Category: Blather

Part 3: Our Eye-beams Begin to Twist

This continues the introduction started here. You can find an index to the entire series here.

Doing Nothing, the Twisted Way

Eventually we are going to re-implement our asynchronous poetry client using Twisted. But first let’s write a few really simple Twisted programs just to get the flavor of things. As I mentioned in Part 2, I developed these examples using Twisted 8.2.0. Twisted APIs do change, but the core APIs we are going to use will likely change slowly, if at all, so I expect these examples to work for many future releases. If you don’t have Twisted installed you can obtain it here.

The absolute simplest Twisted program is listed below, and is also available in basic-twisted/simple.py in the base directory of the twisted-intro example code.

from twisted.internet import reactor
reactor.run()

You can run it like this:

python basic-twisted/simple.py

As we saw in Part 2, Twisted is an implementation of the Reactor Pattern and thus contains an object that represents the reactor, or event loop, that is the heart of any Twisted program. The first line of our program imports the reactor object so we can use it, and the second line tells the reactor to start running the loop.

This program just sits there doing nothing. You’ll have to stop it by pressing Control-C, otherwise it will just sit there forever. Normally we would have given the loop one or more file descriptors (connected to, say, a poetry server) that we want to monitor for I/O. We’ll see how to do that later, but for now our reactor loop is stuck. Note that this is not a busy loop which keeps cycling over and over. If you happen to have a CPU meter on your screen, you won’t see any spikes caused by this technically infinite loop. In fact, our program isn’t using any CPU at all. Instead, the reactor is stuck at the top cycle of Figure 5, waiting for an event that will never come (to be specific, waiting on a select call with no file descriptors).

That might make for a compelling metaphor of Hamletian inaction, but it’s still a pretty boring program. We’re about to make it more interesting, but we can already draw a few conclusions:

1. Twisted’s reactor loop doesn’t start until told to. You start it by calling reactor.run().
2. The reactor loop runs in the same thread it was started in. In this case, it runs in the main (and only) thread.
3. Once the loop starts up, it just keeps going. The reactor is now “in control” of the program (or the specific thread it was started in).
4. If it doesn’t have anything to do, the reactor loop does not consume CPU.
5. The reactor isn’t created explicitly, just imported.

That last point is worth elaborating on. In Twisted, the reactor is basically a Singleton. There is only one reactor object and it is created implicitly when you import it. If you open the reactor module in the twisted.internet package you will find very little code. The actual implementation resides in other files (see, for example, twisted.internet.selectreactor).

Twisted actually contains multiple reactor implementations. As mentioned in Part 2, the select call is just one method of waiting on file descriptors. Twisted includes several reactor implementations that use a variety of different methods. For example, twisted.internet.pollreactor uses the poll system call instead of select.

To use a specific reactor, you must install it before importing twisted.internet.reactor. Here is how you install the pollreactor:

from twisted.internet import pollreactor
pollreactor.install()

If you import twisted.internet.reactor without first installing a specific reactor implementation, then Twisted will install the default reactor for you. The particular one you get will depend on the operating system and Twisted version you are using. For that reason, it is general practice not to import the reactor at the top level of modules to avoid accidentally installing the default reactor. Instead, import the reactor in the same scope in which you use it.

Note: as of this writing, Twisted has been moving gradually towards an architecture which would allow multiple reactors to co-exist. In this scheme, a reactor object would be passed around as a reference rather than imported from a module.

Note: not all operating systems support the poll call. If that is the case for your system, this example will not work.

Now we can re-implement our first Twisted program using the pollreactor, as found in basic-twisted/simple-poll.py:

from twisted.internet import pollreactor
pollreactor.install()

from twisted.internet import reactor
reactor.run()

And we have a poll loop that does nothing at all instead of a select loop that does nothing at all. Neato.

We’re going to stick with the default reactor for the rest of this introduction. For the purposes of learning Twisted, all the reactors do the same thing.

Hello, Twisted

Let’s make a Twisted program that at least does something. Here’s one that prints a message to the terminal window, after the reactor loop starts up:

def hello():
print 'Hello from the reactor loop!'
print 'Lately I feel like I\'m stuck in a rut.'

from twisted.internet import reactor

reactor.callWhenRunning(hello)

print 'Starting the reactor.'
reactor.run()

This program is in basic-twisted/hello.py. If you run it, you will see this output:

Starting the reactor.
Hello from the reactor loop!
Lately I feel like I'm stuck in a rut.

You’ll still have to kill the program yourself, since it gets stuck again after printing those lines.

Notice the hello function is called after the reactor starts running. That means it is called by the reactor itself, so Twisted code must be calling our function. We arrange for this to happen by invoking the reactor method callWhenRunning with a reference to the function we want Twisted to call. And, of course, we have to do that before we start the reactor.

We use the term callback to describe the reference to the hello function. A callback is a function reference that we give to Twisted (or any other framework) that Twisted will use to “call us back” at the appropriate time, in this case right after the reactor loop starts up. Since Twisted’s loop is separate from our code, most interactions between the reactor core and our business logic will begin with a callback to a function we gave to Twisted using various APIs.

We can see how Twisted is calling our code using this program:

import traceback

def stack():
    print 'The python stack:'
    traceback.print_stack()

from twisted.internet import reactor
reactor.callWhenRunning(stack)
reactor.run()

You can find it in basic-twisted/stack.py and it prints out something like this:

The python stack:
...
  reactor.run() <-- This is where we called the reactor
...
...  <-- A bunch of Twisted function calls
...
  traceback.print_stack() <-- The second line in the stack function

Don’t worry about all the Twisted calls in between. Just notice the relationship between the reactor.run() call and our callback.

What’s the deal with callbacks?

Twisted is not the only reactor framework that uses callbacks. The older asynchronous Python frameworks Medusa and asyncore also use them. As do the GUI toolkits GTK and QT, both based, like many GUI frameworks, on a reactor loop.

The developers of reactive systems sure love callbacks. Maybe they should just marry them. Maybe they already did. But consider this:

1. The reactor pattern is single-threaded.
2. A reactive framework like Twisted implements the reactor loop so our code doesn’t have to.
3. Our code still needs to get called to implement our business logic.
4. Since it is “in control” of the single thread, the reactor loop will have to call our code.
5. The reactor can’t know in advance which part of our code needs to be called.

In this situation callbacks are not just one option — they are the only real game in town.

Figure 6 shows what happens during a callback:

Figure 6 illustrates some important properties of callbacks:

1. Our callback code runs in the same thread as the Twisted loop.
2. When our callbacks are running, the Twisted loop is not running.
3. And vice versa.
4. The reactor loop resumes when our callback returns.

During a callback, the Twisted loop is effectively “blocked” on our code. So we should make sure our callback code doesn’t waste any time. In particular, we should avoid making blocking I/O calls in our callbacks. Otherwise, we would be defeating the whole point of using the reactor pattern in the first place. Twisted will not take any special precautions to prevent our code from blocking, we just have to make sure not to do it. As we will eventually see, for the common case of network I/O we don’t have to worry about it as we let Twisted do the asynchronous communication for us.

Other examples of potentially blocking operations include reading or writing from a non-socket file descriptor (like a pipe) or waiting for a subprocess to finish. Exactly how you switch from blocking to non-blocking operations is specific to what you are doing, but there is often a Twisted API that will help you do it. Note that many standard Python functions have no way to switch to a non-blocking mode. For example, the os.system function will always block until the subprocess is finished. That’s just how it works. So when using Twisted, you will have to eschew os.system in favor of the Twisted API for launching subprocesses.

Goodbye, Twisted

It turns out you can tell the Twisted reactor to stop running by using the reactor’s stop method. But once stopped the reactor cannot be restarted, so it’s generally something you do only when your program needs to exit.

Note: there has been past discussion on the Twisted mailing list about making the reactor “restartable” so it could be started and stopped as you like. But as of version 8.2.0, you can only start (and thus stop) the reactor once.

Here’s a program, listed in basic-twisted/countdown.py, which stops the reactor after a 5 second countdown:

class Countdown(object):

    counter = 5

    def count(self):
        if self.counter == 0:
            reactor.stop()
        else:
            print self.counter, '...'
            self.counter -= 1
            reactor.callLater(1, self.count)

from twisted.internet import reactor

reactor.callWhenRunning(Countdown().count)

print 'Start!'
reactor.run()
print 'Stop!'

This program uses the callLater API to register a callback with Twisted. With callLater the callback is the second argument and the first argument is the number of seconds in the future you would like your callback to run. You can use a floating point number to specify a fractional number of seconds, too.

