-
Notifications
You must be signed in to change notification settings - Fork 81
/
Copy pathparallelism.html
638 lines (557 loc) · 42.4 KB
/
parallelism.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
<!DOCTYPE html>
<!-- data-theme below is forced to be "light" but should be changed if we use pydata-theme-sphinx in the future -->
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" data-content_root="../" data-theme="light"> <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" data-content_root="../" data-theme="light"> <!--<![endif]-->
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1" />
<meta property="og:title" content="8.3. Parallelism, resource management, and configuration" />
<meta property="og:type" content="website" />
<meta property="og:url" content="https://scikit-learn/stable/computing/parallelism.html" />
<meta property="og:site_name" content="scikit-learn" />
<meta property="og:description" content="Parallelism: Some scikit-learn estimators and utilities parallelize costly operations using multiple CPU cores. Depending on the type of estimator and sometimes the values of the constructor parame..." />
<meta property="og:image" content="https://scikit-learn.org/stable/_static/scikit-learn-logo-small.png" />
<meta property="og:image:alt" content="scikit-learn" />
<meta name="description" content="Parallelism: Some scikit-learn estimators and utilities parallelize costly operations using multiple CPU cores. Depending on the type of estimator and sometimes the values of the constructor parame..." />
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>8.3. Parallelism, resource management, and configuration — scikit-learn 1.4.2 documentation</title>
<link rel="canonical" href="https://scikit-learn.org/stable/computing/parallelism.html" />
<link rel="shortcut icon" href="../_static/favicon.ico"/>
<link rel="stylesheet" href="../_static/css/vendor/bootstrap.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/pygments.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<link rel="stylesheet" href="../_static/copybutton.css" type="text/css" />
<link rel="stylesheet" href="../_static/plot_directive.css" type="text/css" />
<link rel="stylesheet" href="../https://fonts.googleapis.com/css?family=Vibur" type="text/css" />
<link rel="stylesheet" href="../_static/jupyterlite_sphinx.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-binder.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-dataframe.css" type="text/css" />
<link rel="stylesheet" href="../_static/sg_gallery-rendered-html.css" type="text/css" />
<link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
<script id="documentation_options" data-url_root="../" src="../_static/documentation_options.js"></script>
<script src="../_static/js/vendor/jquery-3.6.3.slim.min.js"></script>
<script src="../_static/js/details-permalink.js"></script>
</head>
<body>
<nav id="navbar" class="sk-docs-navbar navbar navbar-expand-md navbar-light bg-light py-0">
<div class="container-fluid sk-docs-container px-0">
<a class="navbar-brand py-0" href="../index.html">
<img
class="sk-brand-img"
src="../_static/scikit-learn-logo-small.png"
alt="logo"/>
</a>
<button
id="sk-navbar-toggler"
class="navbar-toggler"
type="button"
data-toggle="collapse"
data-target="#navbarSupportedContent"
aria-controls="navbarSupportedContent"
aria-expanded="false"
aria-label="Toggle navigation"
>
<span class="navbar-toggler-icon"></span>
</button>
<div class="sk-navbar-collapse collapse navbar-collapse" id="navbarSupportedContent">
<ul class="navbar-nav mr-auto">
<li class="nav-item">
<a class="sk-nav-link nav-link" href="../install.html">Install</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link" href="../user_guide.html">User Guide</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link" href="../modules/classes.html">API</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link" href="../auto_examples/index.html">Examples</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link" target="_blank" rel="noopener noreferrer" href="https://blog.scikit-learn.org/">Community</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../getting_started.html" >Getting Started</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../tutorial/index.html" >Tutorial</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../whats_new/v1.4.html" >What's new</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../glossary.html" >Glossary</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="https://scikit-learn.org/dev/developers/index.html" target="_blank" rel="noopener noreferrer">Development</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../faq.html" >FAQ</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../support.html" >Support</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../related_projects.html" >Related packages</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../roadmap.html" >Roadmap</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../governance.html" >Governance</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="../about.html" >About us</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="https://github.com/scikit-learn/scikit-learn" >GitHub</a>
</li>
<li class="nav-item">
<a class="sk-nav-link nav-link nav-more-item-mobile-items" href="https://scikit-learn.org/dev/versions.html" >Other Versions and Download</a>
</li>
<li class="nav-item dropdown nav-more-item-dropdown">
<a class="sk-nav-link nav-link dropdown-toggle" href="#" id="navbarDropdown" role="button" data-toggle="dropdown" aria-haspopup="true" aria-expanded="false">More</a>
<div class="dropdown-menu" aria-labelledby="navbarDropdown">
