parallel-utils 1.1.0

This library implements a class Monitor, as defined by Per Brinch Hansen and C.A.R. Hoare, for synchronization and concurrent management of threads and processes in Python. It also provides other functions to ease the creation and collection of results for both threads and processes.

parallel-utils

This library implements a class Monitor, as defined by Per Brinch Hansen and C.A.R. Hoare, for synchronization and concurrent management of threads and processes in Python. It also provides other functions to ease the creation and collection of results for both threads and processes.

Table of contents

• Installation
• Usage
  • Monitor class
    • First example
      • @synchronized
    • Second example
      • @synchronized_priority
  • StaticMonitor class
  • Launching threads and processes
• Contributing
• License

Installation

Use the package manager pip to install parallel-utils.

pip3 install parallel-utils

Usage

Monitor class

There are two implementations of the Monitor class: one is located in the thread module and the other in the process module of parallel-utils.

Although it's safe to always use the Monitor class located in the process module, even if you're working only with threads, you will get slightly better performance when using the one located in the thread module. Therefore, it is recommended to use each one for what it is intended for.

From now on, for ease of reading, every time we say thread we will also be including process unless stated otherwise.

Also, from now until the end of this section, when we say function, we will not only refer to "whole" functions but we will also include pieces of code contained within a function.

A monitor essentially does two things:

1. It controls the maximum number of threads that can simultaneously access a function.

2. It organizes a set of functions so that they follow a strict order in their execution, regardless of the thread from which they are called.

First example

1. It allows controlling the maximum number of threads that can simultaneously access a function.

To achieve this first goal, the Monitor class includes the following couple of functions:

def lock_code(self, uid: Union[str, int], max_threads: int = 1)

def unlock_code(self, uid: Union[str, int])

The first one, lock_code, must be called at the beginning of the piece of code for which we want to control the maximum number of threads that can access it simultaneously.

The unlock_code function sets the limit of the scope of the lock_code function.

To do this, both functions must share a same unique identifier (uid), that can be either a string or an integer number. Let's see an example:

import concurrent.futures
from time import sleep
from parallel_utils.thread import Monitor, create_thread

m = Monitor()

def print_and_sleep():
    print('Hello')
    m.lock_code('example', max_threads=1)
    sleep(2)
    m.unlock_code('example')
    print('Goodbye')

th1 = create_thread(print_and_sleep)
th2 = create_thread(print_and_sleep)
th3 = create_thread(print_and_sleep)

concurrent.futures.wait([th1, th2, th3])

The example shown above takes 6 seconds to finalize its execution, since we have a lock_code that only allows one thread each time to execute the sleep function, and we are launching three threads.

If we set the lock_code to allow up to three threads at the same time, then the code only needs 2 seconds to finalize its execution, since all the three threads can make the sleep blocking call at the same time.

We'll se more about the create_thread and create_process functions later.

The last line, concurrent.futures.wait is a blocking call that waits until all the three threads finish running.

@synchronized

In the previous example, we just were protecting the piece of code wrapping the sleep call. But, what if we want to wrap the entire function?

Of course we could call lock_code on the first line, and unlock_code on the last one, and that would work just fine. Like this:

def print_and_sleep():
   m.lock_code('example', 1)
   print('Hello')
   sleep(2)
   print('Goodbye')
   m.unlock_code('example')

But, to simplify life for the programmer and for improving readability, there's some syntactic sugar we could use. And this is where the @synchronized decorator comes in and turns the above code into this:

from parallel_utils.thread import synchronized

@synchronized()
def print_and_sleep():
    print('Hello')
    sleep(2)
    print('Goodbye')

Let's see the decorator prototype:

@synchronized(max_threads: int = 1)

As you can see, the @synchronized decorator doesn't need an identifier, and by default only allows one thread to enter the function at the same time. However, we can override that default behavior with the optional max_threads argument.

If we want, for example, to allow up to two threads to enter the function at the same time, we only need to write:

@synchronized(2)
def print_and_sleep():
    print('Hello')
    sleep(2)
    print('Goodbye')

Note that this decorator has its own namespace for uids, which is completely independent of the namespace of the lock_code and unlock_code functions.

Second example

 2. It organizes a set of functions so that they follow a strict order in their execution, regardless of the thread from which they are called.

