/pandas

PERF: Removed the GIL from parts of the TextReader class #11272

Merged
merged 1 commit into from Nov 4, 2015

Conversation

Reviewers
No reviews
Assignees
No one assigned
Labels
IO CSV Performance
Projects
None yet
Milestone
  0.17.1
3 participants
@jdeschenes @jreback @mrocklin
@jdeschenes
Contributor

jdeschenes commented Oct 9, 2015

The GIL was removed around the tokenizer functions and the conversion function(_string_convert excluded).

Benchmark:

Data Generation:

import pandas as pd
import numpy as np

df = pd.DataFrame(np.random.randn(1000000,10))
df.to_csv('test.csv')

Benchmark Code:

import pandas as pd
from pandas.util.testing import test_parallel

def f():
    for i in range(4):
        pd.read_csv('test.csv', index_col=0)

@test_parallel(4)
def g():
    pd.read_csv('test.csv', index_col=0)

Before:

In [4]: %timeit pd.read_csv('test.csv', index_col=0)
1 loops, best of 3: 2.3 s per loop

In [7]: %timeit f()
1 loops, best of 3: 9.15 s per loop

In [8]: %timeit g()
1 loops, best of 3: 9.25 s per loop

After:

In [6]: %timeit pd.read_csv('test.csv', index_col=0)
1 loops, best of 3: 2.35 s per loop

In [9]: %timeit f()
1 loops, best of 3: 9.55 s per loop

In [10]: %timeit g()
1 loops, best of 3: 4.38 s per loop

@jreback jreback added Performance CSV labels Oct 9, 2015

jreback added this to the 0.17.1 milestone Oct 9, 2015

@jreback jreback commented on an outdated diff Oct 9, 2015

asv_bench/benchmarks/gil.py
@@ -320,3 +325,21 @@ def time_nogil_kth_smallest(self):
def run(arr):
algos.kth_smallest(arr, self.k)
run()
+
+
+class nogil_read_csv(object):
+ number = 1
+ repeat = 5
+
+ def setup(self):
+ if (not have_real_test_parallel):
+ raise NotImplementedError
+ # Using the values
+ self.data = '0.1213700904466425978256438611,0.0525708283766902484401839501,0.4174092731488769913994474336\n 0.4096341697147408700274695547,0.1587830198973579909349496119,0.1292545832485494372576795285\n 0.8323255650024565799327547210,0.9694902427379478160318626578,0.6295047811546814475747169126\n 0.4679375305798131323697930383,0.2963942381834381301075609371,0.5268936082160610157032465394\n 0.6685382761849776311890991564,0.6721207066140679753374342908,0.6519975277021627935170045020\n '
@jreback

jreback Oct 9, 2015

Contributor

I would just do use random here. e.g. DataFrame(np.random.randn(10000,10)) (or whatever makes this a reasonable time

also can you add a tests that uses object and another that has a datetime parse in the index_col

@jreback jreback commented on an outdated diff Oct 9, 2015

pandas/parser.pyx
cdef int status
status = tokenize_nrows(self.parser, nrows)
if self.parser.warn_msg != NULL:
- print >> sys.stderr, self.parser.warn_msg
+ # Need to rewrite this without the GIL
@jreback

jreback Oct 9, 2015

Contributor

this is ok, as doesn't happen that often
@jreback
Contributor

jreback commented Oct 9, 2015

some windows cythoning errors.

odd they don't show up for you, what platform are you testing on?

[pandas3.5] C:\Users\Jeff Reback\pandas3.5>git checkout -b jdeschenes-nogil_csv master
Switched to a new branch 'jdeschenes-nogil_csv'

[pandas3.5] C:\Users\Jeff Reback\pandas3.5>git pull https://github.com/jdeschenes/pandas.git nogil_csv
remote: Counting objects: 7, done.
remote: Compressing objects: 100% (7/7), done.
remote: Total 7 (delta 0), reused 0 (delta 0), pack-reused 0
Unpacking objects: 100% (7/7), done.
From https://github.com/jdeschenes/pandas
 * branch            nogil_csv  -> FETCH_HEAD
Updating 26db172..2da353e
Fast-forward
 asv_bench/benchmarks/gil.py |  23 ++++
 pandas/parser.pyx           | 269 +++++++++++++++++++++++++++++++-------------
 2 files changed, 213 insertions(+), 79 deletions(-)

[pandas3.5] C:\Users\Jeff Reback\pandas3.5>make
python setup.py build_ext --inplace
make: python: Command not found
Makefile:2: recipe for target `tseries' failed
make: *** [tseries] Error 127

[pandas3.5] C:\Users\Jeff Reback\pandas3.5>python setup.py build_ext --inplace
running build_ext
skipping 'pandas\index.c' Cython extension (up-to-date)
skipping 'pandas\tslib.c' Cython extension (up-to-date)
skipping 'pandas\hashtable.c' Cython extension (up-to-date)
skipping 'pandas\algos.c' Cython extension (up-to-date)
cythoning pandas\parser.pyx to pandas\parser.c

Error compiling Cython file:
------------------------------------------------------------
...
            # in the hash table
            if k != na_hashset.n_buckets:
                na_count[0] += 1
                data[0] = NA
            else:
                data[0] = parser.converter(word, &p_end, parser.decimal, parser.sci,
                                         ^
------------------------------------------------------------

pandas\parser.pyx:1538:42: Calling gil-requiring function not allowed without gil

Error compiling Cython file:
------------------------------------------------------------
...
                        data[0] = NA
            data += 1
    else:
        for i in range(lines):
            COLITER_NEXT(it, word)
            data[0] = parser.converter(word, &p_end, parser.decimal, parser.sci,
                                     ^
------------------------------------------------------------

pandas\parser.pyx:1557:38: Calling gil-requiring function not allowed without gil

Error compiling Cython file:
------------------------------------------------------------
...
                na_count[0] += 1
                data[0] = NA
                data += 1
                continue

            error = to_boolean(word, data)
                             ^
------------------------------------------------------------

pandas\parser.pyx:1680:30: Calling gil-requiring function not allowed without gil

Error compiling Cython file:
------------------------------------------------------------
...
            data += 1
    else:
        for i in range(lines):
            COLITER_NEXT(it, word)

            error = to_boolean(word, data)
                             ^
------------------------------------------------------------

pandas\parser.pyx:1688:30: Calling gil-requiring function not allowed without gil

Error compiling Cython file:
------------------------------------------------------------
...
            if k != false_hashset.n_buckets:
                data[0] = 0
                data += 1
                continue

            error = to_boolean(word, data)
                             ^
------------------------------------------------------------

pandas\parser.pyx:1758:30: Calling gil-requiring function not allowed without gil

Error compiling Cython file:
------------------------------------------------------------
...
            if k != false_hashset.n_buckets:
                data[0] = 0
                data += 1
                continue

            error = to_boolean(word, data)
                             ^
------------------------------------------------------------

pandas\parser.pyx:1778:30: Calling gil-requiring function not allowed without gil
building 'pandas.parser' extension
C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\BIN\amd64\cl.exe /c /nologo /Ox /W3 /GL /DNDEBUG /MD -Ipandas/src/klib -Ipandas/src -IC:\Miniconda\envs\pandas3.5\lib\site-packages\numpy\core\in
clude -IC:\Miniconda\envs\pandas3.5\include -IC:\Miniconda\envs\pandas3.5\include "-IC:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\INCLUDE" "-IC:\Program Files (x86)\Microsoft Visual Studio 1
4.0\VC\ATLMFC\INCLUDE" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.10240.0\ucrt" "-IC:\Program Files (x86)\Windows Kits\NETFXSDK\4.6\include\um" "-IC:\Program Files (x86)\Windows Kits\10\in
clude\10.0.10240.0\shared" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.10240.0\um" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.10240.0\winrt" /Tcpandas\parser.c /Fobuild\temp.win
-amd64-3.5\Release\pandas\parser.obj
parser.c
pandas\parser.c(1): fatal error C1189: #error:  Do not use this file, it is the result of a failed Cython compilation.
error: command 'C:\\Program Files (x86)\\Microsoft Visual Studio 14.0\\VC\\BIN\\amd64\\cl.exe' failed with exit status 2

