Repository navigation
re._compiled_typed's lru_cache causes significant degradation of the mako_v2 bench #60593
Description
Activity
bpo-9396 replaced a few caches in the stdlib w/ lru_cache, this made the mako_v2 benchmark on Python 3 almost 3x slower than 2.7
The benchmark results are good now that Mako was changed to cache the re itself, but the problem still stands that lru_cache seems to hurt the perf of inline res compared to 2.7. The fix for Mako did not affect the 2.7 benchmark numbers
See more info here:
http://mail.python.org/pipermail/python-dev/2012-November/122521.html
lru_cache() seems to use a complicated make_key() function, which is invoked on each cache hit. The LRU logic is probably on the slow side too, compared to a hand-coded logic which would favour lookup cost over insertion / eviction cost.
Would be interesting to know what speed difference would occur if the statistics gathering was optional and turned off.
As for _make_key(), I wonder if
(args, tuple(sorted(kwd.items())))as a key would be any faster as a tuple's hash is derived from its contents and not the tuple itself (if I remember correctly). You could even special case when len(kwds) == 0 to skip the sorted(kwd.items()) overhead if it is worth it performance-wise.Ditching the statistics only sped up regex_compile by 2%.
Ditching the statistics only sped up regex_compile by 2%.
Does explicit compiling even go through the cache?
Regardless, the issue here is with performance of cache hits, not cache
misses. By construction, you cache something which is costly to compute,
so the overhead of a cache miss won't be very noticeable.re.compile() calls _compile() which has the lru_cache decorator so it will trigger it. But you make a good point, Antoine, that it's the hit overhead here that we care about as long as misses don't get worse as the calculation of is to be cached should overwhelm anything the LRU does.
With a simplified _make_key() I can get regex_compile w/ cache clearing turned off to be 1.28x faster by making it be::
if not typed: if len(kwds) == 0: return args, () else: return args, tuple(sorted(kwds.items())) else: if len(kwds) == 0: return (tuple((type(arg), arg) for arg in args), ()) else: return (tuple((type(arg), arg) for arg in args), tuple((type(v), (k, v)) for k, v in kwds.items()))
That might not be the fastest way to handle keyword arguments (since regex_compile w/ caching and leaving out the len(kwds) trick out becomes 1.13x slower), but at least for the common case of positional arguments it seems faster and the code is easier to read IMO.
Did you try moving the existing single-argument fast path to before the main if statement in _make_key? That is:
if not kwds and len(args) == 1: key = args[0] key_type = type(key) if key_type in fasttypes: if typed: return key, key_type return key
Such a special case is already present, but it's *after* a lot of the other processing *and* it doesn't fire when typed==True.
So instead of the simple 2-tuple creation above, you instead do the relatively wasteful:
args + tuple(type(v) for v in args)re.compile() calls _compile() which has the lru_cache decorator so it
will trigger it.What's the point of using the lru_cache for compiled regexes?
Unless I'm missing something, re.compile() should just return the compiled regex without going though lru_cache and needlessly wasting time and cache's slots.zzzeek commented
on Nov 3, 2012 zzzeekmannequinMannequinin response to ezio, I poked around the source here, since I've never been sure if re.compile() cached its result or not. It seems to be the case in 2.7 and 3.2 also - 2.7 uses a local caching scheme and 3.2 uses functools.lru_cache, yet we don't see as much of a slowdown with 3.2.
so it seems like the caching behavior is precedent here, but I would revert re.py's caching scheme to the one used in 2.7 if the functools.lru_cache can't be sped up very significantly. ideally lru_cache would be native.
also does python include any kind of benchmarking unit tests ? over in SQLA we have an approach that fails if the call-counts of various functions, as measured by cProfile, fall outside of a known range. it's caught many issues like these for me.
Now that Brett has a substantial portion of the benchmark suite running on Py3k, we should see a bit more progress on the PyPy-inspired speed.python.org project (which should make it much easier to catch this kind of regression before it hits a production release).
In this case, as I noted in my earlier comment, I think the 3.3 changes to make_key broke an important single-argument fast path that the re module was previously relying on, thus the major degradation in performance on a cache hit. I haven't looked into setting up the benchmark suite on my own machine though, so we won't know for sure until either I get around to doing that, or someone with it already set up tries the change I suggested above.