<a class="sk-nav-dropdown-item dropdown-item" href="../getting_started.html" >Getting Started</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../tutorial/index.html" >Tutorial</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../whats_new/v1.4.html" >What's new</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../glossary.html" >Glossary</a>
<a class="sk-nav-dropdown-item dropdown-item" href="https://scikit-learn.org/dev/developers/index.html" target="_blank" rel="noopener noreferrer">Development</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../faq.html" >FAQ</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../support.html" >Support</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../related_projects.html" >Related packages</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../roadmap.html" >Roadmap</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../governance.html" >Governance</a>
<a class="sk-nav-dropdown-item dropdown-item" href="../about.html" >About us</a>
<a class="sk-nav-dropdown-item dropdown-item" href="https://github.com/scikit-learn/scikit-learn" >GitHub</a>
<a class="sk-nav-dropdown-item dropdown-item" href="https://scikit-learn.org/dev/versions.html" >Other Versions and Download</a>
</div>
</li>
</ul>
<div id="searchbox" role="search">
<div class="searchformwrapper">
<form class="search" action="../search.html" method="get">
<input class="sk-search-text-input" type="text" name="q" aria-labelledby="searchlabel" />
<input class="sk-search-text-btn" type="submit" value="Go" />
</form>
</div>
</div>
</div>
</div>
</nav>
<div class="d-flex" id="sk-doc-wrapper">
<input type="checkbox" name="sk-toggle-checkbox" id="sk-toggle-checkbox">
<label id="sk-sidemenu-toggle" class="sk-btn-toggle-toc btn sk-btn-primary" for="sk-toggle-checkbox">Toggle Menu</label>
<div id="sk-sidebar-wrapper" class="border-right">
<div class="sk-sidebar-toc-wrapper">
<div class="btn-group w-100 mb-2" role="group" aria-label="rellinks">
<a href="computational_performance.html" role="button" class="btn sk-btn-rellink py-1" sk-rellink-tooltip="8.2. Computational Performance">Prev</a><a href="../computing.html" role="button" class="btn sk-btn-rellink py-1" sk-rellink-tooltip="8. Computing with scikit-learn">Up</a>
<a href="../model_persistence.html" role="button" class="btn sk-btn-rellink py-1" sk-rellink-tooltip="9. Model persistence">Next</a>
</div>
<div class="alert alert-danger p-1 mb-2" role="alert">
<p class="text-center mb-0">
<strong>scikit-learn 1.4.2</strong><br/>
<a href="https://scikit-learn.org/dev/versions.html">Other versions</a>
</p>
</div>
<div class="alert alert-warning p-1 mb-2" role="alert">
<p class="text-center mb-0">
Please <a class="font-weight-bold" href="../about.html#citing-scikit-learn"><string>cite us</string></a> if you use the software.
</p>
</div>
<div class="sk-sidebar-toc">
<ul>
<li>
<a href="../user_guide.html" class="sk-toc-active">User Guide</a>
</li>
<ul>
<li>
<a href="../supervised_learning.html" class="">1. Supervised learning</a>
</li>
<li>
<a href="../unsupervised_learning.html" class="">2. Unsupervised learning</a>
</li>
<li>
<a href="../model_selection.html" class="">3. Model selection and evaluation</a>
</li>
<li>
<a href="../inspection.html" class="">4. Inspection</a>
</li>
<li>
<a href="../visualizations.html" class="">5. Visualizations</a>
</li>
<li>
<a href="../data_transforms.html" class="">6. Dataset transformations</a>
</li>
<li>
<a href="../datasets.html" class="">7. Dataset loading utilities</a>
</li>
<li>
<a href="../computing.html" class="sk-toc-active">8. Computing with scikit-learn</a>
<ul>
<li class="sk-toctree-l3">
<a href="scaling_strategies.html">8.1. Strategies to scale computationally: bigger data</a>
</li>
<li class="sk-toctree-l3">
<a href="computational_performance.html">8.2. Computational Performance</a>
</li>
<li class="sk-toctree-l3">
<a href="">8.3. Parallelism, resource management, and configuration</a>
</li>
</ul>
</li>
<li>
<a href="../model_persistence.html" class="">9. Model persistence</a>
</li>
<li>
<a href="../common_pitfalls.html" class="">10. Common pitfalls and recommended practices</a>
</li>
<li>
<a href="../dispatching.html" class="">11. Dispatching</a>
</li>
</ul>
</ul>
</div>
</div>
</div>
<div id="sk-page-content-wrapper">
<div class="sk-page-content container-fluid body px-md-3" role="main">
<section id="parallelism-resource-management-and-configuration">
<h1><span class="section-number">8.3. </span>Parallelism, resource management, and configuration<a class="headerlink" href="#parallelism-resource-management-and-configuration" title="Link to this heading">¶</a></h1>
<section id="parallelism">
<span id="id1"></span><h2><span class="section-number">8.3.1. </span>Parallelism<a class="headerlink" href="#parallelism" title="Link to this heading">¶</a></h2>
<p>Some scikit-learn estimators and utilities parallelize costly operations
using multiple CPU cores.</p>
<p>Depending on the type of estimator and sometimes the values of the
constructor parameters, this is either done:</p>
<ul class="simple">
<li><p>with higher-level parallelism via <a class="reference external" href="https://joblib.readthedocs.io/en/latest/">joblib</a>.</p></li>
<li><p>with lower-level parallelism via OpenMP, used in C or Cython code.</p></li>
<li><p>with lower-level parallelism via BLAS, used by NumPy and SciPy for generic operations
on arrays.</p></li>
</ul>
<p>The <code class="docutils literal notranslate"><span class="pre">n_jobs</span></code> parameters of estimators always controls the amount of parallelism
managed by joblib (processes or threads depending on the joblib backend).