[pandas3.5] C:\Users\Jeff Reback\pandas3.5>
@jdeschenes
Contributor

jdeschenes commented Oct 9, 2015

I fixed a few issues with the build
@jreback
Contributor

jreback commented Oct 9, 2015

builds ,couple of errors on windows

[pandas3.5] C:\Users\Jeff Reback\pandas3.5>nosetests -A "not network" pandas\io\tests\test_parsers.py
...........S.........................................................................F......................................S....................S..........................S...........................
................................F........S..................................................................................................................................S...........................
..............................................
======================================================================
FAIL: test_parse_bools (pandas.io.tests.test_parsers.TestCParserHighMemory)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "C:\Users\Jeff Reback\pandas3.5\pandas\io\tests\test_parsers.py", line 1177, in test_parse_bools
    self.assertEqual(data['A'].dtype, np.bool_)
AssertionError: dtype('O') != <class 'numpy.bool_'>

======================================================================
FAIL: test_parse_bools (pandas.io.tests.test_parsers.TestCParserLowMemory)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "C:\Users\Jeff Reback\pandas3.5\pandas\io\tests\test_parsers.py", line 1177, in test_parse_bools
    self.assertEqual(data['A'].dtype, np.bool_)
AssertionError: dtype('O') != <class 'numpy.bool_'>

----------------------------------------------------------------------
Ran 446 tests in 12.620s

FAILED (SKIP=6, failures=2)
@jdeschenes
Contributor

jdeschenes commented Oct 9, 2015

This was an issue with python3 and not limited to windows.

@jreback jreback and 1 other commented on an outdated diff Oct 9, 2015

asv_bench/benchmarks/gil.py
+ self.df.to_csv('__test__.csv')
+
+ self.rng = date_range('1/1/2000', periods=10000)
+ self.df_date_time = DataFrame(np.random.randn(10000, 10), index=self.rng)
+ self.df_date_time.to_csv('__test_datetime__.csv')
+
+ self.df_object = DataFrame('foo', index=self.df.index, columns=self.create_cols('object'))
+ self.df_object.to_csv('__test_object__.csv')
+
+ def create_cols(self, name):
+ return [('%s%03d' % (name, i)) for i in range(5)]
+
+ def time_nogil_read_csv(self):
+ @test_parallel(num_threads=2)
+ def run(arr):
+ read_csv('__test__.csv', sep=',', header=None, float_precision=None)
@jreback

jreback Oct 9, 2015

Contributor

is their a cleanup function on these? (e.g. to remove the created csvs)?

@qwhelan ?

@jreback

jreback Oct 25, 2015

Contributor

@jdeschenes @qwhelan ?

@jdeschenes

jdeschenes Oct 27, 2015

Contributor

There isn't. I had a look at a few tests and they don't seem to cleanup either... see test.h5

@jreback jreback and 1 other commented on an outdated diff Oct 9, 2015

pandas/parser.pyx
@@ -1452,6 +1450,20 @@ cdef _to_fw_string(parser_t *parser, int col, int line_start,
result = np.empty(line_end - line_start, dtype='|S%d' % width)
data = <char*> result.data
+ with nogil:
+ error = _to_fw_string_internal(parser, col, line_start, line_end, width, data)
+ if error != 0:
+ raise CParserError("Unknown error raised")
@jreback

jreback Oct 9, 2015

Contributor

are these ever hit?

@jdeschenes

jdeschenes Oct 13, 2015

Contributor

No.

_to_fw_string_internal function is essentially a wrapper around strncpy. It can never fail.

I wanted to cover the bases if the internal function implementation is changed. I can remove it.

@jreback

jreback Oct 25, 2015

Contributor

ok
@jreback
Contributor

jreback commented Oct 13, 2015

@jdeschenes

  • couple of questions for you above.
  • can you post the asv results here.
  • pls add a note in the performance section (use this PR number)
  • squash
@jdeschenes
Contributor

jdeschenes commented Oct 13, 2015

@jreback

Where do I need to add information in the performance section? Is it in the what's new file in the documentation?
@jreback
Contributor

jreback commented Oct 13, 2015

whatsnew/v0.17.1 (Performance section)
@jreback
Contributor

jreback commented Oct 25, 2015

@jdeschenes looks good. can you

  • add a note in whatsnew (performance section)
  • rebase
  • squash
@jreback
Contributor