This is not only 3.3 regression, this is also 3.2 regression. 3.1, 3.2 and 3.3 have different caching implementation.
Mikrobenchmark:
$ ./python -m timeit -s "import re" "re.match('', '')"Results:
3.1: 2.61 usec per loop
3.2: 5.77 usec per loop
3.3: 11.8 usec per loopHere is a patch which reverts 3.1 implementation (and adds some optimization).
Microbenchmark:
$ ./python -m timeit -s "import re" "re._compile('', 0)"Results:
3.1: 1.45 usec per loop
3.2: 4.45 usec per loop
3.3: 9.91 usec per loop
3.4patched: 0.89 usec per loopAttached a proof of concept that removes the caching for re.compile, as suggested in msg174599.
5 remaining items
Since switching from a simple custom cache to the generalized lru cache made a major slowdown, I think the change should be reverted. A dict + either occasional clearing or a circular queue and a first-in, first-out discipline is quite sufficient. There is no need for the extra complexity of a last-used, first out discipline.
For 3.4 bpo-14373 might solve the issue.
A few thoughts:
-
The LRU cache was originally intended for IO bound calls not for tight, frequently computationally bound calls like re.compile.
-
The Py3.3 version of lru_cache() favors size optimizations (i.e. it uses only one dictionary instead of the two used by OrderedDict and keyword arguments are flattened into a single list instead of a nested structure). Also, the 3.3 version assures that __hash__ is not called more than one for a given key (this change helps objects that have a slow hash function and it helps solve a reentrancy problem with recursive cached function calls). The cost of these changes is that _make_key is slower than it was before.
-
I had hoped to get in a C version of _make_key before Py3.3 went out but I didn't have time. Going forward, the lru_cache() will likely have a C-implementation that is blindingly fast.
-
For the re module, it might make sense to return to custom logic in the re modue that implements size limited caching without the overhead of 1) LRU logic, 2) general purpose argument handling, 3) reentrancy or locking logic, and 4) without statistics tracking.
-
Until the lru_cache can be sped-up significantly, I recommend just accepting Serhiy's patch to go back to 3.2 logic in the regex module.
In the meantime, I'll continue to work on improving speed of _make_key().
Raymond's plan sounds good to me.
We may also want to tweak the 3.3 lru_cache docs to note the trade-offs involved in using it. Perhaps something like:
"As a general purpose cache, lru_cache needs to be quite pessimistic in deriving non-conflicting keys from the supplied arguments. When caching the results of CPU-bound calculations, the cost of deriving non-conflicting keys may need be assessed against the typical cost of the underlying calculation."
Which does give me a thought - perhaps lru_cache in 3.4 could accept a "key" argument that is called as "key(*args, **kwds)" to derive the cache key? (that would be a separate issue, of course)
Serhiy, please go ahead an apply your patch. Be sure to restore the re cache tests that existed in Py3.2 as well.
Thank you.
Raymond, actually my patch reverts 3.1 logic. lru_cache used since 3.2.
There are no any additional re cache tests in 3.2 or 3.1.
sbt commented
on Mar 11, 2013 sbtmannequinMannequinWhich does give me a thought - perhaps lru_cache in 3.4 could accept a
"key" argument that is called as "key(*args, **kwds)" to derive the cache
key? (that would be a separate issue, of course)Agreed. I suggested the same in an earlier post.
python-dev commented
on Mar 16, 2013 python-devmannequinMannequinNew changeset 6951d7b8d3ad by Serhiy Storchaka in branch '3.2':
Issue bpo-16389: Fixed an issue number in previos commit.
http://hg.python.org/cpython/rev/6951d7b8d3adNew changeset 7b737011d822 by Serhiy Storchaka in branch '3.3':
Issue bpo-16389: Fixed an issue number in previos commit.
http://hg.python.org/cpython/rev/7b737011d822New changeset 6898e1afc216 by Serhiy Storchaka in branch 'default':
Issue bpo-16389: Fixed an issue number in previos commit.
http://hg.python.org/cpython/rev/6898e1afc216
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: