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 : A = numpy.zeros(2) In : def setter(a): ...: a=1 ...: return a In : rc.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, just remember to copy the array if it’s read-only.
ndarray.flags.writeable flag will tell you if you can write to an array.
In : A = numpy.zeros(2) In : def setter(a): ...: """only copy read-only arrays""" ...: if not a.flags.writeable: ...: a=a.copy() ...: a=1 ...: return a In : rc.apply_sync(setter, A) Out: array([ 1., 0.]) # note that results will also be read-only: In : _.flags.writeable Out: 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 : A = numpy.random.random((1024,1024)) In : view.track=True In : ar = view.apply_async(lambda x: 2*x, A) In : ar.sent Out: False In : 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:
bytes objects, and
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
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
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' a=5 g2() # returns 5
g1 and g2 will be sendable with IPython, and will treat the engine’s namespace as
pull() method is implemented based on this principle. If we did not
provide pull, you could implement it yourself with apply, by simply returning objects out
of the global namespace:
In : view.apply(lambda : a) # is equivalent to In : view.pull('a')
You can send functions with closures if you enable using dill or cloudpickle:
In : rc[:].use_cloudpickle()
which will use a more advanced pickling library, which covers things like closures.
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
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.
- f : function
- The function to be called remotely
- args : tuple/list
- The positional arguments passed to f
- kwargs : dict
- The keyword arguments passed to f
flags for all views:
- block : bool (default: view.block)
Whether to wait for the result, or return immediately.
- returns AsyncResult
returns actual result(s) of
- if multiple targets:
- list of results, matching targets
- track : bool [default view.track]
- whether to track non-copying sends.
- targets : int,list of ints, ‘all’, None [default view.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
LoadBalancedView uses targets to restrict possible destinations.
LoadBalanced calls will always execute in just one location.
flags only in LoadBalancedViews:
- after : Dependency or collection of msg_ids
- 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 : Dependency or collection of msg_ids
- 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 : float/int or None
- 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
run() method, which rather than take functions and arguments, take simple strings.
execute simply takes a string of Python code to execute, and sends it to the Engine(s). run
is the same as execute, but for a file, rather than a string. It is simply a wrapper that
does something very similar to
TODO: Examples for execute and run
The principal extension of the
Client is the
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.
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 simply by accessing the client by engine id:
In : rc Out: <DirectView 0>
You can also create a DirectView with a list of engines:
In : rc[0,1,2] Out: <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 : rc[-1] Out: <DirectView 3> # or slicing: In : rc[::2] Out: <DirectView [0,2]>
are always the same as:
In : rc[rc.ids[-1]] Out: <DirectView 3> In : rc[rc.ids[::2]] Out: <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 : import os In : dview.apply_sync(os.getpid) Out: [1354, 1356, 1358, 1360]
Or to see the hostname of the machine they are on:
In : import socket In : dview.apply_sync(socket.gethostname) Out: ['tesla', 'tesla', 'edison', 'edison', 'edison']
TODO: expand on direct execution
Data movement via DirectView¶
Since a Python namespace is just a
DirectView objects provide
dictionary-style access by key and methods such as
update() for convenience. This make the remote namespaces of the engines
appear as a local dictionary. Underneath, these methods call
In : dview['a']=['foo','bar'] In : dview['a'] Out: [ ['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
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 : dview.scatter('a',range(16)) Out: [None,None,None,None] In : dview['a'] Out: [ [0, 1, 2, 3], [4, 5, 6, 7], [8, 9, 10, 11], [12, 13, 14, 15] ] In : dview.gather('a') Out: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]
Push and pull¶
TODO: write this section
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.
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
object, based on the object of the same name in the built-in
module. Our version provides a superset of that interface.
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:
The stdlib AsyncResult spec
Wait until the result is available or until timeout seconds pass. This method always returns
Return whether the call has completed.
Return whether the call completed without raising an exception. Will raise
AssertionErrorif the result is not ready.
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
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() will always return None, and
should never raise an error.
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
If you actually want the result of the call, you can use
behaves just like
wait(), in that it will block until the result is ready, or until a
timeout is met. However, unlike
get() will raise a
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.
For more information on remote exceptions, see the section in the Direct Interface.
Other extensions of the AsyncResult interface include convenience wrappers for
AsyncResults have a property,
result, with the short alias
r, which simply 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
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 actually 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
- 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
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 actually 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 actually make any network requests. It simply 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
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
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
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
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 the block
and bound keyword arguments, just like
apply(), but only as keywords.
- iter, map_async, reduce
Decorators and RemoteFunctions¶
TODO: write this section
TODO: write this section