The thread-level parallelism managed by OpenMP in scikit-learn’s own Cython code
or by BLAS & LAPACK libraries used by NumPy and SciPy operations used in scikit-learn
is always controlled by environment variables or <code class="docutils literal notranslate"><span class="pre">threadpoolctl</span></code> as explained below.
Note that some estimators can leverage all three kinds of parallelism at different
points of their training and prediction methods.</p>
<p>We describe these 3 types of parallelism in the following subsections in more details.</p>
<section id="higher-level-parallelism-with-joblib">
<h3><span class="section-number">8.3.1.1. </span>Higher-level parallelism with joblib<a class="headerlink" href="#higher-level-parallelism-with-joblib" title="Link to this heading">¶</a></h3>
<p>When the underlying implementation uses joblib, the number of workers
(threads or processes) that are spawned in parallel can be controlled via the
<code class="docutils literal notranslate"><span class="pre">n_jobs</span></code> parameter.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>Where (and how) parallelization happens in the estimators using joblib by
specifying <code class="docutils literal notranslate"><span class="pre">n_jobs</span></code> is currently poorly documented.
Please help us by improving our docs and tackle <a class="reference external" href="https://github.com/scikit-learn/scikit-learn/issues/14228">issue 14228</a>!</p>
</div>
<p>Joblib is able to support both multi-processing and multi-threading. Whether
joblib chooses to spawn a thread or a process depends on the <strong>backend</strong>
that it’s using.</p>
<p>scikit-learn generally relies on the <code class="docutils literal notranslate"><span class="pre">loky</span></code> backend, which is joblib’s
default backend. Loky is a multi-processing backend. When doing
multi-processing, in order to avoid duplicating the memory in each process
(which isn’t reasonable with big datasets), joblib will create a <a class="reference external" href="https://docs.scipy.org/doc/numpy/reference/generated/numpy.memmap.html">memmap</a>
that all processes can share, when the data is bigger than 1MB.</p>
<p>In some specific cases (when the code that is run in parallel releases the
GIL), scikit-learn will indicate to <code class="docutils literal notranslate"><span class="pre">joblib</span></code> that a multi-threading
backend is preferable.</p>
<p>As a user, you may control the backend that joblib will use (regardless of
what scikit-learn recommends) by using a context manager:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kn">from</span> <span class="nn">joblib</span> <span class="kn">import</span> <span class="n">parallel_backend</span>
<span class="k">with</span> <span class="n">parallel_backend</span><span class="p">(</span><span class="s1">'threading'</span><span class="p">,</span> <span class="n">n_jobs</span><span class="o">=</span><span class="mi">2</span><span class="p">):</span>
<span class="c1"># Your scikit-learn code here</span>
</pre></div>
</div>
<p>Please refer to the <a class="reference external" href="https://joblib.readthedocs.io/en/latest/parallel.html#thread-based-parallelism-vs-process-based-parallelism">joblib’s docs</a>
for more details.</p>
<p>In practice, whether parallelism is helpful at improving runtime depends on
many factors. It is usually a good idea to experiment rather than assuming
that increasing the number of workers is always a good thing. In some cases
it can be highly detrimental to performance to run multiple copies of some
estimators or functions in parallel (see oversubscription below).</p>
</section>
<section id="lower-level-parallelism-with-openmp">
<h3><span class="section-number">8.3.1.2. </span>Lower-level parallelism with OpenMP<a class="headerlink" href="#lower-level-parallelism-with-openmp" title="Link to this heading">¶</a></h3>
<p>OpenMP is used to parallelize code written in Cython or C, relying on
multi-threading exclusively. By default, the implementations using OpenMP
will use as many threads as possible, i.e. as many threads as logical cores.</p>
<p>You can control the exact number of threads that are used either:</p>
<ul>
<li><p>via the <code class="docutils literal notranslate"><span class="pre">OMP_NUM_THREADS</span></code> environment variable, for instance when:
running a python script:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><style type="text/css">
span.prompt1:before {
content: "$ ";
}
</style><span class="prompt1"><span class="nv">OMP_NUM_THREADS</span><span class="o">=</span><span class="m">4</span><span class="w"> </span>python<span class="w"> </span>my_script.py</span>
