Re: [Python-Dev] Program runs in 12s on Python 2.7, but 5s on Python 3.5 -- why so much difference?
On Mon, Jul 24, 2017 at 1:49 PM Wang, Peter Xihong < peter.xihong.w...@intel.com> wrote: > I believe we have evaluated clang vs gcc before (long time ago), and gcc > won at that time. > > > > PGO might have overshadowed impact from computed goto, and thus the latter > may no longer be needed. > Computed goto is still needed. PGO does not magically replace it. A PGO build with computed goto is faster than one without computed goto. ... as tested on gcc 4.9 a couple years ago. I doubt that has changed or changes between compilers; PGO and computed goto are different types of optimizations. -gps ___ Python-Dev mailing list Python-Dev@python.org https://mail.python.org/mailman/listinfo/python-dev Unsubscribe: https://mail.python.org/mailman/options/python-dev/archive%40mail-archive.com
Re: [Python-Dev] Program runs in 12s on Python 2.7, but 5s on Python 3.5 -- why so much difference?
I believe we have evaluated clang vs gcc before (long time ago), and gcc won at that time. PGO might have overshadowed impact from computed goto, and thus the latter may no longer be needed. When the performance difference is as large as 50%, there could be various options to nail down the root cause, including bytecode analysis. However, coming down to 3.6 sec vs 3.4 sec, a delta of ~5%, it could be hard to find out. Internally we use sampling based performance tools for micro-architecture level analysis. Or generic Linux based and open source tool “perf” is very good to use. You could also do a disassembly analysis/comparison of the object files such as the main loop, ceval.o, looking at the efficiency of the generated codes (which gives generic info regarding to Python2 and 3, but may not tell you the run time behavior with respect your specific app, pentomino.py). Hope that helps. Peter From: Ben Hoyt [mailto:benh...@gmail.com] Sent: Monday, July 24, 2017 12:35 PM To: Wang, Peter Xihong Cc: Nathaniel Smith ; Python-Dev Subject: Re: [Python-Dev] Program runs in 12s on Python 2.7, but 5s on Python 3.5 -- why so much difference? Thanks for testing. Oddly, I just tested it in Linux (Ubuntu), and get the same results as you -- Python 2.7.13 outperforms 3 (3.5.3 in my case) by a few percent. And even under a Virtualbox VM it takes 3.4 and 3.6 seconds, compared to ~5s on the host macOS operating system. Very odd. I guess that means Virtualbox is very good, and that clang/LLVM is not as good at optimizing the Python VM as gcc is. I can't find anything majorly different about my macOS Python 2 and 3 builds. Both look like they have PGO turned on (from sysconfig.get_config_vars()). Both have HAVE_COMPUTED_GOTOS=1 but USE_COMPUTED_GOTOS=0 for some reason. My Python 2 version is the macOS system version (/usr/local/bin/python2), whereas my Python3 version is from "brew install", so that's probably the difference, though still doesn't explain exactly why. -Ben On Mon, Jul 24, 2017 at 1:49 PM, Wang, Peter Xihong mailto:peter.xihong.w...@intel.com>> wrote: Hi Ben, Out of curiosity with a quick experiment, I ran your pentomino.py with 2.7.12 PGO+LTO build (Ubuntu OS 16.04.2 LTS default at /usr/bin/python), and compared with 3.7.0 alpha1 PGO+LTO (which I built a while ago), on my SkyLake processor based desktop, and 2.7 outperforms 3.7 by 3.5%. On your 2.5 GHz i7 system, I'd recommend making sure the 2 Python binaries you are comparing are in equal footings (compiled with same optimization PGO+LTO). Thanks, Peter -Original Message- From: Python-Dev [mailto:python-dev-bounces+peter.xihong.wang<mailto:python-dev-bounces%2Bpeter.xihong.wang>=intel@python.org<mailto:intel@python.org>] On Behalf Of Nathaniel Smith Sent: Tuesday, July 18, 2017 7:00 PM To: Ben Hoyt mailto:benh...