greenlet local storage on greenlet 0.4.0

Posted on by saghul Leave a comment

Greenlet 0.4.0 brought an interesting new feature: an instance dictionary on each greenlet object, which makes it a lot simpler to implement greenlet local storage. Here is how greenlet local storage is currently implemented in Eventlet and in Gevent.

As it can be seen, the implementation is not particularly straightforward, mainly due to the fact that the actual information needs to be stored in a separate entity and mapped to each greenlet.

Thanks to the instance dictionary added in 0.4.0, we can use some attribute in it to keep the locally stored objects. The plan is to use a dictionary called __local_dict__ and store the greenlet local attributes there. Here is how it looks like:

[gist]https://gist.github.com/4151154[/gist]

Hope it’s of use.

:wq

Using functools.partial instead of saving arguments

Posted on by saghul 4 Comments

I’m a big fan of functools.partial myself. It allows you to take a function, preset some of its arguments and return another function which you can call with more arguments. Not sure if currying is the right term here, but I’ve heard people refer to functools.partial like that.

Today while browsing the source code for the futures module I came across this class:

[gist]https://gist.github.com/4048428[/gist]

The moment I saw that func, args and kwargs were saved as attributes in the instance I thought: “why not use partial and save a single attribute?”. Then I thought that maybe performance had something to do here, so I wrote a dead simple stupid test to check it out:

[gist]https://gist.github.com/4048405[/gist]

Here are the results using CPython 2.7.3:

In [3]: %timeit testpartial.do_test(testpartial.WorkItem)
100000 loops, best of 3: 5.26 us per loop

In [4]: %timeit testpartial.do_test(testpartial.WorkItemPartial)
100000 loops, best of 3: 2.65 us per loop

We are talking microseconds here, but still the version with partial is almost twice as fast.

Now, lets see how PyPy performs:

In [9]: %timeit -n 10000 testpartial.do_test(testpartial.WorkItem)
10000 loops, best of 3: 154 ns per loop
Compiler time: 554050.78 s

In [10]: %timeit -n 10000 testpartial.do_test(testpartial.WorkItemPartial)
10000 loops, best of 3: 2.49 us per loop

Fun, the version without partial goes into the nanoseconds! And the one with partial doesn’t improve much with regards to CPython. Interesting.

So what’s the take here? Well, whenever I see fit I use partial, code looks nicer and it’s apparently faster, so why not? 🙂

:wq

Fast(er) locks in Python?

Posted on by saghul 2 Comments

While searching for some information on Python locks I recently ran across this great post by David Beazley. In it, he explains how the syncronization primitives in the Python standard library are implemented. Basically, Lock is implemented as a binary semaphore in C, and the rest are implemented in pure Python. Even if the post is from 2009 this is still the case. UPDATE: As Antoine Pitrou points out in the comments, starting with CPython 3.2, RLock is now implemented in C.

This got me thinking. As you know I’ve created pyuv, a Python wrapper for libuv, and libuv includes cross-platform implementations for mutexes, semaphores, conditions rwlocks and barriers, which I never bothered to add to pyuv. I just didn’t add them because I thought they didn’t add any value, but after reading David’s article I decided to do a quick test: implement wrappers for a mutex and a condition variable and use them in a Queue implementation in order to see if there was any difference in performance. Not that I ever ran into performance issues related to that, but I was curious anyway 🙂

Someone may think “oh, but given that Python has the GIL, how does using multiple threads and speeding up the locks matter?”. The GIL is released whenever a IO operation is performed, so if your Python application is multithreaded and it mainly deals with IO-bound tasks, they GIL is not that relevant. If your application is CPU bound, however, better have a look at the multiprocessing module.

So, lets get into the code! I implemented Barrier, Condition, Mutex, RWLock and Semaphore in this pyuv branch, which directly wrap their libuv counterpats. Then I copied the Queue implementation from the standard library and used the freshly wrapped synchronization primitives:

[gist]https://gist.github.com/3997233[/gist]

For the testing part, I used the timeit function from IPython with 5 runs. Not sure if it’s the best way, but results suggest it is a good way 🙂 Here is the test script:

[gist]https://gist.github.com/3997244[/gist]

Here are the results:

The tests were run with 2, 4 and 100 threads, and since I was testing performance, I added PyPy to the mix. Now, as you can see in the results, the custom Queue is about 33% faster than the one in the standard library, so I was pretty happy about that. Until I tested PyPy. It just beats the shit out of both, which is awesome 🙂

The performance increase on CPython is nice, there is a downside, however: libuv treats errors in this primitives a bit “abruptly”, it calls abort(). This means that if you use the locks incorrectly your program will core dump. I personally like it, because it helps you find and fix the problem right away, but not everyone may like it.

:wq