</pre></div></div></li>
<li><p>or via <code class="docutils literal notranslate"><span class="pre">threadpoolctl</span></code> as explained by <a class="reference external" href="https://github.com/joblib/threadpoolctl/#setting-the-maximum-size-of-thread-pools">this piece of documentation</a>.</p></li>
</ul>
</section>
<section id="parallel-numpy-and-scipy-routines-from-numerical-libraries">
<h3><span class="section-number">8.3.1.3. </span>Parallel NumPy and SciPy routines from numerical libraries<a class="headerlink" href="#parallel-numpy-and-scipy-routines-from-numerical-libraries" title="Link to this heading">¶</a></h3>
<p>scikit-learn relies heavily on NumPy and SciPy, which internally call
multi-threaded linear algebra routines (BLAS & LAPACK) implemented in libraries
such as MKL, OpenBLAS or BLIS.</p>
<p>You can control the exact number of threads used by BLAS for each library
using environment variables, namely:</p>
<ul class="simple">
<li><p><code class="docutils literal notranslate"><span class="pre">MKL_NUM_THREADS</span></code> sets the number of thread MKL uses,</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">OPENBLAS_NUM_THREADS</span></code> sets the number of threads OpenBLAS uses</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">BLIS_NUM_THREADS</span></code> sets the number of threads BLIS uses</p></li>
</ul>
<p>Note that BLAS & LAPACK implementations can also be impacted by
<code class="docutils literal notranslate"><span class="pre">OMP_NUM_THREADS</span></code>. To check whether this is the case in your environment,
you can inspect how the number of threads effectively used by those libraries
is affected when running the following command in a bash or zsh terminal
for different values of <code class="docutils literal notranslate"><span class="pre">OMP_NUM_THREADS</span></code>:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span class="prompt1"><span class="nv">OMP_NUM_THREADS</span><span class="o">=</span><span class="m">2</span><span class="w"> </span>python<span class="w"> </span>-m<span class="w"> </span>threadpoolctl<span class="w"> </span>-i<span class="w"> </span>numpy<span class="w"> </span>scipy</span>
</pre></div></div><div class="admonition note">
<p class="admonition-title">Note</p>
<p>At the time of writing (2022), NumPy and SciPy packages which are
distributed on pypi.org (i.e. the ones installed via <code class="docutils literal notranslate"><span class="pre">pip</span> <span class="pre">install</span></code>)
and on the conda-forge channel (i.e. the ones installed via
<code class="docutils literal notranslate"><span class="pre">conda</span> <span class="pre">install</span> <span class="pre">--channel</span> <span class="pre">conda-forge</span></code>) are linked with OpenBLAS, while
NumPy and SciPy packages packages shipped on the <code class="docutils literal notranslate"><span class="pre">defaults</span></code> conda
channel from Anaconda.org (i.e. the ones installed via <code class="docutils literal notranslate"><span class="pre">conda</span> <span class="pre">install</span></code>)
are linked by default with MKL.</p>
</div>
</section>
<section id="oversubscription-spawning-too-many-threads">
<h3><span class="section-number">8.3.1.4. </span>Oversubscription: spawning too many threads<a class="headerlink" href="#oversubscription-spawning-too-many-threads" title="Link to this heading">¶</a></h3>
<p>It is generally recommended to avoid using significantly more processes or
threads than the number of CPUs on a machine. Over-subscription happens when
a program is running too many threads at the same time.</p>
<p>Suppose you have a machine with 8 CPUs. Consider a case where you’re running
a <a class="reference internal" href="../modules/generated/sklearn.model_selection.GridSearchCV.html#sklearn.model_selection.GridSearchCV" title="sklearn.model_selection.GridSearchCV"><code class="xref py py-class docutils literal notranslate"><span class="pre">GridSearchCV</span></code></a> (parallelized with joblib)
with <code class="docutils literal notranslate"><span class="pre">n_jobs=8</span></code> over a
<a class="reference internal" href="../modules/generated/sklearn.ensemble.HistGradientBoostingClassifier.html#sklearn.ensemble.HistGradientBoostingClassifier" title="sklearn.ensemble.HistGradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingClassifier</span></code></a> (parallelized with
OpenMP). Each instance of
