FastRLock

This is a C-level implementation of a fast, re-entrant, optimistic lock for CPython. It is a drop-in replacement for threading.RLock. FastRLock is implemented in Cython and also provides a C-API for direct use from Cython code via from fastrlock cimport rlock or from cython.cimports.fastrlock import rlock. See the pxd interface file.

Under normal conditions, it is about 10x faster than threading.RLock in Python 2.7 because it avoids all locking unless two or more threads try to acquire it at the same time. Under congestion, it is still about 10% faster than RLock due to being implemented in Cython.

This is mostly equivalent to the revised RLock implementation in Python 3.2, but still faster due to being implemented in Cython. However, in Python 3.4 and later, the threading.RLock implementation in the stdlib tends to be as fast or even faster than the lock provided by this package, when called through the Python API. FastRLock is still faster also on these systems when called through its Cython API from other Cython modules.

It was initially published as a code recipe here: https://code.activestate.com/recipes/577336-fast-re-entrant-optimistic-lock-implemented-in-cyt/

FastRLock has been used and tested in Lupa for several years.

How does it work?

The FastRLock implementation optimises for the non-congested case. It works by exploiting the availability of the GIL. Since it knows that it holds the GIL when the acquire()/release() methods are called, it can safely check the lock for being held by other threads and just count any re-entries as long as it is always the same thread that acquires it. This is a lot faster than actually acquiring the underlying lock.

When a second thread wants to acquire the lock as well, it first checks the lock count and finds out that the lock is already owned. If the underlying lock is also held by another thread already, it then just frees the GIL and asks for acquiring the lock, just like RLock does. If the underlying lock is not held, however, it acquires it immediately and basically hands over the ownership by telling the current owner to free it when it’s done. Then, it falls back to the normal non-owner behaviour that asks for the lock and will eventually acquire it when it gets released. This makes sure that the real lock is only acquired when at least two threads want it.

All of these operations are basically atomic because any thread that modifies the lock state always holds the GIL. Note that the implementation must not call any Python code while handling the lock, as calling into Python may lead to a context switch which hands over the GIL to another thread and thus breaks atomicity. Therefore, the code misuses Cython’s ‘nogil’ annotation to make sure that no Python code slips in accidentally.

The smart reliance on the GIL is also the reason why freethreading Python is not supported.

How fast is it?

Here are some timings for the following scenarios:

1. five acquire-release cycles (‘lock_unlock’)

2. five acquire calls followed by five release calls (nested locking, ‘reentrant_lock_unlock’)

3. a mixed and partly nested sequence of acquire and release calls (‘mixed_lock_unlock’)

4. five acquire-release cycles that do not block (‘lock_unlock_nonblocking’)

All four are benchmarked for the single threaded case and the multi threaded case with 10 threads. I also tested it with 20 threads only to see that it then takes about twice the time for both versions. Note also that the congested case is substantially slower for both locks and the benchmark includes the thread creation time, so I only looped 100x here to get useful timings instead of 1000x for the single threaded case.

The results here are mixed. Depending on the optimisation of the CPython installation, it can be faster, about the same speed, or somewhat slower. In any case, the direct Cython interface is always faster than going through the Python API, because it avoids the Python call overhead and executes a C call instead.

On Linux x86-64:

Testing RLock (3.14.8)

sequential (x1000):
lock_unlock              :      0.47 msec
reentrant_lock_unlock    :      0.47 msec
mixed_lock_unlock        :      0.43 msec
lock_unlock_nonblocking  :      0.60 msec
context_manager          :      2.48 msec

threaded 10T (x100):
lock_unlock              :    356.06 msec
reentrant_lock_unlock    :    371.20 msec
mixed_lock_unlock        :    355.51 msec
lock_unlock_nonblocking  :    372.06 msec
context_manager          :    386.38 msec

Testing FastRLock (0.8.4)

sequential (x1000):
lock_unlock              :      0.57 msec
reentrant_lock_unlock    :      0.35 msec
mixed_lock_unlock        :      0.53 msec
lock_unlock_nonblocking  :      0.53 msec
context_manager          :      1.65 msec

threaded 10T (x100):
lock_unlock              :    382.98 msec
reentrant_lock_unlock    :    384.84 msec
mixed_lock_unlock        :    388.89 msec
lock_unlock_nonblocking  :    396.14 msec
context_manager          :    393.30 msec

On Linux Aarch64 (ARM64):

Testing RLock (3.14.8)

sequential (x1000):
lock_unlock              :      0.43 msec
reentrant_lock_unlock    :      0.38 msec
mixed_lock_unlock        :      0.39 msec
lock_unlock_nonblocking  :      0.57 msec
context_manager          :      2.19 msec

threaded 10T (x100):
lock_unlock              :     73.74 msec
reentrant_lock_unlock    :     77.39 msec
mixed_lock_unlock        :     76.54 msec
lock_unlock_nonblocking  :     77.25 msec
context_manager          :     79.46 msec

Testing FastRLock (0.8.4)

sequential (x1000):
lock_unlock              :      0.25 msec
reentrant_lock_unlock    :      0.36 msec
mixed_lock_unlock        :      0.25 msec
lock_unlock_nonblocking  :      0.28 msec
context_manager          :      1.26 msec

threaded 10T (x100):
lock_unlock              :     78.71 msec
reentrant_lock_unlock    :     78.44 msec
mixed_lock_unlock        :     78.19 msec
lock_unlock_nonblocking  :     77.46 msec
context_manager          :     79.48 msec

Comparing Python RLock to the Cython interface of FastRLock (0.8.4):

Testing RLock (3.14.4)

sequential (x1000):
lock_unlock              :      0.22 msec
reentrant_lock_unlock    :      0.19 msec
mixed_lock_unlock        :      0.20 msec
lock_unlock_nonblocking  :      0.31 msec
context_manager          :      1.02 msec

threaded 10T (x100):
lock_unlock              :     42.05 msec
reentrant_lock_unlock    :     37.20 msec
mixed_lock_unlock        :     41.53 msec
lock_unlock_nonblocking  :     39.51 msec
context_manager          :     40.21 msec

Testing Cython interface of FastRLock (0.8.4)

sequential (x1000):
lock_unlock              :      0.07 msec
reentrant_lock_unlock    :      0.06 msec
mixed_lock_unlock        :      0.05 msec
lock_unlock_nonblocking  :      0.06 msec

threaded 10T (x100):
lock_unlock              :     38.54 msec
reentrant_lock_unlock    :     40.38 msec
mixed_lock_unlock        :     36.31 msec
lock_unlock_nonblocking  :     40.80 msec