Details of Parallel Computing with IPython#


There are still many sections to fill out in this doc


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:

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 ndarray.flags.writeable flag will tell you if you can write to an array.

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 sent property, and wait_on_send() method for checking and waiting for 0MQ to finish with a buffer.

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.


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:

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:

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'
g2() # returns 5

g1 and g2 will be sendable with IPython, and will treat the engine’s namespace as globals(). The 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:

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:

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.


The principal method of remote execution is apply(), of View objects. The Client provides the full execution and communication API for engines via its low-level send_apply_message() method, which is used by all higher level methods of its Views.


The function to be called remotely


The positional arguments passed to f


The keyword arguments passed to f

flags for all views:


Whether to wait for the result, or return immediately.


returns AsyncResult


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

if multiple targets:

list of results, matching targets


whether to track non-copying sends.


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


LoadBalancedView uses targets to restrict possible destinations. LoadBalanced calls will always execute on exactly one engine.

flags only in LoadBalancedViews:


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.


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.


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, DirectView s also provide an execute() and a 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()).


TODO: Examples for execute and run


The principal extension of the Client is the 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 View objects for working with engines in various ways.


The DirectView is the class for the IPython Multiplexing Interface.

Creating a DirectView#

DirectViews can be created in two ways, by index access to a client, or by a client’s 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:

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

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

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 ids list, so:

# 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:

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:

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:

In [5]: import socket

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


TODO: expand on direct execution

Data movement via DirectView#

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

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 Client class, scatter() is from the interactive IPython session to the engines and gather() is from the engines back to the interactive IPython session. For scatter/gather operations between engines, MPI should be used:

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#




TODO: write this section


The 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 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.



Our primary representation of the results of remote execution is the AsyncResult object, based on the object of the same name in the built-in multiprocessing.pool module. Our version provides a superset of that interface, and starting in 6.0 is a subclass of 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 multiprocessing.pool, and consists of four methods:

class AsyncResult#

The stdlib AsyncResult spec


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


Return whether the call has completed.


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


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

While an AsyncResult is not done, you can check on it with its ready() method, which will return whether the AR is done. You can also wait on an AsyncResult with its 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 wait(). wait() will always return None, and should never raise an error.

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

If you want the result of the call, you can use get(). Initially, get() behaves just like wait(), in that it will block until the result is ready, or until a timeout is met. However, unlike wait(), get() will raise a TimeoutError if the timeout is reached and the result is still not ready. If the result arrives before the timeout is reached, then 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.

See also

For more information on remote exceptions, see the section in the Direct Interface.

Extended interface#

Other extensions of the AsyncResult interface include convenience wrappers for get(). AsyncResults have a property, result, with the short alias r, which call 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 get_dict(), which is also a wrapper on 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 abort() method. This will instruct engines to not execute the job when it arrives.

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

Metadata keys:



When the task left the Client


When the task started execution on the engine


When execution finished on the engine


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 Client.spin(), so in interactive use, this may not be strictly informative.

Information about the engine


The integer id


The UUID of the engine

output of the call


Python exception, if there was one


The code (str) that was executed


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


stderr stream


stdout (e.g. print) stream

And some extended information


either ‘ok’ or ‘error’


The UUID of the message


For tasks: the time-based msg_id dependencies


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 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 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.


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


check on 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


Sometimes you may want to prevent a job you have submitted from running. The method for this is 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.


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


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.


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 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 wait() method, which will call 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.


TODO wait example


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 map(), which does exactly this, and many parallel execution tools in Python, such as the built-in multiprocessing.Pool object provide implementations of map. All View objects provide a 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.

class ipyparallel.DirectView, *sequences, block=None, track=False, return_exceptions=False)

Parallel version of builtin map, using this View’s targets.

There will be one task per target, so work will be chunked if the sequences are longer than targets.

Results can be iterated as they are ready, but will become available in chunks.

New in version 7.0: return_exceptions

  • f (callable) – function to be mapped

  • *sequences (one or more sequences of matching length) – the sequences to be distributed and passed to f

  • block (bool [default self.block]) – whether to wait for the result or not

  • track (bool [default False]) – Track underlying zmq send to indicate when it is safe to modify memory. Only for zero-copy sends such as numpy arrays that are going to be modified in-place.

  • return_exceptions (bool [default False]) – Return remote Exceptions in the result sequence instead of raising them.


  • If block=False – An AsyncMapResult instance. An object like AsyncResult, but which reassembles the sequence of results into a single list. AsyncMapResults can be iterated through before all results are complete.

  • else – A list, the result of map(f,*sequences)

class ipyparallel.LoadBalancedView, *sequences, block=None, chunksize=1, ordered=True, return_exceptions=False)

Parallel version of builtin map, load-balanced by this View.

Each chunksize elements will be a separate task, and will be load-balanced. This lets individual elements be available for iteration as soon as they arrive.

New in version 7.0: return_exceptions

  • f (callable) – function to be mapped

  • *sequences (one or more sequences of matching length) – the sequences to be distributed and passed to f

  • block (bool [default self.block]) – whether to wait for the result or not

  • chunksize (int [default 1]) – how many elements should be in each task.

  • ordered (bool [default True]) –

    Whether the results should be gathered as they arrive, or enforce the order of submission.

    Only applies when iterating through AsyncMapResult as results arrive. Has no effect when block=True.

  • return_exceptions (bool [default False]) – Return Exceptions instead of raising on the first exception.


  • if block=False – An AsyncMapResult instance. An object like AsyncResult, but which reassembles the sequence of results into a single list. AsyncMapResults can be iterated through before all results are complete.

  • else – A list, the result of map(f,*sequences)

LoadBalancedView.imap(f, *sequences, ordered=True, max_outstanding='auto', return_exceptions=False)

Parallel version of lazily-evaluated imap, load-balanced by this View.

ordered, and max_outstanding can be specified by keyword only.

Unlike other map functions in IPython Parallel, this one does not consume the full iterable before submitting work, returning a single ‘AsyncMapResult’ representing the full computation.

Instead, it consumes iterables as they come, submitting up to max_outstanding tasks to the cluster before waiting on results (default: one task per engine). This allows it to work with infinite generators, and avoid potentially expensive read-ahead for large streams of inputs that may not fit in memory all at once.

New in version 7.0.

  • f (callable) – function to be mapped

  • *sequences (one or more sequences of matching length) – the sequences to be distributed and passed to f

  • ordered (bool [default True]) – Whether the results should be yielded on a first-come-first-yield basis, or preserve the order of submission.

  • max_outstanding (int [default len(engines)]) –

    The maximum number of tasks to be outstanding.

    max_outstanding=0 will greedily consume the whole generator (map_async may be more efficient).

    A limit of 1 should be strictly worse than running a local map, as there will be no parallelism.

    Use this to tune how greedily input generator should be consumed.

  • return_exceptions (bool [default False]) – Return Exceptions instead of raising them.


  • lazily-evaluated generator, yielding results of f on each item of sequences.

  • Yield-order depends on ordered argument.

Decorators and RemoteFunctions#


TODO: write this section







TODO: write this section