@gmail.com>> Cc: Python-Dev mailto:python-dev@python.org>> Subject: Re: [Python-Dev] Program runs in 12s on Python 2.7, but 5s on Python 3.5 -- why so much difference? I'd probably start with a regular C-level profiler, like perf or callgrind. They're not very useful for comparing two versions of code written in Python, but here the Python code is the same (modulo changes in the stdlib), and it's changes in the interpreter's C code that probably make the difference. On Tue, Jul 18, 2017 at 9:03 AM, Ben Hoyt mailto:benh...@gmail.com>> wrote: > Hi folks, > > (Not entirely sure this is the right place for this question, but > hopefully it's of interest to several folks.) > > A few days ago I posted a note in response to Victor Stinner's > articles on his CPython contributions, noting that I wrote a program > that ran in 11.7 seconds on Python 2.7, but only takes 5.1 seconds on > Python 3.5 (on my 2.5 GHz macOS i7), more than 2x as fast. Obviously > this is a Good Thing, but I'm curious as to why there's so much difference. > > The program is a pentomino puzzle solver, and it works via code > generation, generating a ton of nested "if" statements, so I believe > it's exercising the Python bytecode interpreter heavily. Obviously > there have been some big optimizations to make this happen, but I'm > curious what the main improvements are that are causing this much difference. > > There's a writeup about my program here, with benchmarks at the bottom: > http://benhoyt.com/writings/python-pentomino/ > > This is the generated Python code that's being exercised: > https://github.com/benhoyt/python-pentomino/blob/master/generated_solv > e.py > > For reference, on Python 3.6 it runs in 4.6 seconds (same on Python > 3.7 alpha). This smallish increase from Python 3.5 to Python 3.6 was > more expected to me due to the bytecode changing to wordcode in 3.6. > > I tried usin
Re: [Python-Dev] Program runs in 12s on Python 2.7, but 5s on Python 3.5 -- why so much difference?
Thanks for testing. Oddly, I just tested it in Linux (Ubuntu), and get the same results as you -- Python 2.7.13 outperforms 3 (3.5.3 in my case) by a few percent. And even under a Virtualbox VM it takes 3.4 and 3.6 seconds, compared to ~5s on the host macOS operating system. Very odd. I guess that means Virtualbox is very good, and that clang/LLVM is not as good at optimizing the Python VM as gcc is. I can't find anything majorly different about my macOS Python 2 and 3 builds. Both look like they have PGO turned on (from sysconfig.get_config_vars()). Both have HAVE_COMPUTED_GOTOS=1 but USE_COMPUTED_GOTOS=0 for some reason. My Python 2 version is the macOS system version (/usr/local/bin/python2), whereas my Python3 version is from "brew install", so that's probably the difference, though still doesn't explain exactly why. -Ben On Mon, Jul 24, 2017 at 1:49 PM, Wang, Peter Xihong < peter.xihong.w...