Note
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.
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, just 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
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, str/bytes objects, and numpy arrays. These are handled specially by IPython in order to prevent the copying 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 principal 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 echo
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 echo
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 pull() method is implemented based on this principal. If we did not provide pull, you could implement it yourself with apply, by simply returning objects out of the global namespace:
In [10]: view.apply(lambda : a)
# is equivalent to
In [11]: view.pull('a')
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.
flags for all views:
Whether to wait for the result, or return immediately. False:
returns AsyncResult
Specify the destination of the job. if ‘all’ or None:
Run on all active engines
Note that LoadBalancedView uses targets to restrict possible destinations. LoadBalanced calls will always execute in just one location.
flags only in LoadBalancedViews:
For executing strings of Python code, DirectView ‘s also provide an execute() and a 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 execute(open(f).read()).
Note
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.
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 [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.
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']
Note
TODO: expand on direct execution
Since a Python namespace is just 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'] ]
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]
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.
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.
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 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 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.
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 actually 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.
Other extensions of the AsyncResult interface include convenience wrappers for get(). 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 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 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 execution.
Metadata keys:
timestamps
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
output of the call
And some extended information
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
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)
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 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.
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.
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 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.
Note
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 the block and bound keyword arguments, just like apply(), but only as keywords.
dview.map(*sequences, block=None)