So how does Twisted arrange to execute the callback at the right time? Since this program doesn’t listen on any file descriptors, why doesn’t it get stuck in the select loop like the others? The select call, and the others like it, also accepts an optional timeout value. If a timeout value is supplied and no file descriptors have become ready for I/O within the specified time then the select call will return anyway. Incidentally, by passing a timeout value of zero you can quickly check (or “poll”) a set of file descriptors without blocking at all.

You can think of a timeout as another kind of event the event loop of Figure 5 is waiting for. And Twisted uses timeouts to make sure any “timed callbacks” registered with callLater get called at the right time. Or rather, at approximately the right time. If another callback takes a really long time to execute, a timed callback may be delayed past its schedule. Twisted’s callLater mechanism cannot provide the sort of guarantees required in a hard real-time system.

Here is the output of our countdown program:

Start!
5 ...
4 ...
3 ...
2 ...
1 ...
Stop!

Note the “Stop!” line at the ends shows us that when the reactor exits, the reactor.run call returns. And we have a program that stops all by itself.

Take That, Twisted

Since Twisted often ends up calling our code in the form of callbacks, you might wonder what happens when a callback raises an exception. Let’s try it out. The program in basic-twisted/exception.py raises an exception in one callback, but behaves normally in another:

def falldown():
    raise Exception('I fall down.')

def upagain():
    print 'But I get up again.'
    reactor.stop()

from twisted.internet import reactor

reactor.callWhenRunning(falldown)
reactor.callWhenRunning(upagain)

print 'Starting the reactor.'
reactor.run()

When you run it at the command line, you will see this output:

Starting the reactor.
Traceback (most recent call last):
  ... # I removed most of the traceback
exceptions.Exception: I fall down.
But I get up again.

Notice the second callback runs after the first, even though we see the traceback from the exception the first raised. And if you comment out the reactor.stop() call, the program will just keep running forever. So the reactor will keep going even when our callbacks fail (though it will report the exception).

Network servers generally need to be pretty robust pieces of software. They’re not supposed to crash whenever any random bug shows its head. That’s not to say we should be lackadaisical when it comes to handling our own errors, but it’s nice to know Twisted has our back.

Poetry, Please

Now we’re ready to grab some poetry with Twisted. In Part 4, we will implement a Twisted version of our asynchronous poetry client.

Suggested Exercises

1. Update the countdown.py program to have three independently running counters going at different rates. Stop the reactor when all counters have finished.
2. Consider the LoopingCall class in twisted.internet.task. Rewrite the countdown program above to use LoopingCall. You only need the start and stop methods and you don’t need to use the “deferred” return value in any way. We’ll learn what a “deferred” value is in a later Part.

Part 2: Slow Poetry and the Apocalypse

This continues the introduction started here. And if you read it, welcome back. Now we’re going to get our hands dirty and write some code. But first, let’s get some assumptions out of the way.

My Assumptions About You

I will proceed as if you have a basic working knowledge of writing synchronous programs in Python, and know at least a little bit about Python socket programming. If you have never used sockets before, you might read the socket module documentation now, especially the example code towards the end. If you’ve never used Python before, then the rest of this introduction is probably going to be rather opaque.

My Assumptions About Your Computer

My experience with Twisted is mainly on Linux systems, and it is a Linux system on which I developed the examples. And while I won’t intentionally make the code Linux-dependent, some of it, and some of what I say, may only apply to Linux and other UNIX-like systems (like Mac OSX or FreeBSD). Windows is a strange, murky place and, if you are hacking in it, I can’t offer you much more beyond my heartfelt sympathies.

I will assume you have installed relatively recent versions of Python and Twisted. The examples were developed with Python 2.5 and Twisted 8.2.0.

Also, you can run all the examples on a single computer, although you can configure them to run on a network of systems as well. But for learning the basic mechanics of asynchronous programming, a single computer will do fine.

Getting the example code

The example code is available as a zip or tar file or as a clone of my public git repository. If you can use git or another version control system that can read git repositories, then I recommend using that method as I will update the examples over time and it will be easier for you to stay current. As a bonus, it includes the SVG source files used to generate the figures. Here is the git command to clone the repository:

git clone git://github.com/jdavisp3/twisted-intro.git

The rest of this tutorial will assume you have the latest copy of the example code and you have multiple shells open in its top-level directory (the one with the README file).

Slow Poetry

Although CPUs are much faster than networks, most networks are still a lot faster than your brain, or at least faster than your eyeballs. So it can be challenging to get the “cpu’s-eye-view” of network latency, especially when there’s only one machine and the bytes are whizzing past at full speed on the loopback interface. What we need is a slow server, one with artificial delays we can vary to see the effect. And since servers have to serve something, ours will serve poetry. The example code includes a sub-directory called poetry with one poem each by John Donne, W.B. Yeats, and Edgar Allan Poe. Of course, you are free to substitute your own poems for the server to dish up.

The basic slow poetry server is implemented in blocking-server/slowpoetry.py. You can run one instance of the server like this:

python blocking-server/slowpoetry.py poetry/ecstasy.txt

That command will start up the blocking server with John Donne’s poem “Ecstasy” as the poem to serve. Go ahead and look at the source code to the blocking server now. As you can see, it does not use Twisted, only basic Python socket operations. It also sends a limited number of bytes at a time, with a fixed time delay between them. By default, it sends 10 bytes every 0.1 seconds, but you can change these parameters with the –num-bytes and –delay command line options. For example, to send 50 bytes every 5 seconds:

python blocking-server/slowpoetry.py --num-bytes 50 --delay 5 poetry/ecstasy.txt

When the server starts up it prints out the port number it is listening on. By default, this is a random port that happens to be available on your machine. When you start varying the settings, you will probably want to use the same port number over again so you don’t have to adjust the client command. You can specify a particular port like this:

python blocking-server/slowpoetry.py --port 10000 poetry/ecstasy.txt

If you have the netcat program available, you could test the above command like this:

netcat localhost 10000

If the server is working, you will see the poem slowly crawl its way down your screen. Ecstasy! You will also notice the server prints out a line each time it sends some bytes. Once the complete poem has been sent, the server closes the connection.

By default, the server only listens on the local “loopback” interface. If you want to access the server from another machine, you can specify the interface to listen on with the –iface option.

Not only does the server send each poem slowly, if you read the code you will find that while the server is sending poetry to one client, all other clients must wait for it to finish before getting even the first line. It is truly a slow server, and not much use except as a learning device.

Or is it?

On the other hand, if the more pessimistic of the Peak Oil folks are right and our world is heading for a global energy crisis and planet-wide societal meltdown, then perhaps one day soon a low-bandwidth, low-power poetry server could be just what we need. Imagine, after a long day of tending your self-sufficient gardens, making your own clothing, serving on your commune’s Central Organizing Committee, and fighting off the radioactive zombies that roam the post-apocalyptic wastelands, you could crank up your generator and download a few lines of high culture from a vanished civilization. That’s when our little server will really come into its own.

The Blocking Client

Also in the example code is a blocking client which can download poems from multiple servers, one after another. Let’s give our client three tasks to perform, as in Figure 1 from Part 1. First we’ll start three servers, serving three different poems. Run these commands in three different terminal windows:

python blocking-server/slowpoetry.py --port 10000 poetry/ecstasy.txt --num-bytes 30
python blocking-server/slowpoetry.py --port 10001 poetry/fascination.txt
python blocking-server/slowpoetry.py --port 10002 poetry/science.txt

You can choose different port numbers if one or more of the ones I chose above are already being used on your system. Note I told the first server to use chunks of 30 bytes instead of the default 10 since that poem is about three times as long as the others. That way they all finish around the same time.

Now we can use the blocking client in blocking-client/get-poetry.py to grab some poetry. Run the client like this:

python blocking-client/get-poetry.py 10000 10001 10002

Change the port numbers here, too, if you used different ones for your servers. Since this is the blocking client, it will download one poem from each port number in turn, waiting until a complete poem is received until starting the next. Instead of printing out the poems, the blocking client produces output like this:

Task 1: get poetry from: 127.0.0.1:10000
Task 1: got 3003 bytes of poetry from 127.0.0.1:10000 in 0:00:10.126361
Task 2: get poetry from: 127.0.0.1:10001
Task 2: got 623 bytes of poetry from 127.0.0.1:10001 in 0:00:06.321777
Task 3: get poetry from: 127.0.0.1:10002
Task 3: got 653 bytes of poetry from 127.0.0.1:10002 in 0:00:06.617523
Got 3 poems in 0:00:23.065661

This is basically a text version of Figure 1, where each task is downloading a single poem. Your times may be a little different, and will vary as you change the timing parameters of the servers. Try changing those parameters to see the effect on the download times.

You might take a look at the source code to the blocking server and client now, and locate the points in the code where each blocks while sending or receiving network data.

The Asynchronous Client

Now let’s take a look at a simple asynchronous client written without Twisted. First let’s run it. Get a set of three servers going on the same ports like we did above. If the ones you ran earlier are still going, you can just use them again. Now we can run the asynchronous client, located in async-client/get-poetry.py, like this:

python async-client/get-poetry.py 10000 10001 10002

And you should get some output like this:

Task 1: got 30 bytes of poetry from 127.0.0.1:10000
Task 2: got 10 bytes of poetry from 127.0.0.1:10001
Task 3: got 10 bytes of poetry from 127.0.0.1:10002
Task 1: got 30 bytes of poetry from 127.0.0.1:10000
Task 2: got 10 bytes of poetry from 127.0.0.1:10001
...
Task 1: 3003 bytes of poetry
Task 2: 623 bytes of poetry
Task 3: 653 bytes of poetry
Got 3 poems in 0:00:10.133169

This time the output is much longer because the asynchronous client prints a line each time it downloads some bytes from any server, and these slow poetry servers just dribble out the bytes little by little. Notice that the individual tasks are mixed together just like in Figure 3 from Part 1.

Try varying the delay settings for the servers (e.g., by making one server slower than the others) to see how the asynchronous client automatically “adjusts” to the speed of the slower servers while still keeping up with the faster ones. That’s asynchronicity in action.

Also notice that, for the server settings we chose above, the asynchronous client finishes in about 10 seconds while the synchronous client needs around 23 seconds to get all the poems. Now recall the differences between Figure 3 and Figure 4 in Part 1. By spending less time blocking, our asynchronous client can download all the poems in a shorter overall time. Now, our asynchronous client does block some of the time. Our slow server is slow.  It’s just that the asynchronous client spends a lot less time blocking than the “blocking” client does, because it can switch back and forth between all the servers.

Technically, our asynchronous client is performing a blocking operation: it’s writing to the standard output file descriptor with those print statements! This isn’t a problem for our examples. On a local machine with a terminal shell that’s always willing to accept more output the print statements won’t really block, and execute quickly relative to our slow servers. But if we wanted our program to be part of a process pipeline and still execute asynchronously, we would need to use asynchronous I/O for standard input and output, too. Twisted includes support for doing just that, but to keep things simple we’re just going to use print statements, even in our Twisted programs.

A Closer Look

Now take a look at the source code for the asynchronous client. Notice the main differences between it and the synchronous client:

1. Instead of connecting to one server at a time, the asynchronous client connects to all the servers at once.
2. The socket objects used for communication are placed in non-blocking mode with the call to setblocking(0).
3. The select method in the select module is used to wait (block) until any of the sockets are ready to give us some data.
4. When reading data from the servers, we read only as much as we can until the socket would block, and then move on to the next socket with data to read (if any). This means we have to keep track of the poetry we’ve received from each server so far.

The core of the asynchronous client is the top-level loop in the get_poetry function. This loop can be broken down into steps:

1. Wait (block) on all open sockets using select until one (or more) sockets has data to be read.
2. For each socket with data to be read, read it, but only as much as is available now. Don’t block.
3. Repeat, until all sockets have been closed.

The synchronous client had a loop as well (in the main function), but each iteration of the synchronous loop downloaded one complete poem. In one iteration of the asynchronous client we might download pieces of all the poems we are working on, or just some of them. And we don’t know which ones we will work on in a given iteration, or how much data we will get from each one. That all depends on the relative speeds of the servers and the state of the network. We just let select tell us which ones are ready to go, and then read as much data as we can from each socket without blocking.

If the synchronous client always contacted a fixed number of servers (say 3), it wouldn’t need an outer loop at all, it could just call its blocking get_poetry function three times in succession. But the asynchronous client can’t do without an outer loop — to gain the benefits of asynchronicity, we need to wait on all of our sockets at once, and only process as much data as each is capable of delivering in any given iteration.

This use of a loop which waits for events to happen, and then handles them, is so common that it has achieved the status of a design pattern: the reactor pattern. It is visualized in Figure 5 below:

The loop is a “reactor” because it waits for and then reacts to events. For that reason it is also known as an event loop. And since reactive systems are often waiting on I/O, these loops are also sometimes called select loops, since the select call is used to wait for I/O. So in a select loop, an “event” is when a socket becomes available for reading or writing. Note that select is not the only way to wait for I/O, it is just one of the oldest methods (and thus widely available). There are several newer APIs, available on different operating systems, that do the same thing as select but offer (hopefully) better performance. But leaving aside performance, they all do the same thing: take a set of sockets (really file descriptors) and block until one or more of them is ready to do I/O.

Note that it’s possible to use select and its brethren to simply check whether a set of file descriptors is ready for I/O without blocking. This feature permits a reactive system to perform non-I/O work inside the loop. But in reactive systems it is often the case that all work is I/O-bound, and thus blocking on all file descriptors conserves CPU resources.

Strictly speaking, the loop in our asynchronous client is not the reactor pattern because the loop logic is not implemented separately from the “business logic” that is specific to the poetry servers. They are all just mixed together. A real implementation of the reactor pattern would implement the loop as a separate abstraction with the ability to:

1. Accept a set of file descriptors you are interested in performing I/O with.
2. Tell you, repeatedly, when any file descriptors are ready for I/O.

And a really good implementation of the reactor pattern would also:

1. Handle all the weird corner cases that crop up on different systems.
2. Provide lots of nice abstractions to help you use the reactor with the least amount of effort.
3. Provide implementations of public protocols that you can use out of the box.

Well that’s just what Twisted is — a robust, cross-platform implementation of the Reactor Pattern with lots of extras. And in Part 3 we will start writing some simple Twisted programs as we move towards a Twisted version of Get Poetry Now!.

Suggested Exercises

1. Do some timing experiments with the blocking and asynchronous clients by varying the number and settings of the poetry servers.
2. Could the asynchronous client provide a get_poetry function that returned the text of the poem? Why not?
3. If you wanted a get_poetry function in the asynchronous client that was analogous to the synchronous version of get_poetry, how could it work? What arguments and return values might it have?

Part 1: In Which We Begin at the Beginning

Preface

Someone recently posted to the Twisted mailing list asking for something like the “Twisted introduction for people on a deadline”. Full disclosure: this isn’t it. On the spectrum of introductions to Twisted and asynchronous programming in Python, it may be on the exact opposite end. So if you don’t have any time, or any patience, this isn’t the introduction you are looking for.

However, I also believe that if you are new to asynchronous programming, a quick introduction is simply not possible, at least if you are not a genius. I’ve used Twisted successfully for a number of years and having thought about how I initially learned it (slowly), and what I found difficult, I’ve come to the conclusion that much of the challenge does not stem from Twisted per se, but rather in the acquisition of the “mental model” required to write and understand asynchronous code. Most of the Twisted source code is clear and well written, and the online documentation is good, at least by the standards of most free software. But without that mental model, reading the Twisted codebase, or code that uses Twisted, or even much of the documentation, will result in confusion and headache.

So the first parts of this introduction are designed to help you acquire that model and only later on will we introduce the features of Twisted. In fact, we will start without using Twisted at all, instead using simple Python programs to illustrate how an asynchronous system works. And once we get into Twisted, we will begin with very low-level aspects that you would not normally use in day-to-day programming. Twisted is a highly abstracted system and this gives you tremendous leverage when you use it to solve problems. But when you are learning Twisted, and particularly when you are trying to understand how Twisted actually works, the many levels of abstraction can cause troubles. So we will go from the inside-out, starting with the basics.

And once you have the mental model in place, I think you will find reading the Twisted documentation, or just browsing the source code, to be much easier. So let’s begin.

The Models

We will start by reviewing two (hopefully) familiar models in order to contrast them with the asynchronous model. By way of illustration we will imagine a program that consists of three conceptually distinct tasks which must be performed to complete the program. We will make these tasks more concrete later on, but for now we won’t say anything about them except the program must perform them. Note I am using “task” in the non-technical sense of “something that needs to be done”.

The first model we will look at is the single-threaded synchronous model, in Figure 1 below:

This is the simplest style of programming. Each task is performed one at a time, with one finishing completely before another is started. And if the tasks are always performed in a definite order, the implementation of a later task can assume that all earlier tasks have finished without errors, with all their output available for use — a definite simplification in logic.

We can contrast the synchronous model with another one, the threaded model illustrated in Figure 2:

In this model, each task is performed in a separate thread of control. The threads are managed by the operating system and may, on a system with multiple processors or multiple cores, run truly concurrently, or may be interleaved together on a single processor. The point is, in the threaded model the details of execution are handled by the OS and the programmer simply thinks in terms of independent instruction streams which may run simultaneously. Although the diagram is simple, in practice threaded programs can be quite complex because of the need for threads to coordinate with one another. Thread communication and coordination is an advanced programming topic and can be difficult to get right.

