File: details.md

package info (click to toggle)
ipyparallel 8.8.0-6
  • links: PTS, VCS
  • area: main
  • in suites: sid
  • size: 12,412 kB
  • sloc: python: 21,991; javascript: 267; makefile: 29; sh: 28
file content (695 lines) | stat: -rw-r--r-- 21,343 bytes parent folder | download | duplicates (2)
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
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
(parallel-details)=

# Details of Parallel Computing with IPython

```{note}
There are still many sections to fill out in this doc
```

## Caveats

First, some caveats about the detailed workings of parallel computing with 0MQ and IPython.

### Non-copying sends and numpy arrays

When numpy arrays are passed as arguments to apply or via data-movement methods, they are not
copied. This means that you must be careful if you are sending an array that you intend to work
on. PyZMQ does allow you to track when a message has been sent so you can know when it is safe
to edit the buffer, but IPython only allows for this.

It is also important to note that the non-copying receive of a message is _read-only_. That
means that if you intend to work in-place on an array that you have sent or received, you must
copy it. This is true for both numpy arrays sent to engines and numpy arrays retrieved as
results.

The following will fail:

```ipython
In [3]: A = numpy.zeros(2)

In [4]: def setter(a):
   ...:   a[0]=1
   ...:   return a

In [5]: rc[0].apply_sync(setter, A)
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)<string> in <module>()
<ipython-input-12-c3e7afeb3075> in setter(a)
RuntimeError: array is not writeable
```

If you do need to edit the array in-place, remember to copy the array if it's read-only.
The {attr}`ndarray.flags.writeable` flag will tell you if you can write to an array.

```ipython
In [3]: A = numpy.zeros(2)

In [4]: def setter(a):
   ...:     """only copy read-only arrays"""
   ...:     if not a.flags.writeable:
   ...:         a=a.copy()
   ...:     a[0]=1
   ...:     return a

In [5]: rc[0].apply_sync(setter, A)
Out[5]: array([ 1.,  0.])

# note that results will also be read-only:
In [6]: _.flags.writeable
Out[6]: False
```

If you want to safely edit an array in-place after _sending_ it, you must use the `track=True`
flag. IPython always performs non-copying sends of arrays, which return immediately. You must
instruct IPython track those messages _at send time_ in order to know for sure that the send has
completed. AsyncResults have a {attr}`sent` property, and {meth}`wait_on_send` method for
checking and waiting for 0MQ to finish with a buffer.

```ipython
In [5]: A = numpy.random.random((1024,1024))

In [6]: view.track=True

In [7]: ar = view.apply_async(lambda x: 2*x, A)

In [8]: ar.sent
Out[8]: False

In [9]: ar.wait_on_send() # blocks until sent is True
```

### What is sendable?

If IPython doesn't know what to do with an object, it will pickle it. There is a short list of
objects that are not pickled: `buffers/memoryviews`, `bytes` objects, and `numpy`
arrays. These are handled specially by IPython in order to prevent extra in-memory copies of data. Sending
bytes or numpy arrays will result in exactly zero in-memory copies of your data (unless the data
is very small).

If you have an object that provides a Python buffer interface, then you can always send that
buffer without copying - and reconstruct the object on the other side in your own code. It is
possible that the object reconstruction will become extensible, so you can add your own
non-copying types, but this does not yet exist.

#### Closures

Just about anything in Python is pickleable. The one notable exception is objects (generally
functions) with _closures_. Closures can be a complicated topic, but the basic principle is that
functions that refer to variables in their parent scope have closures.

An example of a function that uses a closure:

```python
def f(a):
    def inner():
        # inner will have a closure
        return a
    return inner

f1 = f(1)
f2 = f(2)
f1() # returns 1
f2() # returns 2
```

`f1` and `f2` will have closures referring to the scope in which `inner` was defined,
because they use the variable 'a'. As a result, you would not be able to send `f1` or `f2`
with IPython. Note that you _would_ be able to send `f`. This is only true for interactively
defined functions (as are often used in decorators), and only when there are variables used
inside the inner function, that are defined in the outer function. If the names are _not_ in the
outer function, then there will not be a closure, and the generated function will look in
`globals()` for the name:

```python
def g(b):
    # note that `b` is not referenced in inner's scope
    def inner():
        # this inner will *not* have a closure
        return a
    return inner
g1 = g(1)
g2 = g(2)
g1() # raises NameError on 'a'
a=5
g2() # returns 5
```

`g1` and `g2` _will_ be sendable with IPython, and will treat the engine's namespace as
globals(). The {meth}`pull` method is implemented based on this principle. If we did not
provide pull, you could implement it yourself with `apply`, by returning objects out
of the global namespace:

```ipython
In [10]: view.apply(lambda : a)

# is equivalent to
In [11]: view.pull('a')
```

You can send functions with closures if you enable using dill or cloudpickle:

```ipython
In [10]: rc[:].use_cloudpickle()
```

which will use a more advanced pickling library, which covers things like closures.

## Running Code

There are two principal units of execution in Python: strings of Python code (e.g. 'a=5'),
and Python functions. IPython is designed around the use of functions via the core
Client method, called `apply`.

### Apply

The principal method of remote execution is {meth}`apply`, of
{class}`~ipyparallel.client.view.View` objects. The Client provides the full execution and
communication API for engines via its low-level {meth}`send_apply_message` method, which is used
by all higher level methods of its Views.

f

: The function to be called remotely

args

: The positional arguments passed to `f`

kwargs

: The keyword arguments passed to `f`

flags for all views:

block

: Whether to wait for the result, or return immediately.

False:

: returns AsyncResult

True:

: returns actual result(s) of `f(*args, **kwargs)`

    if multiple targets:

    : list of results, matching `targets`

track

: whether to track non-copying sends.

targets

: Specify the destination of the job.

if 'all' or None:

: Run on all active engines

if list:

: Run on each specified engine

if int:

: Run on single engine

```{note}
{class}`LoadBalancedView` uses targets to restrict possible destinations.
LoadBalanced calls will always execute on exactly one engine.
```

flags only in LoadBalancedViews:

after

: Only for load-balanced execution (targets=None)
Specify a list of msg ids as a time-based dependency.
This job will only be run _after_ the dependencies
have been met.

follow

: Only for load-balanced execution (targets=None)
Specify a list of msg_ids as a location-based dependency.
This job will only be run on an engine where this dependency
is met.

timeout

: Only for load-balanced execution (targets=None)
Specify an amount of time (in seconds) for the scheduler to
wait for dependencies to be met before failing with a
DependencyTimeout.

### execute and run

For executing strings of Python code, {class}`~.DirectView` s also provide an {meth}`~.DirectView.execute` and
a {meth}`~.DirectView.run` method, which rather than take functions and arguments, take Python strings.
`execute` takes a string of Python code to execute, and sends it to the Engine(s). `run`
is the same as `execute`, but for a _filename_ rather than a string. It is a wrapper that
does something very similar to `execute(open(f).read())`.

```{note}
TODO: Examples for execute and run
```

## Views

The principal extension of the {class}`~parallel.Client` is the {class}`~parallel.View`
class. The client is typically a singleton for connecting to a cluster, and presents a
low-level interface to the Hub and Engines. Most real usage will involve creating one or more
{class}`~parallel.View` objects for working with engines in various ways.

### DirectView

The {class}`.DirectView` is the class for the IPython {ref}`Multiplexing Interface <parallel-direct>`.

