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`
|