@intel.com> wrote: > Hi Ben, > > Out of curiosity with a quick experiment, I ran your pentomino.py with > 2.7.12 PGO+LTO build (Ubuntu OS 16.04.2 LTS default at /usr/bin/python), > and compared with 3.7.0 alpha1 PGO+LTO (which I built a while ago), on my > SkyLake processor based desktop, and 2.7 outperforms 3.7 by 3.5%. > On your 2.5 GHz i7 system, I'd recommend making sure the 2 Python binaries > you are comparing are in equal footings (compiled with same optimization > PGO+LTO). > > Thanks, > > Peter > > > > -Original Message- > From: Python-Dev [mailto:python-dev-bounces+peter.xihong.wang=intel.com@ > python.org] On Behalf Of Nathaniel Smith > Sent: Tuesday, July 18, 2017 7:00 PM > To: Ben Hoyt > Cc: Python-Dev > Subject: Re: [Python-Dev] Program runs in 12s on Python 2.7, but 5s on > Python 3.5 -- why so much difference? > > I'd probably start with a regular C-level profiler, like perf or > callgrind. They're not very useful for comparing two versions of code > written in Python, but here the Python code is the same (modulo changes in > the stdlib), and it's changes in the interpreter's C code that probably > make the difference. > > On Tue, Jul 18, 2017 at 9:03 AM, Ben Hoyt wrote: > > Hi folks, > > > > (Not entirely sure this is the right place for this question, but > > hopefully it's of interest to several folks.) > > > > A few days ago I posted a note in response to Victor Stinner's > > articles on his CPython contributions, noting that I wrote a program > > that ran in 11.7 seconds on Python 2.7, but only takes 5.1 seconds on > > Python 3.5 (on my 2.5 GHz macOS i7), more than 2x as fast. Obviously > > this is a Good Thing, but I'm curious as to why there's so much > difference. > > > > The program is a pentomino puzzle solver, and it works via code > > generation, generating a ton of nested "if" statements, so I believe > > it's exercising the Python bytecode interpreter heavily. Obviously > > there have been some big optimizations to make this happen, but I'm > > curious what the main improvements are that are causing this much > difference. > > > > There's a writeup about my program here, with benchmarks at the bottom: > > http://benhoyt.com/writings/python-pentomino/ > > > > This is the generated Python code that's being exercised: > > https://github.com/benhoyt/python-pentomino/blob/master/generated_solv > > e.py > > > > For reference, on Python 3.6 it runs in 4.6 seconds (same on Python > > 3.7 alpha). This smallish increase from Python 3.5 to Python 3.6 was > > more expected to me due to the bytecode changing to wordcode in 3.6. > > > > I tried using cProfile on both Python versions, but that didn't say > > much, because the functions being called aren't taking the majority of > the time. > > How does one benchmark at a lower level, or otherwise explain what's > > going on here? > > > > Thanks, > > Ben > > > > ___ > > Python-Dev mailing list > > Python-Dev@python.org > > https://mail.python.org/mailman/listinfo/python-dev > > Unsubscribe: > > https://mail.python.org/mailman/options/python-dev/njs%40pobox.com > > > > > > -- > Nathaniel J. Smith -- https://vorpus.org __ > _ > Python-Dev mailing list > Python-Dev@python.org > https://mail.python.org/mailman/listinfo/python-dev > Unsubscribe: https://mail.python.org/mailman/options/python-dev/ > peter.xihong.wang%40intel.com > ___ Python-Dev mailing list Python-Dev@python.org https://mail.python.org/mailman/listinfo/python-dev Unsubscribe: https://mail.python.org/mailman/options/python-dev/archive%40mail-archive.com
Re: [Python-Dev] Program runs in 12s on Python 2.7, but 5s on Python 3.5 -- why so much difference?