<a class="reference internal" href="../modules/generated/sklearn.ensemble.HistGradientBoostingClassifier.html#sklearn.ensemble.HistGradientBoostingClassifier" title="sklearn.ensemble.HistGradientBoostingClassifier"><code class="xref py py-class docutils literal notranslate"><span class="pre">HistGradientBoostingClassifier</span></code></a> will spawn 8 threads
(since you have 8 CPUs). That’s a total of <code class="docutils literal notranslate"><span class="pre">8</span> <span class="pre">*</span> <span class="pre">8</span> <span class="pre">=</span> <span class="pre">64</span></code> threads, which
leads to oversubscription of threads for physical CPU resources and thus
to scheduling overhead.</p>
<p>Oversubscription can arise in the exact same fashion with parallelized
routines from MKL, OpenBLAS or BLIS that are nested in joblib calls.</p>
<p>Starting from <code class="docutils literal notranslate"><span class="pre">joblib</span> <span class="pre">>=</span> <span class="pre">0.14</span></code>, when the <code class="docutils literal notranslate"><span class="pre">loky</span></code> backend is used (which
is the default), joblib will tell its child <strong>processes</strong> to limit the
number of threads they can use, so as to avoid oversubscription. In practice
the heuristic that joblib uses is to tell the processes to use <code class="docutils literal notranslate"><span class="pre">max_threads</span>
<span class="pre">=</span> <span class="pre">n_cpus</span> <span class="pre">//</span> <span class="pre">n_jobs</span></code>, via their corresponding environment variable. Back to
our example from above, since the joblib backend of
<a class="reference internal" href="../modules/generated/sklearn.model_selection.GridSearchCV.html#sklearn.model_selection.GridSearchCV" title="sklearn.model_selection.GridSearchCV"><code class="xref py py-class docutils literal notranslate"><span class="pre">GridSearchCV</span></code></a> is <code class="docutils literal notranslate"><span class="pre">loky</span></code>, each process will
only be able to use 1 thread instead of 8, thus mitigating the
oversubscription issue.</p>
<p>Note that:</p>
<ul class="simple">
<li><p>Manually setting one of the environment variables (<code class="docutils literal notranslate"><span class="pre">OMP_NUM_THREADS</span></code>,
<code class="docutils literal notranslate"><span class="pre">MKL_NUM_THREADS</span></code>, <code class="docutils literal notranslate"><span class="pre">OPENBLAS_NUM_THREADS</span></code>, or <code class="docutils literal notranslate"><span class="pre">BLIS_NUM_THREADS</span></code>)
will take precedence over what joblib tries to do. The total number of
threads will be <code class="docutils literal notranslate"><span class="pre">n_jobs</span> <span class="pre">*</span> <span class="pre"><LIB>_NUM_THREADS</span></code>. Note that setting this
limit will also impact your computations in the main process, which will
only use <code class="docutils literal notranslate"><span class="pre"><LIB>_NUM_THREADS</span></code>. Joblib exposes a context manager for
finer control over the number of threads in its workers (see joblib docs
linked below).</p></li>
<li><p>When joblib is configured to use the <code class="docutils literal notranslate"><span class="pre">threading</span></code> backend, there is no
mechanism to avoid oversubscriptions when calling into parallel native
libraries in the joblib-managed threads.</p></li>
<li><p>All scikit-learn estimators that explicitly rely on OpenMP in their Cython code
always use <code class="docutils literal notranslate"><span class="pre">threadpoolctl</span></code> internally to automatically adapt the numbers of
threads used by OpenMP and potentially nested BLAS calls so as to avoid
oversubscription.</p></li>
</ul>
<p>You will find additional details about joblib mitigation of oversubscription
in <a class="reference external" href="https://joblib.readthedocs.io/en/latest/parallel.html#avoiding-over-subscription-of-cpu-resources">joblib documentation</a>.</p>
<p>You will find additional details about parallelism in numerical python libraries
in <a class="reference external" href="https://thomasjpfan.github.io/parallelism-python-libraries-design/">this document from Thomas J. Fan</a>.</p>
</section>
</section>
<section id="configuration-switches">
<h2><span class="section-number">8.3.2. </span>Configuration switches<a class="headerlink" href="#configuration-switches" title="Link to this heading">¶</a></h2>
<section id="python-api">
<h3><span class="section-number">8.3.2.1. </span>Python API<a class="headerlink" href="#python-api" title="Link to this heading">¶</a></h3>