#### Creating a DirectView

DirectViews can be created in two ways, by index access to a client, or by a client's
{meth}`view` method. Index access to a Client works in a few ways. First, you can create
DirectViews to single engines by accessing the client by engine id:

```ipython
In [2]: rc[0]
Out[2]: <DirectView 0>
```

You can also create a DirectView with a list of engines:

```ipython
In [2]: rc[0,1,2]
Out[2]: <DirectView [0,1,2]>
```

Other methods for accessing elements, such as slicing and negative indexing, work by passing
the index directly to the client's {attr}`ids` list, so:

```ipython
# negative index
In [2]: rc[-1]
Out[2]: <DirectView 3>

# or slicing:
In [3]: rc[::2]
Out[3]: <DirectView [0,2]>
```

are always the same as:

```ipython
In [2]: rc[rc.ids[-1]]
Out[2]: <DirectView 3>

In [3]: rc[rc.ids[::2]]
Out[3]: <DirectView [0,2]>
```

Also note that the slice is evaluated at the time of construction of the DirectView, so the
targets will not change over time if engines are added/removed from the cluster.

#### Execution via DirectView

The DirectView is the simplest way to work with one or more engines directly (hence the name).

For instance, to get the process ID of all your engines:

```ipython
In [5]: import os

In [6]: dview.apply_sync(os.getpid)
Out[6]: [1354, 1356, 1358, 1360]
```

Or to see the hostname of the machine they are on:

```ipython
In [5]: import socket

In [6]: dview.apply_sync(socket.gethostname)
Out[6]: ['tesla', 'tesla', 'edison', 'edison', 'edison']
```

```{note}
TODO: expand on direct execution
```

#### Data movement via DirectView

Since a Python namespace is a {class}`dict`, {class}`DirectView` objects provide
dictionary-style access by key and methods such as {meth}`get` and
{meth}`update` for convenience. This make the remote namespaces of the engines
appear as a local dictionary. Underneath, these methods call {meth}`apply`:

```ipython
In [51]: dview['a']=['foo','bar']

In [52]: dview['a']
Out[52]: [ ['foo', 'bar'], ['foo', 'bar'], ['foo', 'bar'], ['foo', 'bar'] ]
```

### Scatter and gather

Sometimes it is useful to partition a sequence and push the partitions to
different engines. In MPI language, this is know as scatter/gather and we
follow that terminology. However, it is important to remember that in
IPython's {class}`Client` class, {meth}`scatter` is from the
interactive IPython session to the engines and {meth}`gather` is from the
engines back to the interactive IPython session. For scatter/gather operations
between engines, MPI should be used:

```ipython
In [58]: dview.scatter('a',range(16))
Out[58]: [None,None,None,None]

In [59]: dview['a']
Out[59]: [ [0, 1, 2, 3], [4, 5, 6, 7], [8, 9, 10, 11], [12, 13, 14, 15] ]

In [60]: dview.gather('a')
Out[60]: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]
```

### Push and pull

{meth}`~ipyparallel.client.view.DirectView.push`

{meth}`~ipyparallel.client.view.DirectView.pull`

```{note}
TODO: write this section
```

### LoadBalancedView

The {class}`~.LoadBalancedView` is the class for load-balanced execution via the task scheduler.
These views always run tasks on exactly one engine, but let the scheduler determine where that
should be, allowing load-balancing of tasks. The LoadBalancedView does allow you to specify
restrictions on where and when tasks can execute, for more complicated load-balanced workflows.

## Data Movement

Since the {class}`~.LoadBalancedView` does not know where execution will take place, explicit
data movement methods like push/pull and scatter/gather do not make sense, and are not provided.

## Results

### AsyncResults

Our primary representation of the results of remote execution is the {class}`~.AsyncResult`
object, based on the object of the same name in the built-in {py:mod}`multiprocessing.pool`
module.
Our version provides a superset of that interface,
and starting in 6.0 is a subclass of {class}`concurrent.futures.Future`.

The basic principle of the AsyncResult is the encapsulation of one or more results not yet completed.
Execution methods (including data movement, such as push/pull) will all return AsyncResults when `block=False`.

### The mp.pool.AsyncResult interface

The basic interface of the AsyncResult is exactly that of the AsyncResult in {py:mod}`multiprocessing.pool`, and consists of four methods:

% AsyncResult spec directly from docs.python.org

```{eval-rst}
.. class:: AsyncResult

   The stdlib AsyncResult spec

   .. method:: wait([timeout])

      Wait until the result is available or until *timeout* seconds pass. This
      method always returns ``None``.

   .. method:: ready()

      Return whether the call has completed.

   .. method:: successful()

      Return whether the call completed without raising an exception.  Will
      raise :exc:`AssertionError` if the result is not ready.

   .. method:: get([timeout])

      Return the result when it arrives.  If *timeout* is not ``None`` and the
      result does not arrive within *timeout* seconds then
      :exc:`TimeoutError` is raised.  If the remote call raised
      an exception then that exception will be reraised as a :exc:`RemoteError`
      by :meth:`get`.

```

While an AsyncResult is not done, you can check on it with its {meth}`ready` method, which will
return whether the AR is done. You can also wait on an AsyncResult with its {meth}`wait` method.
This method blocks until the result arrives. If you don't want to wait forever, you can pass a
timeout (in seconds) as an argument to {meth}`wait`. {meth}`wait` will _always return None_, and
should never raise an error.

{meth}`ready` and {meth}`wait` are insensitive to the success or failure of the call. After a
result is done, {meth}`successful` will tell you whether the call completed without raising an
exception.

If you want the result of the call, you can use {meth}`get`. Initially, {meth}`get`
behaves just like {meth}`wait`, in that it will block until the result is ready, or until a
timeout is met. However, unlike {meth}`wait`, {meth}`get` will raise a {exc}`TimeoutError` if
the timeout is reached and the result is still not ready. If the result arrives before the
timeout is reached, then {meth}`get` will return the result itself if no exception was raised,
and will raise an exception if there was.

Here is where we start to expand on the multiprocessing interface. Rather than raising the
original exception, a RemoteError will be raised, encapsulating the remote exception with some
metadata. If the AsyncResult represents multiple calls (e.g. any time `targets` is plural), then
a CompositeError, a subclass of RemoteError, will be raised.

```{seealso}
For more information on remote exceptions, see {ref}`the section in the Direct Interface <parallel-exceptions>`.
```

#### Extended interface

Other extensions of the AsyncResult interface include convenience wrappers for {meth}`get`.
AsyncResults have a property, {attr}`result`, with the short alias {attr}`r`, which call
{meth}`get`. Since our object is designed for representing _parallel_ results, it is expected
that many calls (any of those submitted via DirectView) will map results to engine IDs. We
provide a {meth}`get_dict`, which is also a wrapper on {meth}`get`, which returns a dictionary
of the individual results, keyed by engine ID.

You can also prevent a submitted job from executing, via the AsyncResult's
{meth}`abort` method. This will instruct engines to not execute the job when it arrives.

The larger extension of the AsyncResult API is the {attr}`metadata` attribute. The metadata
is a dictionary (with attribute access) that contains, logically enough, metadata about the
execution.

Metadata keys:

timestamps

submitted

: When the task left the Client

started

: When the task started execution on the engine

completed

: When execution finished on the engine

received

: When the result arrived on the Client

note that it is not known when the result arrived in 0MQ on the client, only when it
arrived in Python via {meth}`Client.spin`, so in interactive use, this may not be
strictly informative.

Information about the engine

engine_id

: The integer id

engine_uuid

: The UUID of the engine

output of the call

error

: Python exception, if there was one

execute_input

: The code (str) that was executed

execute_result

: Python output of an execute request (not apply),
as a Jupyter message dictionary.

stderr

: stderr stream

stdout

: stdout (e.g. print) stream

And some extended information

status

: either 'ok' or 'error'

msg_id

: The UUID of the message

after

: For tasks: the time-based msg_id dependencies

follow

: For tasks: the location-based msg_id dependencies

While in most cases, the Clients that submitted a request will be the ones using the results,
other Clients can also request results directly from the Hub. This is done via the Client's
{meth}`get_result` method. This method will _always_ return an AsyncResult object. If the call
was not submitted by the client, then it will be a subclass, called {class}`AsyncHubResult`.
These behave in the same way as an AsyncResult, but if the result is not ready, waiting on an
AsyncHubResult polls the Hub, which is much more expensive than the passive polling used
in regular AsyncResults.