Hi Ben, Out of curiosity with a quick experiment, I ran your pentomino.py with 2.7.12 PGO+LTO build (Ubuntu OS 16.04.2 LTS default at /usr/bin/python), and compared with 3.7.0 alpha1 PGO+LTO (which I built a while ago), on my SkyLake processor based desktop, and 2.7 outperforms 3.7 by 3.5%. On your 2.5 GHz i7 system, I'd recommend making sure the 2 Python binaries you are comparing are in equal footings (compiled with same optimization PGO+LTO). Thanks, Peter -Original Message- From: Python-Dev [mailto:python-dev-bounces+peter.xihong.wang=intel@python.org] On Behalf Of Nathaniel Smith Sent: Tuesday, July 18, 2017 7:00 PM To: Ben Hoyt Cc: Python-Dev Subject: Re: [Python-Dev] Program runs in 12s on Python 2.7, but 5s on Python 3.5 -- why so much difference? I'd probably start with a regular C-level profiler, like perf or callgrind. They're not very useful for comparing two versions of code written in Python, but here the Python code is the same (modulo changes in the stdlib), and it's changes in the interpreter's C code that probably make the difference. On Tue, Jul 18, 2017 at 9:03 AM, Ben Hoyt wrote: > Hi folks, > > (Not entirely sure this is the right place for this question, but > hopefully it's of interest to several folks.) > > A few days ago I posted a note in response to Victor Stinner's > articles on his CPython contributions, noting that I wrote a program > that ran in 11.7 seconds on Python 2.7, but only takes 5.1 seconds on > Python 3.5 (on my 2.5 GHz macOS i7), more than 2x as fast. Obviously > this is a Good Thing, but I'm curious as to why there's so much difference. > > The program is a pentomino puzzle solver, and it works via code > generation, generating a ton of nested "if" statements, so I believe > it's exercising the Python bytecode interpreter heavily. Obviously > there have been some big optimizations to make this happen, but I'm > curious what the main improvements are that are causing this much difference. > > There's a writeup about my program here, with benchmarks at the bottom: > http://benhoyt.com/writings/python-pentomino/ > > This is the generated Python code that's being exercised: > https://github.com/benhoyt/python-pentomino/blob/master/generated_solv > e.py > > For reference, on Python 3.6 it runs in 4.6 seconds (same on Python > 3.7 alpha). This smallish increase from Python 3.5 to Python 3.6 was > more expected to me due to the bytecode changing to wordcode in 3.6. > > I tried using cProfile on both Python versions, but that didn't say > much, because the functions being called aren't taking the majority of the > time. > How does one benchmark at a lower level, or otherwise explain what's > going on here? > > Thanks, > Ben > > ___ > Python-Dev mailing list > Python-Dev@python.org > https://mail.python.org/mailman/listinfo/python-dev > Unsubscribe: > https://mail.python.org/mailman/options/python-dev/njs%40pobox.com > -- Nathaniel J. Smith -- https://vorpus.org ___ Python-Dev mailing list Python-Dev@python.org https://mail.python.org/mailman/listinfo/python-dev Unsubscribe: https://mail.python.org/mailman/options/python-dev/peter.xihong.wang%40intel.com ___ Python-Dev mailing list Python-Dev@python.org https://mail.python.org/mailman/listinfo/python-dev Unsubscribe: https://mail.python.org/mailman/options/python-dev/archive%40mail-archive.com
Re: [Python-Dev] Program runs in 12s on Python 2.7, but 5s on Python 3.5 -- why so much difference?
I'd probably start with a regular C-level profiler, like perf or callgrind. They're not very useful for comparing two versions of code written in Python, but here the Python code is the same (modulo changes in the stdlib), and it's changes in the interpreter's C code that probably make the difference. On Tue, Jul 18, 2017 at 9:03 AM, Ben Hoyt wrote: > Hi folks, > > (Not entirely sure this is the right place for this question, but hopefully > it's of interest to several folks.) > > A few days ago I posted a note in response to Victor Stinner's articles on > his CPython contributions, noting that I wrote a program that ran in 11.7 > seconds on Python 2.7, but only takes 5.1 seconds on Python 3.5 (on my 2.5 > GHz macOS i7), more than 2x as fast. Obviously this is a Good Thing, but I'm > curious as to why there's so much difference. > > The program is a pentomino puzzle solver, and it works via code generation, > generating a ton of nested "if" statements, so I believe it's exercising the > Python bytecode interpreter heavily. Obviously there have been some big > optimizations to make this happen, but I'm curious what the main > improvements are that are causing this much difference. > > There's a writeup about my program here, with benchmarks at the bottom: > http://benhoyt.com/writings/python-pentomino/ > > This is the generated Python code that's being exercised: > https://github.com/benhoyt/python-pentomino/blob/master/generated_solve.py > > For reference, on Python 3.6 it runs in 4.6 seconds (same on Python 3.7 > alpha). This smallish increase from Python 3.5 to Python 3.6 was more > expected to me due to the bytecode changing to wordcode in 3.6. > > I tried using cProfile on both Python versions, but that didn't say much, > because the functions being called aren't taking the majority of the time. > How does one benchmark at a lower level, or otherwise explain what's going > on here? > > Thanks, > Ben > > ___ > Python-Dev mailing list > Python-Dev@python.org > https://mail.python.org/mailman/listinfo/python-dev > Unsubscribe: > https://mail.python.org/mailman/options/python-dev/njs%40pobox.com > -- Nathaniel J. Smith -- https://vorpus.org ___ Python-Dev mailing list Python-Dev@python.org https://mail.python.org/mailman/listinfo/python-dev Unsubscribe: https://mail.python.org/mailman/options/python-dev/archive%40mail-archive.com
Re: [Python-Dev] Program runs in 12s on Python 2.7, but 5s on Python 3.5 -- why so much difference?
Thanks, Nick -- that's interesting. I just saw the extra JUMP_FORWARD and JUMP_ABSOLUTE instructions on my commute home (I guess those are something Python 3.x optimizes away). VERY strangely, on Windows Python 2.7 is faster! Comparing 64-bit Python 2.7.12 against Python 3.5.3 on my Windows 10 laptop: * Python 2.7.12: 4.088s * Python 3.5.3: 5.792s I'm pretty sure MSVC/Windows doesn't support computed gotos, but that doesn't explain why 3.5 is so much faster than 2.7 on Mac. I have yet to try it on Linux. -Ben On Tue, Jul 18, 2017 at 9:35 PM, Nick Coghlan wrote: > On 19 July 2017 at 02:18, Antoine Pitrou wrote: > > On Tue, 18 Jul 2017 12:03:36 -0400 > > Ben Hoyt wrote: > >> The program is a pentomino puzzle solver, and it works via code > generation, > >> generating a ton of nested "if" statements, so I believe it's exercising > >> the Python bytecode interpreter heavily. > > > > A first step would be to see if the generated bytecode has changed > > substantially. > > Scanning over them, the Python 2.7 bytecode appears to have many more > JUMP_FORWARD and JUMP_ABSOLUTE opcodes than appear in the 3.6 version > (I didn't dump them into a Counter instance to tally them properly > though, since 2.7's dis module is missing the structured opcode > iteration APIs). > > With the shift to wordcode, the overall size of the bytecode is also > significantly *smaller*: > > >>> len(co.co_consts[0].co_code) # 2.7 > 14427 > > >>> len(co.co_consts[0].co_code) # 3.6 > 11850 > > However, I'm not aware of any Python profilers that currently offer > opcode level profiling - the closest would probably be VMProf's JIT > profiling, and that aspect of VMProf is currently PyPy specific > (although could presumably be extended to CPython 3.6+ by way of the > opcode evaluation hook). > > Cheers, > Nick. > > -- > Nick Coghlan | ncogh...@gmail.com | Brisbane, Australia > ___ > Python-Dev mailing list > Python-Dev@python.org > https://mail.python.org/mailman/listinfo/python-dev > Unsubscribe: https://mail.python.org/mailman/options/python-dev/ > benhoyt%40gmail.com > ___ Python-Dev mailing list Python-Dev@python.org https://mail.python.org/mailman/listinfo/python-dev Unsubscribe: https://mail.python.org/mailman/options/python-dev/archive%40mail-archive.com
Re: [Python-Dev] Program runs in 12s on Python 2.7, but 5s on Python 3.5 -- why so much difference?
On 19 July 2017 at 02:18, Antoine Pitrou wrote: > On Tue, 18 Jul 2017 12:03:36 -0400 > Ben Hoyt wrote: >> The program is a pentomino puzzle solver, and it works via code generation, >> generating a ton of nested "if" statements, so I believe it's exercising >> the Python bytecode interpreter heavily. > > A first step would be to see if the generated bytecode has changed > substantially. Scanning over them, the Python 2.7 bytecode appears to have many more JUMP_FORWARD and JUMP_ABSOLUTE opcodes than appear in the 3.6 version (I didn't dump them into a Counter instance to tally them properly though, since 2.7's dis module is missing the structured opcode iteration APIs). With the shift to wordcode, the overall size of the bytecode is also significantly *smaller*: >>> len(co.co_consts[0].co_code) # 2.7 14427 >>> len(co.co_consts[0].co_code) # 3.6 11850 However, I'm not aware of any Python profilers that currently offer opcode level profiling - the closest would probably be VMProf's JIT profiling, and that aspect of VMProf is currently PyPy specific (although could presumably be extended to CPython 3.6+ by way of the opcode evaluation hook). Cheers, Nick. -- Nick Coghlan | ncogh...@gmail.com | Brisbane, Australia ___ Python-Dev mailing list Python-Dev@python.org https://mail.python.org/mailman/listinfo/python-dev Unsubscribe: https://mail.python.org/mailman/options/python-dev/archive%40mail-archive.com
Re: [Python-Dev] Program runs in 12s on Python 2.7, but 5s on Python 3.5 -- why so much difference?
On Tue, 18 Jul 2017 12:03:36 -0400 Ben Hoyt wrote: > Hi folks, > > (Not entirely sure this is the right place for this question, but hopefully > it's of interest to several folks.) > > A few days ago I posted a note in response to Victor Stinner's articles on > his CPython contributions, noting that I wrote a program that ran in 11.7 > seconds on Python 2.7, but only takes 5.1 seconds on Python 3.5 (on my 2.5 > GHz macOS i7), more than 2x as fast. Obviously this is a Good Thing, but > I'm curious as to why there's so much difference. > > The program is a pentomino puzzle solver, and it works via code generation, > generating a ton of nested "if" statements, so I believe it's exercising > the Python bytecode interpreter heavily. A first step would be to see if the generated bytecode has changed substantially. Otherwise, you can try to comment out parts of the function until the performance difference has been nullified. Regards Antoine. ___ Python-Dev mailing list Python-Dev@python.org https://mail.python.org/mailman/listinfo/python-dev Unsubscribe: https://mail.python.org/mailman/options/python-dev/archive%40mail-archive.com
[Python-Dev] Program runs in 12s on Python 2.7, but 5s on Python 3.5 -- why so much difference?
Hi folks, (Not entirely sure this is the right place for this question, but hopefully it's of interest to several folks.) A few days ago I posted a note in response to Victor Stinner's articles on his CPython contributions, noting that I wrote a program that ran in 11.7 seconds on Python 2.7, but only takes 5.1 seconds on Python 3.5 (on my 2.5 GHz macOS i7), more than 2x as fast. Obviously this is a Good Thing, but I'm curious as to why there's so much difference. The program is a pentomino puzzle solver, and it works via code generation, generating a ton of nested "if" statements, so I believe it's exercising the Python bytecode interpreter heavily. Obviously there have been some big optimizations to make this happen, but I'm curious what the main improvements are that are causing this much difference. There's a writeup about my program here, with benchmarks at the bottom: http://benhoyt.com/writings/python-pentomino/ This is the generated Python code that's being exercised: https://github.com/benhoyt/python-pentomino/blob/master/generated_solve.py For reference, on Python 3.6 it runs in 4.6 seconds (same on Python 3.7 alpha). This smallish increase from Python 3.5 to Python 3.6 was more expected to me due to the bytecode changing to wordcode in 3.6. I tried using cProfile on both Python versions, but that didn't say much, because the functions being called aren't taking the majority of the time. How does one benchmark at a lower level, or otherwise explain what's going on here? Thanks, Ben ___ Python-Dev mailing list Python-Dev@python.org https://mail.python.org/mailman/listinfo/python-dev Unsubscribe: https://mail.python.org/mailman/options/python-dev/archive%40mail-archive.com