jreback commented Nov 2, 2015

can you update according to comments
@mrocklin

mrocklin commented Nov 3, 2015

What's the status on this @jdeschenes? I'd like to include this work in a talk happening tomorrow. It'd be awesome to be able to say that this was in master rather than in a branch.
@jreback
Contributor

jreback commented Nov 3, 2015

you can say slated for 0.17.1 :)
@jdeschenes
Contributor

jdeschenes commented Nov 3, 2015

I will get the final changes tomorrow.
@jdeschenes
Contributor

jdeschenes commented Nov 4, 2015

@jreback, The changes have been implemented. Let me know if there is anything else that needs to be done.

@jreback jreback commented on an outdated diff Nov 4, 2015

pandas/parser.pyx
@@ -1452,6 +1449,18 @@ cdef _to_fw_string(parser_t *parser, int col, int line_start,
result = np.empty(line_end - line_start, dtype='|S%d' % width)
data = <char*> result.data
+ with nogil:
+ _to_fw_string_internal(parser, col, line_start, line_end, width, data)
@jreback

jreback Nov 4, 2015

Contributor

call this: _to_fw_string_nogil instead

@jreback jreback commented on an outdated diff Nov 4, 2015

doc/source/whatsnew/v0.17.1.txt
@@ -60,6 +60,8 @@ Performance Improvements
- Release the GIL on most datetime field operations (e.g. ``DatetimeIndex.year``, ``Series.dt.year``), normalization, and conversion to and from ``Period``, ``DatetimeIndex.to_period`` and ``PeriodIndex.to_timestamp`` (:issue:`11263`)
- Release the GIL on some srolling algos (``rolling_median``, ``rolling_mean``, ``rolling_max``, ``rolling_min``, ``rolling_var``, ``rolling_kurt``, `rolling_skew`` (:issue:`11450`)
+- Release the GIL when reading a file using the TextReader class(``read_csv``, ``read_table``).
+ The GIL is now released around the tokenizer functions and the conversion functions(_string_convert excluded).
@jreback

jreback Nov 4, 2015

Contributor

just say: Release the GIL when reading and parsing text files inpd.read_csv, (:issue:12272)

@jreback jreback commented on an outdated diff Nov 4, 2015

pandas/parser.pyx
lines = line_end - line_start
result = np.empty(lines, dtype=np.float64)
data = <double *> result.data
+ na_fset = kset_float64_from_list(na_flist)
+ with nogil:
+ error = _try_double_internal(parser, col, line_start, line_end,
+ na_filter, na_hashset, use_na_flist, na_fset, NA, data, &na_count)
@jreback

jreback Nov 4, 2015

Contributor

similarly name this
@jreback
Contributor

jreback commented Nov 4, 2015

@jdeschenes thanks, just some small comments. ping when pushed. pls also post a short benchmark in the top of the PR (you can just run before/after in ipython via timeit if you want), mainly for posterity.
@jdeschenes jdeschenes PERF: Released the GIL from parts of the TextReader class 
The GIL was released around the tokenizer functions and the conversion function(_string_convert excluded).
091df3e
@jdeschenes
Contributor

jdeschenes commented Nov 4, 2015

@jreback: Added the benchmarks.
@jreback
Contributor

jreback commented Nov 4, 2015

looks good

ping when green
@mrocklin

mrocklin commented Nov 4, 2015

Ping

@jreback jreback added a commit that referenced this pull request Nov 4, 2015

@jreback jreback Merge pull request #11272 from jdeschenes/nogil_csv 
PERF: Removed the GIL from parts of the TextReader class
 774411c

@jreback jreback merged commit 774411c into pandas-dev:master Nov 4, 2015
1 check passed

1 check passed

continuous-integration/travis-ci/pr The Travis CI build passed
Details
@jreback
Contributor

jreback commented Nov 4, 2015

thanks @jdeschenes

and @mrocklin for the pings!

jdeschenes deleted the jdeschenes:nogil_csv branch Nov 4, 2015

@khs26 khs26 added a commit to khs26/pandas that referenced this pull request Nov 6, 2015

@khs26 khs26 Merge pull request #1 from pydata/master 
Merge pull request #11272 from jdeschenes/nogil_csv
 fd4fc5f
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment