New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Use deque instead of list the threading.Condition waiter queue #61587
Comments
Condition variables implement a FIFO queue for waiting threads. The current implementation uses a regular Python list but could use a deque instead. A wait() call appends a new waiter. A notify() call removes the oldest waiter; this is an O(n) operation on list but only an O(1) operation on deques. A notify_all() call is O(n**2) for a list but only O(n) for a deque. If there is interest in this patch, I can add slicing support to collections.deque so that this patch won't need itertools.islice() |
I don't think you need slicing if you rewrite the patch in another way, e.g.: for i in range(n):
try:
waiter = __waiters.popleft()
except IndexError:
break
waiter.release() I think this is safe, since notify() must be called with the lock held: another thread shouldn't be able to mutate the waiters list in the meantime. As for notify_all(), it could be optimized to swap the internal list with an empty one: there's no need to pop the waiters one by one. |
Tim, do you remember why Condition.notify() went to great lengths to act as if the lock could be released after the check for self._is_owned()? It loops over its own a copy of __waiters, and the __waiters.remove(waiter) code is wrapped in a try/except to detect a situation where __waiters mutated during the release-loop. I'm presuming that defensive programming was put there for a reason. |
Actually, wait() calls self._waiters.remove() without holding the lock. But I think it could easily do so after taking the lock (since it takes it anyway before returning). Also, _waiters should better be a set, since wait() needs the associative behaviour when unregistering a waiter. notify() would then look like: for i in range(n):
try:
waiter = self._waiters.pop()
except KeyError:
break
waiter.release() and wait() would look like: waiter = _allocate_lock()
waiter.acquire()
self._waiters.add(waiter)
self._release_save()
try:
return waiter.acquire(timeout)
finally:
self._acquire_restore()
self._waiters.discard(waiter) |
That said, I seem to remember a discussion of Condition's fairness. Also, I can't remember a situation where I made an intensive use of a Condition (say, hundreds of calls per second), as opposed to Lock and RLock which can be heavily invoked to protect the integrity of critical data structures. |
Thanks Antoine. Since the calls are made without a lock, I'll go for a minimal patch and keep the existing fairness logic. Adding Guido to the nosy list since this is his code. FWIW, the heaviest load for condition variables likely arises with use of the Queue module which implements substantially all of its logic around three condition variables and a single lock. |
New changeset 0f86b51f8f8b by Raymond Hettinger in branch 'default': |
Looks fine. I'd say that it would be great to add slicing to deque! There's one oddity in the code, but it was there before the patch -- the local variable name __waiters is pretty silly. It appears to be a micro-optimization to avoid using self._waiters more than once; even if that is worth it (I kind of doubt it), the __private name is wrong and misguided. |
Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.
Show more details
GitHub fields:
bugs.python.org fields:
The text was updated successfully, but these errors were encountered: