Repository navigation
BytesIO copy-on-write #66202
Description
| BPO | 22003 |
|---|---|
| Nosy | @pitrou, @scoder, @benjaminp, @skrah, @hynek, @dw, @serhiy-storchaka |
| Files |
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:
assignee = None
closed_at = <Date 2014-07-29.23:46:30.037>
created_at = <Date 2014-07-17.22:25:37.100>
labels = ['library', 'performance']
title = 'BytesIO copy-on-write'
updated_at = <Date 2015-03-04.21:04:29.471>
user = 'https://github.com/dw'bugs.python.org fields:
activity = <Date 2015-03-04.21:04:29.471>
actor = 'dw'
assignee = 'none'
closed = True
closed_date = <Date 2014-07-29.23:46:30.037>
closer = 'pitrou'
components = ['Library (Lib)']
creation = <Date 2014-07-17.22:25:37.100>
creator = 'dw'
dependencies = []
files = ['35988', '36004', '36005', '36016', '36078', '36137', '38058']
hgrepos = []
issue_num = 22003
keywords = ['patch']
message_count = 37.0
messages = ['223383', '223385', '223386', '223401', '223402', '223522', '223526', '223542', '223581', '223582', '223583', '223586', '223588', '223599', '223600', '223611', '223633', '223681', '223682', '223683', '223692', '223707', '223907', '223962', '224121', '224139', '224164', '224168', '224169', '224238', '224273', '224274', '235596', '235619', '236002', '237207', '237212']
nosy_count = 11.0
nosy_names = ['pitrou', 'scoder', 'benjamin.peterson', 'stutzbach', 'skrah', 'python-dev', 'hynek', 'piotr.dobrogost', 'dw', 'serhiy.storchaka', 'kmike']
pr_nums = []
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'performance'
url = 'https://bugs.python.org/issue22003'
versions = ['Python 3.5']
Activity
dw commentedon Jul 17, 2014
This is a followup to the thread at https://mail.python.org/pipermail/python-dev/2014-July/135543.html , discussing the existing behaviour of BytesIO copying its source object, and how this regresses compared to cStringIO.StringI.
The goal of posting the patch on list was to try and stimulate discussion around the approach. The patch itself obviously isn't ready for review, and I'm not in a position to dedicate time to it just now (although in a few weeks I'd love to give it full attention!).
Ignoring this quick implementation, are there any general comments around the approach?
My only concern is that it might keep large objects alive in a non-intuitive way in certain circumstances, though I can't think of any obvious ones immediately.
Also interested in comments on the second half of that thread: "a natural extension of this is to do something very similar on the write side: instead of generating a temporary private heap allocation, generate (and freely resize) a private PyBytes object until it is exposed to the user, at which point, _getvalue() returns it, and converts its into an IO_SHARED buffer."
There are quite a few interactions with making that work correctly, in particular:
How BytesIO would implement the buffers interface without causing the under-construction Bytes to become readonly
Avoiding redundant copies and resizes -- we can't simply tack 25% slack on the end of the Bytes and then truncate it during getvalue() without likely triggering a copy and move, however with careful measurement of allocator behavior there are various tradeoffs that could be made - e.g. obmalloc won't move a <500 byte allocation if it shrinks by <25%. glibc malloc's rules are a bit more complex though.
Could also add a private _PyBytes_SetSize() API to allow truncation to the final size during getvalue() without informing the allocator. Then we'd simply overallocate by up to 10% or 1-2kb, and write off the loss of the slack space.
Notably, this approach completely differs from the one documented in http://bugs.python.org/issue15381 .. it's not clear to me which is better.
dw commentedon Jul 17, 2014
Submitted contributor agreement. Please consider the demo patch licensed under the Apache 2 licence.
pitrou commentedon Jul 17, 2014
Be careful what happens when the original object is mutable:
I don't know what your patch does in this case.
dw commentedon Jul 18, 2014
Good catch :( There doesn't seem to be way a to ask for an immutable buffer, so perhaps it could just be a little more selective. I think the majority of use cases would still be covered if the sharing behaviour was restricted only to BytesType.
In that case "Py_buffer initialdata" could become a PyObject*, saving a small amount of memory, and allowing reuse of the struct member if BytesIO was also modified to directly write into a private BytesObject
scoder commentedon Jul 18, 2014
Even if there is no way to explicitly request a RO buffer, the Py_buffer struct that you get back actually tells you if it's read-only or not. Shouldn't that be enough to enable this optimisation?
Whether or not implementors of the buffer protocol set this flag correctly is another question, but if not then they need fixing on their side anyway. (And in the vast majority of cases, the implementor will be either CPython or NumPy.)
Also, generally speaking, I think such an optimisation would be nice, even if it only catches some common cases (and doesn't break the others :). It could still copy data if necessary, but try to avoid it if possible.
dw commentedon Jul 20, 2014
This version is tidied up enough that I think it could be reviewed.
Changes are:
Defer `buf' allocation until __init__, rather than __new__ as was previously done. Now upon completion, BytesIO.__new__ returns a valid, closed BytesIO, whereas previously a valid, empty, open BytesIO was returned. Is this interface change permissible?
Move __init__ guts into a "reinit()", for sharing with __setstate__, which also previously caused an unnecessary copy. Additionally gather up various methods for deallocating buffers into a single "reset()" function, called by reinit(), _dealloc(), and _close()
Per Stefan's suggested approach, reinit() now explicitly checks for a read-only buffer, falling back to silently performing a copy if the returned buffer is read-write. That seems vastly preferable to throwing an exception, which would probably be another interface change.
Call
unshare()any place the buffer is about to be modified. If the buffer needs to be made private, it also accepts a size hint indicating how much less/more space the subsequent operation needs, to avoid a redundant reallocation after the unsharing.Outstanding issues:
dw commentedon Jul 20, 2014
New patch also calls unshare() during getbuffer()
dw commentedon Jul 20, 2014
I'm not sure the "read only buffer" test is strong enough: having a readonly view is not a guarantee that the data in the view cannot be changed through some other means, i.e. it is read-only, not immutable.
Pretty sure this approach is broken. What about the alternative approach of specializing for Bytes?
pitrou commentedon Jul 21, 2014
That would certainly sound good enough, to optimize the common case.
Also, it would be nice if you could add some tests to the patch (e.g. to stress the bytearray case). Thank you!
pitrou commentedon Jul 21, 2014
As for whether the "checking for a readonly view" approach is broken, I don't know: that part of the buffer API is still mysterious to me. Stefan, would you have some insight?
skrah commentedon Jul 21, 2014
I think checking for a readonly view is fine. The protocol is this:
Use the PyBUF_WRITABLE flag in the request. Then the provider must
either have a writable buffer or else deny the request entirely.
Omit the PyBUF_WRITABLE flag in the request. Then the provider can
return a writable or a readonly buffer, but must set the readonly flag
correctly AND export the same type of buffer to ALL consumers.
It is not possible to ask for a readonly buffer explicitly, but the
readonly flag in the Py_Buffer struct should always be set correctly.
It is hard to guess the original intention of the PEP-3118 authors, but
in practice "readonly" means "immutable" here. IMO a buffer provider would
be seriously broken if a readonly buffer is mutated in any way.
skrah commentedon Jul 21, 2014
The original wording in the PEP is this:
readonly
--------
an integer variable to hold whether or not the memory is readonly. 1
means the memory is readonly, zero means the memory is writable.
To me this means that a hypothetical compiler that could figure
out at compile time that the readonly flag is set would be allowed
to put the buffer contents into the read-only data section.
dw commentedon Jul 21, 2014
Stefan,
Thanks for digging here. As much as I'd love to follow this interpretation, it simply doesn't match existing buffer implementations, including within the standard library.
For example, mmap.mmap(..., flags=mmap.MAP_SHARED, prot=mmap.PROT_READ) will produce a read-only buffer, yet mutability is entirely at the whim of the operating system. In this case, "immutability" may be apparent for years, until some machine has memory pressure, causing the shared mapping to be be flushed, and refreshed from (say, incoherent NFS storage) on next access.
I thought it would be worth auditing some of the most popular types of buffer just to check your interpretation, and this was the first, most obvious candidate.
Any thoughts? I'm leaning heavily toward the Bytes specialization approach
11 remaining items
dw commentedon Jul 27, 2014
Hey Antoine,
Thanks for the link. I'm having trouble getting reproducible results at present, and running out of ideas as to what might be causing it. Even after totally isolating a CPU for e.g. django_v2 and with frequency scaling disabled, numbers still jump around for the same binary by as much as 3%.
I could not detect any significant change between runs of the old and new binary that could not be described as noise, given the isolation issues above.
pitrou commentedon Jul 27, 2014
That's expected. If the difference doesn't go above 5-10%, then you IMO can pretty much consider your patch didn't have any impact on those benchmarks.