The Client keeps track of all results
history, results, metadata

## Querying the Hub

The Hub sees all traffic that may pass through the schedulers between engines and clients.
It does this so that it can track state, allowing multiple clients to retrieve results of
computations submitted by their peers, as well as persisting the state to a database.

queue_status

> You can check the status of the queues of the engines with this command.

result_status

> check on results

purge_results

> forget results (conserve resources)

## Controlling the Engines

There are a few actions you can do with Engines that do not involve execution. These
messages are sent via the Control socket, and bypass any long queues of waiting execution
jobs

abort

> Sometimes you may want to prevent a job you have submitted from running. The method
> for this is {meth}`abort`. It takes a container of msg_ids, and instructs the Engines to not
> run the jobs if they arrive. The jobs will then fail with an AbortedTask error.

clear

> You may want to purge the Engine(s) namespace of any data you have left in it. After
> running `clear`, there will be no names in the Engine's namespace

shutdown

> You can also instruct engines (and the Controller) to terminate from a Client. This
> can be useful when a job is finished, since you can shutdown all the processes with a
> single command.

## Synchronization

Since the Client is a synchronous object, events do not automatically trigger in your
interactive session - you must poll the 0MQ sockets for incoming messages. Note that
this polling _does not_ make any network requests. It performs a `select`
operation, to check if messages are already in local memory, waiting to be handled.

The method that handles incoming messages is {meth}`spin`. This method flushes any waiting
messages on the various incoming sockets, and updates the state of the Client.

If you need to wait for particular results to finish, you can use the {meth}`wait` method,
which will call {meth}`spin` until the messages are no longer outstanding. Anything that
represents a collection of messages, such as a list of msg_ids or one or more AsyncResult
objects, can be passed as argument to wait. A timeout can be specified, which will prevent
the call from blocking for more than a specified time, but the default behavior is to wait
forever.

The client also has an `outstanding` attribute - a `set` of msg ids that are awaiting
replies. This is the default if wait is called with no arguments - i.e. wait on _all_
outstanding messages.

```{note}
TODO wait example
```

## Map

Many parallel computing problems can be expressed as a `map`, or running a single program with
a variety of different inputs. Python has a built-in {py:func}`map`, which does exactly this,
and many parallel execution tools in Python, such as the built-in
{py:class}`multiprocessing.Pool` object provide implementations of `map`. All View objects
provide a {meth}`map` method as well, but the load-balanced and direct implementations differ.

Views' map methods can be called on any number of sequences,
but they can also take keyword arguments to influence how the work is distributed.
What keyword arguments are available depends on the view being used.

```{eval-rst}
.. class:: ipyparallel.DirectView
   :noindex:

   .. automethod:: map
      :noindex:
```

```{eval-rst}
.. class:: ipyparallel.LoadBalancedView
   :noindex:

   .. automethod:: map
      :noindex:

   .. automethod:: imap
      :noindex:
```

## Decorators and RemoteFunctions

```{note}
TODO: write this section
```

{func}`~ipyparallel.client.remotefunction.parallel`

{func}`~ipyparallel.client.remotefunction.remote`

{class}`~ipyparallel.client.remotefunction.RemoteFunction`

{class}`~ipyparallel.client.remotefunction.ParallelFunction`

## Dependencies

```{note}
TODO: write this section
```

{func}`~ipyparallel.controller.dependency.depend`

{func}`~ipyparallel.controller.dependency.require`

{class}`~ipyparallel.controller.dependency.Dependency`