<p><a class="reference internal" href="../modules/generated/sklearn.set_config.html#sklearn.set_config" title="sklearn.set_config"><code class="xref py py-func docutils literal notranslate"><span class="pre">sklearn.set_config</span></code></a> and <a class="reference internal" href="../modules/generated/sklearn.config_context.html#sklearn.config_context" title="sklearn.config_context"><code class="xref py py-func docutils literal notranslate"><span class="pre">sklearn.config_context</span></code></a> can be used to change
parameters of the configuration which control aspect of parallelism.</p>
</section>
<section id="environment-variables">
<span id="environment-variable"></span><h3><span class="section-number">8.3.2.2. </span>Environment variables<a class="headerlink" href="#environment-variables" title="Link to this heading">¶</a></h3>
<p>These environment variables should be set before importing scikit-learn.</p>
<section id="sklearn-assume-finite">
<h4><span class="section-number">8.3.2.2.1. </span><code class="docutils literal notranslate"><span class="pre">SKLEARN_ASSUME_FINITE</span></code><a class="headerlink" href="#sklearn-assume-finite" title="Link to this heading">¶</a></h4>
<p>Sets the default value for the <code class="docutils literal notranslate"><span class="pre">assume_finite</span></code> argument of
<a class="reference internal" href="../modules/generated/sklearn.set_config.html#sklearn.set_config" title="sklearn.set_config"><code class="xref py py-func docutils literal notranslate"><span class="pre">sklearn.set_config</span></code></a>.</p>
</section>
<section id="sklearn-working-memory">
<h4><span class="section-number">8.3.2.2.2. </span><code class="docutils literal notranslate"><span class="pre">SKLEARN_WORKING_MEMORY</span></code><a class="headerlink" href="#sklearn-working-memory" title="Link to this heading">¶</a></h4>
<p>Sets the default value for the <code class="docutils literal notranslate"><span class="pre">working_memory</span></code> argument of
<a class="reference internal" href="../modules/generated/sklearn.set_config.html#sklearn.set_config" title="sklearn.set_config"><code class="xref py py-func docutils literal notranslate"><span class="pre">sklearn.set_config</span></code></a>.</p>
</section>
<section id="sklearn-seed">
<h4><span class="section-number">8.3.2.2.3. </span><code class="docutils literal notranslate"><span class="pre">SKLEARN_SEED</span></code><a class="headerlink" href="#sklearn-seed" title="Link to this heading">¶</a></h4>
<p>Sets the seed of the global random generator when running the tests, for
reproducibility.</p>
<p>Note that scikit-learn tests are expected to run deterministically with
explicit seeding of their own independent RNG instances instead of relying on
the numpy or Python standard library RNG singletons to make sure that test
results are independent of the test execution order. However some tests might
forget to use explicit seeding and this variable is a way to control the initial
state of the aforementioned singletons.</p>
</section>
<section id="sklearn-tests-global-random-seed">
<h4><span class="section-number">8.3.2.2.4. </span><code class="docutils literal notranslate"><span class="pre">SKLEARN_TESTS_GLOBAL_RANDOM_SEED</span></code><a class="headerlink" href="#sklearn-tests-global-random-seed" title="Link to this heading">¶</a></h4>
<p>Controls the seeding of the random number generator used in tests that rely on
the <code class="docutils literal notranslate"><span class="pre">global_random_seed`</span></code> fixture.</p>
<p>All tests that use this fixture accept the contract that they should
deterministically pass for any seed value from 0 to 99 included.</p>
<p>If the <code class="docutils literal notranslate"><span class="pre">SKLEARN_TESTS_GLOBAL_RANDOM_SEED</span></code> environment variable is set to
<code class="docutils literal notranslate"><span class="pre">"any"</span></code> (which should be the case on nightly builds on the CI), the fixture
will choose an arbitrary seed in the above range (based on the BUILD_NUMBER or
the current day) and all fixtured tests will run for that specific seed. The
goal is to ensure that, over time, our CI will run all tests with different
seeds while keeping the test duration of a single run of the full test suite
limited. This will check that the assertions of tests written to use this
fixture are not dependent on a specific seed value.</p>
<p>The range of admissible seed values is limited to [0, 99] because it is often
not possible to write a test that can work for any possible seed and we want to
avoid having tests that randomly fail on the CI.</p>
<p>Valid values for <code class="docutils literal notranslate"><span class="pre">SKLEARN_TESTS_GLOBAL_RANDOM_SEED</span></code>:</p>
<ul class="simple">
<li><p><code class="docutils literal notranslate"><span class="pre">SKLEARN_TESTS_GLOBAL_RANDOM_SEED="42"</span></code>: run tests with a fixed seed of 42</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">SKLEARN_TESTS_GLOBAL_RANDOM_SEED="40-42"</span></code>: run the tests with all seeds
between 40 and 42 included</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">SKLEARN_TESTS_GLOBAL_RANDOM_SEED="any"</span></code>: run the tests with an arbitrary
seed selected between 0 and 99 included</p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">SKLEARN_TESTS_GLOBAL_RANDOM_SEED="all"</span></code>: run the tests with all seeds
between 0 and 99 included. This can take a long time: only use for individual
tests, not the full test suite!</p></li>
</ul>
<p>If the variable is not set, then 42 is used as the global seed in a
deterministic manner. This ensures that, by default, the scikit-learn test
suite is as deterministic as possible to avoid disrupting our friendly
third-party package maintainers. Similarly, this variable should not be set in
the CI config of pull-requests to make sure that our friendly contributors are
not the first people to encounter a seed-sensitivity regression in a test
unrelated to the changes of their own PR. Only the scikit-learn maintainers who
watch the results of the nightly builds are expected to be annoyed by this.</p>
<p>When writing a new test function that uses this fixture, please use the
following command to make sure that it passes deterministically for all
admissible seeds on your local machine:</p>
<div class="highlight-default notranslate"><div class="highlight"><pre><span class="prompt1"><span class="nv">SKLEARN_TESTS_GLOBAL_RANDOM_SEED</span><span class="o">=</span><span class="s2">"all"</span><span class="w"> </span>pytest<span class="w"> </span>-v<span class="w"> </span>-k<span class="w"> </span>test_your_test_name</span>
</pre></div></div></section>
<section id="sklearn-skip-network-tests">
<h4><span class="section-number">8.3.2.2.5. </span><code class="docutils literal notranslate"><span class="pre">SKLEARN_SKIP_NETWORK_TESTS</span></code><a class="headerlink" href="#sklearn-skip-network-tests" title="Link to this heading">¶</a></h4>
<p>When this environment variable is set to a non zero value, the tests that need
network access are skipped. When this environment variable is not set then
network tests are skipped.</p>
</section>
<section id="sklearn-run-float32-tests">
<h4><span class="section-number">8.3.2.2.6. </span><code class="docutils literal notranslate"><span class="pre">SKLEARN_RUN_FLOAT32_TESTS</span></code><a class="headerlink" href="#sklearn-run-float32-tests" title="Link to this heading">¶</a></h4>
<p>When this environment variable is set to ‘1’, the tests using the
<code class="docutils literal notranslate"><span class="pre">global_dtype</span></code> fixture are also run on float32 data.
When this environment variable is not set, the tests are only run on
float64 data.</p>
</section>
<section id="sklearn-enable-debug-cython-directives">
<h4><span class="section-number">8.3.2.2.7. </span><code class="docutils literal notranslate"><span class="pre">SKLEARN_ENABLE_DEBUG_CYTHON_DIRECTIVES</span></code><a class="headerlink" href="#sklearn-enable-debug-cython-directives" title="Link to this heading">¶</a></h4>
<p>When this environment variable is set to a non zero value, the <code class="docutils literal notranslate"><span class="pre">Cython</span></code>
derivative, <code class="docutils literal notranslate"><span class="pre">boundscheck</span></code> is set to <code class="docutils literal notranslate"><span class="pre">True</span></code>. This is useful for finding
segfaults.</p>
</section>
<section id="sklearn-build-enable-debug-symbols">
<h4><span class="section-number">8.3.2.2.8. </span><code class="docutils literal notranslate"><span class="pre">SKLEARN_BUILD_ENABLE_DEBUG_SYMBOLS</span></code><a class="headerlink" href="#sklearn-build-enable-debug-symbols" title="Link to this heading">¶</a></h4>
<p>When this environment variable is set to a non zero value, the debug symbols
will be included in the compiled C extensions. Only debug symbols for POSIX
systems is configured.</p>
</section>
<section id="sklearn-pairwise-dist-chunk-size">
<h4><span class="section-number">8.3.2.2.9. </span><code class="docutils literal notranslate"><span class="pre">SKLEARN_PAIRWISE_DIST_CHUNK_SIZE</span></code><a class="headerlink" href="#sklearn-pairwise-dist-chunk-size" title="Link to this heading">¶</a></h4>
<p>This sets the size of chunk to be used by the underlying <code class="docutils literal notranslate"><span class="pre">PairwiseDistancesReductions</span></code>
implementations. The default value is <code class="docutils literal notranslate"><span class="pre">256</span></code> which has been showed to be adequate on
most machines.</p>
<p>Users looking for the best performance might want to tune this variable using
powers of 2 so as to get the best parallelism behavior for their hardware,
especially with respect to their caches’ sizes.</p>
</section>
<section id="sklearn-warnings-as-errors">
<h4><span class="section-number">8.3.2.2.10. </span><code class="docutils literal notranslate"><span class="pre">SKLEARN_WARNINGS_AS_ERRORS</span></code><a class="headerlink" href="#sklearn-warnings-as-errors" title="Link to this heading">¶</a></h4>
<p>This environment variable is used to turn warnings into errors in tests and
documentation build.</p>
<p>Some CI (Continuous Integration) builds set <code class="docutils literal notranslate"><span class="pre">SKLEARN_WARNINGS_AS_ERRORS=1</span></code>, for
example to make sure that we catch deprecation warnings from our dependencies
and that we adapt our code.</p>
<p>To locally run with the same “warnings as errors” setting as in these CI builds
you can set <code class="docutils literal notranslate"><span class="pre">SKLEARN_WARNINGS_AS_ERRORS=1</span></code>.</p>
<p>By default, warnings are not turned into errors. This is the case if
<code class="docutils literal notranslate"><span class="pre">SKLEARN_WARNINGS_AS_ERRORS</span></code> is unset, or <code class="docutils literal notranslate"><span class="pre">SKLEARN_WARNINGS_AS_ERRORS=0</span></code>.</p>
<p>This environment variable use specific warning filters to ignore some warnings,
since sometimes warnings originate from third-party libraries and there is not
much we can do about it. You can see the warning filters in the
<code class="docutils literal notranslate"><span class="pre">_get_warnings_filters_info_list</span></code> function in <code class="docutils literal notranslate"><span class="pre">sklearn/utils/_testing.py</span></code>.</p>
<p>Note that for documentation build, <code class="docutils literal notranslate"><span class="pre">SKLEARN_WARNING_AS_ERRORS=1</span></code> is checking
that the documentation build, in particular running examples, does not produce
any warnings. This is different from the <code class="docutils literal notranslate"><span class="pre">-W</span></code> <code class="docutils literal notranslate"><span class="pre">sphinx-build</span></code> argument that
catches syntax warnings in the rst files.</p>
</section>
</section>
</section>
</section>
</div>
<div class="container">
<footer class="sk-content-footer">
© 2007 - 2024, scikit-learn developers (BSD License).
<a href="../_sources/computing/parallelism.rst.txt" rel="nofollow">Show this page source</a>
</footer>
</div>
</div>
</div>
<script src="../_static/js/vendor/bootstrap.min.js"></script>
<script>
window.ga=window.ga||function(){(ga.q=ga.q||[]).push(arguments)};ga.l=+new Date;
ga('create', 'UA-22606712-2', 'auto');
ga('set', 'anonymizeIp', true);
ga('send', 'pageview');
</script>
<script async src='https://www.google-analytics.com/analytics.js'></script>
<script defer data-domain="scikit-learn.org" src="https://views.scientific-python.org/js/script.js">
</script>
<script src="../_static/clipboard.min.js"></script>
<script src="../_static/copybutton.js"></script>
<script>
$(document).ready(function() {
/* Add a [>>>] button on the top-right corner of code samples to hide
* the >>> and ... prompts and the output and thus make the code
* copyable. */
var div = $('.highlight-python .highlight,' +
'.highlight-python3 .highlight,' +
'.highlight-pycon .highlight,' +
'.highlight-default .highlight')
var pre = div.find('pre');
// get the styles from the current theme
pre.parent().parent().css('position', 'relative');
// create and add the button to all the code blocks that contain >>>
div.each(function(index) {
var jthis = $(this);
// tracebacks (.gt) contain bare text elements that need to be
// wrapped in a span to work with .nextUntil() (see later)
jthis.find('pre:has(.gt)').contents().filter(function() {
return ((this.nodeType == 3) && (this.data.trim().length > 0));
}).wrap('<span>');
});
/*** Add permalink buttons next to glossary terms ***/
$('dl.glossary > dt[id]').append(function() {
return ('<a class="headerlink" href="#' +
this.getAttribute('id') +
'" title="Permalink to this term">¶</a>');
});
});
</script>
<script id="MathJax-script" async src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-chtml.js"></script>
<script src="https://scikit-learn.org/versionwarning.js"></script>
</body>
</html>