Some programs implement parallelism using multiple processes instead of multiple threads. Although the programming details are different, for our purposes it is the same model as in Figure 2.

Now we can introduce the asynchronous model in Figure 3:

In this model, the tasks are interleaved with one another, but in a single thread of control. This is simpler than the threaded case because the programmer always knows that when one task is executing, another task is not. Although in a single-processor system a threaded program will also execute in an interleaved pattern, a programmer using threads should still think in terms of Figure 2, not Figure 3, lest the program work incorrectly when moved to a multi-processor system. But a single-threaded asynchronous system will always execute with interleaving, even on a multi-processor system.

There is another difference between the asynchronous and threaded models. In a threaded system the decision to suspend one thread and execute another is largely outside of the programmer’s control. Rather, it is under the control of the operating system, and the programmer must assume that a thread may be suspended and replaced with another at almost any time. In contrast, under the asynchronous model a task will continue to run until it explicitly relinquishes control to other tasks. This is a further simplification from the threaded case.

Note that it is possible to mix the asynchronous and threaded models and use both in the same system. But for most of this introduction, we will stick to “plain vanilla” asynchronous systems with one thread of control.

The Motivation

We’ve seen that the asynchronous model is simpler than the threaded one because there is a single instruction stream and tasks explicitly relinquish control instead of being suspended arbitrarily. But the asynchronous model is clearly more complex than the synchronous case. The programmer must organize each task as a sequence of smaller steps that execute intermittently. And if one task uses the output of another, the dependent task must be written to accept its input as a series of bits and pieces instead of all together. Since there is no actual parallelism, it appears from our diagrams that an asynchronous program will take just as long to execute as a synchronous one, perhaps longer as the asynchronous program might exhibit poorer locality of reference.

So why would you choose to use the asynchronous model? There are at least two reasons. First, if one or more of the tasks are responsible for implementing an interface for a human being, then by interleaving the tasks together the system can remain responsive to user input while still performing other work in the “background”. So while the background tasks may not execute any faster, the system will be more pleasant for the person using it.

However, there is a condition under which an asynchronous system will simply outperform a synchronous one, sometimes dramatically so, in the sense of performing all of its tasks in an overall shorter time. This condition holds when tasks are forced to wait, or block, as illustrated in Figure 4:

In the figure, the gray sections represent periods of time when a particular task is waiting (blocking) and thus cannot make any progress. Why would a task be blocked? A frequent reason is that it is waiting to perform I/O, to transfer data to or from an external device. A typical CPU can handle data transfer rates that are orders of magnitude faster than a disk or a network link is capable of sustaining. Thus, a synchronous program that is doing lots of I/O will spend much of its time blocked while a disk or network catches up. Such a synchronous program is also called a blocking program for that reason.

Notice that Figure 4, a blocking program, looks a bit like Figure 3, an asynchronous program. This is not a coincidence. The fundamental idea behind the asynchronous model is that an asynchronous program, when faced with a task that would normally block in a synchronous program, will instead execute some other task that can still make progress. So an asynchronous program only “blocks” when no task can make progress and is thus called a non-blocking program. And each switch from one task to another corresponds to the first task either finishing, or coming to a point where it would have to block. With a large number of potentially blocking tasks, an asynchronous program can outperform a synchronous one by spending less overall time waiting, while devoting a roughly equal amount of time to real work on the individual tasks.

Compared to the synchronous model, the asynchronous model performs best when:

1. There are a large number of tasks so there is likely always at least one task that can make progress.
2. The tasks perform lots of I/O, causing a synchronous program to waste lots of time blocking when other tasks could be running.
3. The tasks are largely independent from one another so there is little need for inter-task communication (and thus for one task to wait upon another).

These conditions almost perfectly characterize a typical busy network server (like a web server) in a client-server environment. Each task represents one client request with I/O in the form of receiving the request and sending the reply. And client requests (being mostly reads) are largely independent. So a network server implementation is a prime candidate for the asynchronous model and this is why Twisted is first and foremost a networking library.

Onward and Upward

This is the end of Part 1. In Part 2, we will write some network programs, both blocking and non-blocking, as simply as possible (without using Twisted), to get a feel for how an asynchronous Python program actually works.