dw commentedon Jul 28, 2014
Newest patch incorporates Antoine's review comments. The final benchmark results are below. Just curious, what causes e.g. telco to differ up to 7% between runs? That's really huge
Report on Linux k2 3.14-1-amd64 #1 SMP Debian 3.14.9-1 (2014-06-30) x86_64
Total CPU cores: 4
### call_method_slots ###
Min: 0.329869 -> 0.340487: 1.03x slower
Avg: 0.330512 -> 0.341786: 1.03x slower
Significant (t=-216.69)
Stddev: 0.00067 -> 0.00060: 1.1111x smaller
### call_method_unknown ###
Min: 0.351167 -> 0.343961: 1.02x faster
Avg: 0.351731 -> 0.344580: 1.02x faster
Significant (t=238.89)
Stddev: 0.00033 -> 0.00040: 1.2271x larger
### call_simple ###
Min: 0.257487 -> 0.277366: 1.08x slower
Avg: 0.257942 -> 0.277809: 1.08x slower
Significant (t=-845.64)
Stddev: 0.00029 -> 0.00029: 1.0126x smaller
### etree_generate ###
Min: 0.377985 -> 0.365952: 1.03x faster
Avg: 0.381797 -> 0.369452: 1.03x faster
Significant (t=31.15)
Stddev: 0.00314 -> 0.00241: 1.3017x smaller
### etree_iterparse ###
Min: 0.545668 -> 0.565437: 1.04x slower
Avg: 0.554188 -> 0.576807: 1.04x slower
Significant (t=-17.00)
Stddev: 0.00925 -> 0.00956: 1.0340x larger
### etree_process ###
Min: 0.294158 -> 0.286617: 1.03x faster
Avg: 0.296354 -> 0.288877: 1.03x faster
Significant (t=36.22)
Stddev: 0.00149 -> 0.00143: 1.0435x smaller
### fastpickle ###
Min: 0.458961 -> 0.475828: 1.04x slower
Avg: 0.460226 -> 0.481228: 1.05x slower
Significant (t=-109.38)
Stddev: 0.00082 -> 0.00173: 2.1051x larger
### nqueens ###
Min: 0.305883 -> 0.295858: 1.03x faster
Avg: 0.308085 -> 0.297755: 1.03x faster
Significant (t=90.22)
Stddev: 0.00077 -> 0.00085: 1.0942x larger
### silent_logging ###
Min: 0.074152 -> 0.075818: 1.02x slower
Avg: 0.074345 -> 0.076005: 1.02x slower
Significant (t=-96.29)
Stddev: 0.00013 -> 0.00012: 1.0975x smaller
### spectral_norm ###
Min: 0.355738 -> 0.364419: 1.02x slower
Avg: 0.356691 -> 0.365764: 1.03x slower
Significant (t=-126.23)
Stddev: 0.00054 -> 0.00047: 1.1533x smaller
### telco ###
Min: 0.012152 -> 0.013038: 1.07x slower
Avg: 0.012264 -> 0.013157: 1.07x slower
Significant (t=-83.98)
Stddev: 0.00008 -> 0.00007: 1.0653x smaller
The following not significant results are hidden, use -v to show them:
2to3, call_method, chaos, django_v2, etree_parse, fannkuch, fastunpickle, float, formatted_logging, go, hexiom2, iterative_count, json_dump, json_dump_v2, json_load, mako, mako_v2, meteor_contest, nbody, normal_startup, pathlib, pickle_dict, pickle_list, pidigits, raytrace, regex_compile, regex_effbot, regex_v8, richards, simple_logging, startup_nosite, threaded_count, tornado_http, unpack_sequence, unpickle_list.
skrah commentedon Jul 28, 2014
telco.py always varies a lot between runs (up to 10%), even in the
big version "telco.py full":
http://bytereef.org/mpdecimal/quickstart.html#telco-benchmark
Using the average of 10 runs, I can't really see a slowdown.
skrah commentedon Jul 28, 2014
So I wonder why the benchmark suite says that the telco slowdown is significant. :)
dw commentedon Jul 29, 2014
I suspect it's all covered now, but is there anything else I can help with to get this patch pushed along its merry way?
python-dev commentedon Jul 29, 2014
New changeset 79a5fbe2c78f by Antoine Pitrou in branch 'default':
Issue bpo-22003: When initialized from a bytes object, io.BytesIO() now
http://hg.python.org/cpython/rev/79a5fbe2c78f
pitrou commentedon Jul 29, 2014
The latest patch is good indeed. Thank you very much!
kmike commentedon Feb 9, 2015
Shouldn't this fix be mentioned in https://docs.python.org/3.5/whatsnew/3.5.html#optimizations ?
dw commentedon Feb 9, 2015
Attached trivial patch for whatsnew.rst.
python-dev commentedon Feb 14, 2015
New changeset 7ae156f07a90 by Berker Peksag in branch 'default':
Add a whatsnew entry for issue bpo-22003.
https://hg.python.org/cpython/rev/7ae156f07a90
piotrdobrogost commentedon Mar 4, 2015
Why does it abandon buffer interface? Because of the following?
Shouldn't existing buffer implementations be fixed then and this feature made to use buffer interface instead of specialize for Bytes? If so is there at least any information on this in the comments so that one wouldn't wonder why there is specialization instead of relaying on buffer interface?
dw commentedon Mar 4, 2015
Hi Piotr,
There wasn't an obvious fix that didn't involve changing the buffer interface itself. There is presently ambiguity in the interface regarding the difference between a "read only" buffer and an "immutable" buffer, which is crucial for its use in this case.
Fixing the interface, followed by every buffer interface user, is a significantly more complicated task than simply optimizing for the most common case, as done here. FWIW I still think this work is worth doing, though I personally don't have time to approach it just now.
We could have (and possibly should) approach fixing e.g. mmap.mmap() hashability, possibly causing user code regressions, but even if such cases were fixed it still wouldn't be a enough to rely on for the optimization implemented here.