[Python-Dev] Profile Guided Optimization active by-default

Patrascu, Alecsandru alecsandru.patrascu at intel.com
Sat Aug 22 18:58:12 CEST 2015

This target replaces the existing one in the CPython Makefile, which now uses a quick run of pybench and the obtained binary does not perform well on general Python loads. I don't think is a good idea to add a by-default target that does PGO on dedicated workloads, like Django, because then it will perform better on that particular load and poorly on other. 

Of course, if any user has a dedicated workload for which he or she want to get the best benefit over PGO, it will have to run that training separately from the proposed one. Our proposal targets the broader audience that uses Python in various scenarios, and they will see an overall improvement after compiling Python from sources.


From: Brett Cannon [mailto:brett at python.org] 
Sent: Saturday, August 22, 2015 7:25 PM
To: guido at python.org; Patrascu, Alecsandru
Cc: python-dev at python.org
Subject: Re: [Python-Dev] Profile Guided Optimization active by-default

On Sat, Aug 22, 2015, 09:17 Guido van Rossum <guido at python.org> wrote:
How about we first add a new Makefile target that enables PGO, without turning it on by default? Then later we can enable it by default.

I agree. Updating the Makefile so it's easier to use PGO is great, but we should do a release with it as opt-in and go from there.
Also, I have my doubts about regrtest. How sure are we that it represents a typical Python load? Tests are often using a different mix of operations than production code.
That was also my question. You said that "it provides the best performance improvement", but compared to what; what else was tried? And what difference does it make to e.g. a Django app that is trained on their own simulated workload compared to using regrtest? IOW is regrtest displaying the best across-the-board performance because it stresses the largest swath of Python and thus catches generic patterns in the code but individuals could get better performance with a simulated workload?

On Sat, Aug 22, 2015 at 7:46 AM, Patrascu, Alecsandru <alecsandru.patrascu at intel.com> wrote:
Hi All,
This is Alecsandru from Server Scripting Languages Optimization team at Intel Corporation.
I would like to submit a request to turn-on Profile Guided Optimization or PGO as the default build option for Python (both 2.7 and 3.6), given its performance benefits on a wide variety of workloads and hardware.  For instance, as shown from attached sample performance results from the Grand Unified Python Benchmark, >20% speed up was observed.  In addition, we are seeing 2-9% performance boost from OpenStack/Swift where more than 60% of the codes are in Python 2.7. Our analysis indicates the performance gain was mainly due to reduction of icache misses and CPU front-end stalls.
Attached is the Makefile patches that modify the all build target and adds a new one called "disable-profile-opt". We built and tested this patch for Python 2.7 and 3.6 on our Linux machines (CentOS 7/Ubuntu Server 14.04, Intel Xeon Haswell/Broadwell with 18/8 cores).  We use "regrtest" suite for training as it provides the best performance improvement.  Some of the test programs in the suite may fail which leads to build fail.  One solution is to disable the specific failed test using the "-x " flag (as shown in the patch)
Steps to apply the patch:
1.  hg clone https://hg.python.org/cpython cpython
2.  cd cpython
3.  hg update 2.7 (needed for 2.7 only)
4.  Copy *.patch to the current directory
5.  patch < python2.7-pgo.patch (or patch < python3.6-pgo.patch)
6.  ./configure
7.  make
To disable PGO
7b. make disable-profile-opt
In the following, please find our sample performance results from latest XEON machine, XEON Broadwell EP.
Hardware (HW):      Intel XEON (Broadwell) 8 Cores
BIOS settings:      Intel Turbo Boost Technology: false
                    Hyper-Threading: false
Operating System:   Ubuntu 14.04.3 LTS trusty
OS configuration:   CPU freq set at fixed: 2.6GHz by
                        echo 2600000 > /sys/devices/system/cpu/cpu*/cpufreq/scaling_min_freq
                        echo 2600000 > /sys/devices/system/cpu/cpu*/cpufreq/scaling_max_freq
                    Address Space Layout Randomization (ASLR) disabled (to reduce run to run variation) by
                        echo 0 > /proc/sys/kernel/randomize_va_space
GCC version:        gcc version 4.8.4 (Ubuntu 4.8.4-2ubuntu1~14.04)
Benchmark:          Grand Unified Python Benchmark (GUPB)
                    GUPB Source: https://hg.python.org/benchmarks/
Python2.7 results:
    Python source: hg clone https://hg.python.org/cpython cpython
    Python Source: hg update 2.7
    hg id: 0511b1165bb6 (2.7)
    hg id -r 'ancestors(.) and tag()': 15c95b7d81dc (2.7) v2.7.10
    hg --debug id -i: 0511b1165bb6cf40ada0768a7efc7ba89316f6a5
        Benchmarks          Speedup(%)
        simple_logging      20
        raytrace            20
        silent_logging      19
        richards            19
        chaos               16
        formatted_logging   16
        json_dump           15
        hexiom2             13
        pidigits            12
        slowunpickle        12
        django_v2           12
        unpack_sequence     11
        float               11
        mako                11
        slowpickle          11
        fastpickle          11
        django              11
        go                  10
        json_dump_v2        10
        pathlib             10
        regex_compile       10
        pybench             9.9
        etree_process       9
        regex_v8            8
        bzr_startup         8
        2to3                8
        slowspitfire        8
        telco               8
        pickle_list         8
        fannkuch            8
        etree_iterparse     8
        nqueens             8
        mako_v2             8
        etree_generate      8
        call_method_slots   7
        html5lib_warmup     7
        html5lib            7
        nbody               7
        spectral_norm       7
        spambayes           7
        fastunpickle        6
        meteor_contest      6
        chameleon           6
        rietveld            6
        tornado_http        5
        unpickle_list       5
        pickle_dict         4
        regex_effbot        3
        normal_startup      3
        startup_nosite      3
        etree_parse         2
        call_method_unknown 2
        call_simple         1
        json_load           1
        call_method         1
Python3.6 results
    Python source: hg clone https://hg.python.org/cpython cpython
    hg id: 96d016f78726 tip
    hg id -r 'ancestors(.) and tag()': 1a58b1227501 (3.5) v3.5.0rc1
    hg --debug id -i: 96d016f78726afbf66d396f084b291ea43792af1
        Benchmark           Speedup(%)
        fastunpickle        22.94
        fastpickle          21.67
        json_load           17.64
        simple_logging      17.49
        meteor_contest      16.67
        formatted_logging   15.33
        etree_process       14.61
        raytrace            13.57
        etree_generate      13.56
        chaos               12.09
        hexiom2             12
        nbody               11.88
        json_dump_v2        11.24
        richards            11.02
        nqueens             10.96
        fannkuch            10.79
        go                  10.77
        float               10.26
        regex_compile       9.8
        silent_logging      9.63
        pidigits            9.58
        etree_iterparse     9.48
        2to3                8.44
        regex_v8            8.09
        regex_effbot        7.88
        call_simple         7.63
        tornado_http        7.38
        etree_parse         4.92
        spectral_norm       4.72
        normal_startup      4.39
        telco               3.88
        startup_nosite      3.7
        call_method         3.63
        unpack_sequence     3.6
        call_method_slots   2.91
        call_method_unknown 2.59
        iterative_count     0.45
        threaded_count      -2.79
Thank you,
Python-Dev mailing list
Python-Dev at python.org

Unsubscribe: https://mail.python.org/mailman/options/python-dev/guido%40python.org

--Guido van Rossum (python.org/~guido)
Python-Dev mailing list
Python-Dev at python.org
Unsubscribe: https://mail.python.org/mailman/options/python-dev/brett%40python.org

More information about the Python-Dev mailing list