing (Python) code with support for more flexible communications among this family of processes. In particular it's best to build your multiprocessing IPC around the module's Queue objects where possible, but you can also use Event objects and various other features (some of which are, presumably, built around mmap support on the platforms where that support is sufficient).

Python's multiprocessing module is intended to provide interfaces and features which are very similar to threading while allowing CPython to scale your processing among multiple CPUs/cores despite the GIL (Global Interpreter Lock). It leverages all the fine-grained SMP locking and coherency effort that was done by developers of your OS kernel.

The threading option:

threading is for a fairly narrow range of applications which are I/O bound (don't need to scale across multiple CPU cores) and which benefit from the extremely low latency and switching overhead of thread switching (with shared core memory) vs. process/context switching. On Linux this is almost the empty set (Linux process switch times are extremely close to its thread-switches).

threading suffers from two major disadvantages in Python.

One, of course, is implementation specific --- mostly affecting CPython. That's the GIL. For the most part, most CPython programs will not benefit from the availability of more than two CPUs (cores) and often performance will suffer from the GIL locking contention.

The larger issue which is not implementation specific, is that threads share the same memory, signal handlers, file descriptors and certain other OS resources. Thus the programmer must be extremely careful about object locking, exception handling and other aspects of their code which are both subtle and which can kill, stall, or deadlock the entire process (suite of threads).

By comparison the multiprocessing model gives each process its own memory, file descriptors, etc. A crash or unhandled exception in any one of them will only kill that resource and robustly handling the disappearance of a child or sibling process can be considerably easier than debugging, isolating and fixing or working around similar issues in threads.

• (Note: use of threading with major Python systems, such as NumPy, may suffer considerably less from GIL contention than most of your own Python code would. That's because they've been specifically engineered to do so; the native/binary portions of NumPy, for example, will release the GIL when that's safe).

The twisted option:

It's also worth noting that Twisted offers yet another alternative which is both elegant and very challenging to understand. Basically, at the risk of over simplifying to the point where fans of Twisted may storm my home with pitchforks and torches, Twisted provides event-driven co-operative multi-tasking within any (single) process.

To understand how this is possible one should read about the features of select() (which can be built around the select() or poll() or similar OS system calls). Basically it's all driven by the ability to make a request of the OS to sleep pending any activity on a list of file descriptors or some timeout.

Awakening from each of these calls to select() is an event --- either one involving input available (readable) on some number of sockets or file descriptors, or buffering space becoming available on some other (writable) descriptors or sockets, some exceptional conditions (TCP out-of-band PUSH'd packets, for example), or a TIMEOUT.

Thus the Twisted programming model is built around handling these events then looping on the resulting "main" handler, allowing it to dispatch the events to your handlers.

I personally think of the name, Twisted as evocative of the programming model ... since your approach to the problem must be, in some sense, "twisted" inside out. Rather than conceiving of your program as a series of operations on input data and outputs or results, you're writing your program as a service or daemon and defining how it reacts to various events. (In fact the core "main loop" of a Twisted program is (usually? always?) a reactor()).

The major challenges to using Twisted involve twisting your mind around the event driven model and also eschewing the use of any class libraries or toolkits which are not written to co-operate within the Twisted framework. This is why Twisted supplies its own modules for SSH protocol handling, for curses, and its own subprocess/Popen functions, and many other modules and protocol handlers which, at first blush, would seem to duplicate things in the Python standard libraries.

I think it's useful to understand Twisted on a conceptual level even if you never intend to use it. It may give insights into performance, contention, and event handling in your threading, multiprocessing and even subprocess handling as well as any distributed processing you undertake.

(Note: Newer versions of Python 3.x are including asyncio (asynchronous I/O) features such as async def, the @async.coroutine decorator, and the await keyword, and yield from future support. All of these are roughly similar to Twisted from a process (co-operative multitasking) perspective). (For the current status of Twisted support for Python 3, check out: https://twistedmatrix.com/documents/current/core/howto/python3.html)

The distributed option:

Yet another realm of processing you haven't asked about, but which is worth considering, is that of distributed processing. There are many Python tools and frameworks for distributed processing and parallel computation. Personally I think the easiest to use is one which is least often considered to be in that space.

It is almost trivial to build distributed processing around Redis. The entire key store can be used to store work units and results, Redis LISTs can be used as Queue() like object, and the PUB/SUB support can be used for Event-like handling. You can hash your keys and use values, replicated across a loose cluster of Redis instances, to store the topology and hash-token mappings to provide consistent hashing and fail-over for scaling beyond the capacity of any single instance for co-ordinating your workers and marshaling data (pickled, JSON, BSON, or YAML) among them.

Of course as you start to build a larger scale and more sophisticated solution around Redis you are re-implementing many of the features that have already been solved using, Celery, Apache Spark and Hadoop, Zookeeper, etcd, Cassandra and so on. Those all have modules for Python access to their services.

[Update: A couple of resources for consideration if you're considering Python for computationally intensive across distributed systems: IPython Parallel and PySpark. While these are general purpose distributed computing systems, they are particularly accessible and popular subsystems data science and analytics].

Conclusion

There you have the gamut of processing alternatives for Python, from single threaded, with simple synchronous calls to sub-processes, pools of polled subprocesses, threaded and multiprocessing, event-driven co-operative multi-tasking, and out to distributed processing.