To achieve the second goal, the Monitor class includes the following couple of functions:

def lock_priority_code(uid: Union[str, int], order: int = 1, total: int = 1)

def unlock_code(uid: Union[str, int], order: int)

Yes, the unlock_code function is the same as before. And these two functions work quite similarly to the previous example, wrapping the code snippet that we want to control and sharing the same uid between them.

The main difference is that, in this case, we have to specify the order in which the code snippet will run and the total number of functions to sync with the supplied uid.

Let's see an example:

from time import sleep
from parallel_utils.process import Monitor, create_process

m = Monitor()

def say_hello(name):
    print('Entering hello')
    m.lock_priority_code('id1', order=1, total=2)
    print(f'Hello {name}!')
    m.unlock_code('id1')

def say_goodbye(name):
    print('Entering goodbye')
    m.lock_priority_code('id1', order=2)
    print(f'Goodbye {name}!')
    m.unlock_code('id1')

create_process(say_goodbye, 'Peter')
sleep(3)
create_process(say_hello, 'Peter')

this example will always print:

Entering goodbye
. . . (3 secs after)
Entering hello
Hello Peter!
Goodbye Peter!

even if you start the say_goodbye function long before the say_hello function. This is because the snippet in say_goobye has not the first turn, but the second, so it will make a blocking call and wait until say_hello calls unlock_code.

The total argument must be supplied at least once in any of the calls to lock_priority_code.

With these two functions, you can sort the execution of as many code snippets as you need.

Note that lock_code and lock_priority_code share the same namespace for uids.

@synchronized_priority

Just like before, there is a decorator that we can use to wrap a complete function and set its relative order of execution with respect to others.

This is its prototype:

@synchronized_priority(uid: Union[str, int], order: int = 1, total: int = None)

With it, the last example would look like this:

from parallel_utils.process import create_process, synchronized_priority

@synchronized_priority('id1', order=1)
def say_hello(name):
    print('Entering hello')
    print(f'Hello {name}!')

@synchronized_priority('id1', order=2, total=2)
def say_goodbye(name):
    print('Entering goodbye')
    print(f'Goodbye {name}!')

create_process(say_goodbye, 'Peter')
sleep(3)
create_process(say_hello, 'Peter')

Note that this time we've provided the total argument in the second call instead of the first one. Never mind. You could even supply it in both.

The above example will always print:

. . . (sleeps 3 secs)
Entering hello
Hello Peter!
Entering goodbye
Goodbye Peter!

Note that this decorator has its own namespace for uids, which is completely independent of the namespace of the lock_priority_code and unlock_code functions.

StaticMonitor class

The StaticMonitor class has exactly the same methods as the Monitor class. And it also has two implementations: one can be imported from parallel_utils.process and the other can be imported from parallel_utils.thread.

This class saves you the need to instantiate a Monitor, store it in a variable, and then use it. Instead of that, you can just call its methods like this:

from parallel_utils.process import StaticMonitor

def say_hello(name):
    print('Entering hello')
    StaticMonitor.lock_priority_code('id1', order=1, total=2)
    print(f'Hello {name}!')
    StaticMonitor.unlock_code('id1')

Note that this class, since it is static, has a unique namespace for uids that is shared among all calls to its methods.

Launching threads and processes

This library includes two very useful functions to quickly start processes and threads, and retrieve their results, which we have already seen in our examples:

def create_thread(func: Callable, *args: Any, **kwargs: Any) -> Future

def create_process(func: Callable, *args: Any, **kwargs: Any) -> Future

Like the rest of classes and objects in this library, they are located in parallel_utils.thread and parallel_utils.process respectively.

Their first argument is a Callable that, in turn, is called with *args and **kwargs.

They both return a Future object, which encapsulates the asynchronous execution of a callable.

Although the Future class has several interesting methods that you should read in the official documentation, by far the most interesting one is probably the result() method, which makes a blocking call until the thread finishes and returns.

Another interesting feature of these two methods is that you don't need to worry about zombie processes anymore due to a forgotten join() call, since all this logic is handled automatically. Just focus on making calls and retrieving results.

In this example, we compute two factorials, each one in a different process, to bypass the Python GIL and have real parallelism, and then we retrieve and print their results:

from parallel_utils.process import create_process

def factorial(n):
    res = 1
    for i in range(1,n+1):
        res *= i
    return res

f1 = create_process(factorial, 5)
f2 = create_process(factorial, 7)

print(f1.result(), f2.result())

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

License

This library is licensed under the GNU General Public License v3 or later (GPLv3+)