document new AsyncResult properties

docs/source/parallel/asyncresult.txt

@@ -0,0 +1,116 b''
+.. _parallel_asyncresult:
+
+======================
+The AsyncResult object
+======================
+
+In non-blocking mode, :meth:`apply` submits the command to be executed and
+then returns a :class:`~.AsyncResult` object immediately. The
+AsyncResult object gives you a way of getting a result at a later
+time through its :meth:`get` method, but it also collects metadata
+on execution.
+
+
+Beyond multiprocessing's AsyncResult
+====================================
+
+.. Note::
+
+    The :class:`~.AsyncResult` object provides a superset of the interface in 
+    :py:class:`multiprocessing.pool.AsyncResult`.  See the 
+    `official Python documentation <http://docs.python.org/library/multiprocessing#multiprocessing.pool.AsyncResult>`_
+    for more on the basics of this interface.
+
+Our AsyncResult objects add a number of convenient features for working with
+parallel results, beyond what is provided by the original AsyncResult.
+
+
+get_dict
+--------
+
+First, is :meth:`.AsyncResult.get_dict`, which pulls results as a dictionary
+keyed by engine_id, rather than a flat list.  This is useful for quickly
+coordinating or distributing information about all of the engines.
+
+As an example, here is a quick call that gives every engine a dict showing
+the PID of every other engine:
+
+.. sourcecode:: ipython
+
+    In [10]: ar = rc[:].apply_async(os.getpid)
+    In [11]: pids = ar.get_dict()
+    In [12]: rc[:]['pid_map'] = pids
+
+This trick is particularly useful when setting up inter-engine communication,
+as in IPython's :file:`examples/parallel/interengine` examples.
+
+
+Metadata
+========
+
+IPython.parallel tracks some metadata about the tasks, which is stored
+in the :attr:`.Client.metadata` dict.  The AsyncResult object gives you an
+interface for this information as well, including timestamps stdout/err,
+and engine IDs.
+
+
+Timing
+------
+
+IPython tracks various timestamps as :py:class:`.datetime` objects,
+and the AsyncResult object has a few properties that turn these into useful
+times (in seconds as floats).
+
+For use while the tasks are still pending:
+
+* :attr:`ar.elapsed` is just the elapsed seconds since submission, for use
+  before the AsyncResult is complete.
+* :attr:`ar.progress` is the number of tasks that have completed.  Fractional progress
+  would be::
+
+    1.0 * ar.progress / len(ar)
+
+* :meth:`AsyncResult.wait_interactive` will wait for the result to finish, but
+  print out status updates on progress and elapsed time while it waits.
+
+For use after the tasks are done:
+
+* :attr:`ar.serial_time` is the sum of the computation time of all of the tasks
+  done in parallel.
+* :attr:`ar.wall_time` is the time between the first task submitted and last result
+  received.  This is the actual cost of computation, including IPython overhead. 
+  
+
+.. note::
+
+  wall_time is only precise if the Client is waiting for results when
+  the task finished, because the `received` timestamp is made when the result is
+  unpacked by the Client, triggered by the :meth:`~Client.spin` call. If you
+  are doing work in the Client, and not waiting/spinning, then `received` might
+  be artificially high.
+
+An often interesting metric is the time it actually cost to do the work in parallel
+relative to the serial computation, and this can be given simply with
+
+.. sourcecode:: python
+
+    speedup = ar.serial_time / ar.wall_time
+
+
+Map results are iterable!
+=========================
+
+When an AsyncResult object has multiple results (e.g. the :class:`~AsyncMapResult`
+object), you can actually iterate through them, and act on the results as they arrive:
+
+.. literalinclude:: ../../examples/parallel/itermapresult.py
+    :language: python
+    :lines: 20-66
+
+.. seealso::
+
+    When AsyncResult or the AsyncMapResult don't provide what you need (for instance,
+    handling individual results as they arrive, but with metadata), you can always
+    just split the original result's ``msg_ids`` attribute, and handle them as you like.
+    
+    For an example of this, see :file:`docs/examples/parallel/customresult.py`

docs/source/parallel/index.txt

@@ -1,23 +1,24 b''
 .. _parallel_index:
 
 ====================================
 Using IPython for parallel computing
 ====================================
 
 .. toctree::
    :maxdepth: 2
 
    parallel_intro.txt
    parallel_process.txt
    parallel_multiengine.txt
    parallel_task.txt
+   asyncresult.txt
    parallel_mpi.txt
    parallel_db.txt
    parallel_security.txt
    parallel_winhpc.txt
    parallel_demos.txt
    dag_dependencies.txt
    parallel_details.txt
    parallel_transition.txt
 
 

docs/source/parallel/parallel_multiengine.txt

@@ -1,869 +1,865 b''
 .. _parallel_multiengine:
 
 ==========================
 IPython's Direct interface
 ==========================
 
 The direct, or multiengine, interface represents one possible way of working with a set of
 IPython engines. The basic idea behind the multiengine interface is that the
 capabilities of each engine are directly and explicitly exposed to the user.
 Thus, in the multiengine interface, each engine is given an id that is used to
 identify the engine and give it work to do. This interface is very intuitive
 and is designed with interactive usage in mind, and is the best place for
 new users of IPython to begin.
 
 Starting the IPython controller and engines
 ===========================================
 
 To follow along with this tutorial, you will need to start the IPython
 controller and four IPython engines. The simplest way of doing this is to use
 the :command:`ipcluster` command::
 
     $ ipcluster start -n 4
     
 For more detailed information about starting the controller and engines, see
 our :ref:`introduction <parallel_overview>` to using IPython for parallel computing.
 
 Creating a ``DirectView`` instance
 ==================================
 
 The first step is to import the IPython :mod:`IPython.parallel`
 module and then create a :class:`.Client` instance:
 
 .. sourcecode:: ipython
 
     In [1]: from IPython.parallel import Client
     
     In [2]: rc = Client()
 
 This form assumes that the default connection information (stored in
 :file:`ipcontroller-client.json` found in :file:`IPYTHON_DIR/profile_default/security`) is
 accurate. If the controller was started on a remote machine, you must copy that connection
 file to the client machine, or enter its contents as arguments to the Client constructor:
 
 .. sourcecode:: ipython
 
     # If you have copied the json connector file from the controller:
     In [2]: rc = Client('/path/to/ipcontroller-client.json')
     # or to connect with a specific profile you have set up:
     In [3]: rc = Client(profile='mpi')
     
 
 To make sure there are engines connected to the controller, users can get a list
 of engine ids:
 
 .. sourcecode:: ipython
 
     In [3]: rc.ids
     Out[3]: [0, 1, 2, 3]
 
 Here we see that there are four engines ready to do work for us.
 
 For direct execution, we will make use of a :class:`DirectView` object, which can be
 constructed via list-access to the client:
 
 .. sourcecode:: ipython
 
     In [4]: dview = rc[:] # use all engines
 
 .. seealso::
 
     For more information, see the in-depth explanation of :ref:`Views <parallel_details>`.
 
 
 Quick and easy parallelism
 ==========================
 
 In many cases, you simply want to apply a Python function to a sequence of
 objects, but *in parallel*. The client interface provides a simple way
 of accomplishing this: using the DirectView's :meth:`~DirectView.map` method.
 
 Parallel map
 ------------
 
 Python's builtin :func:`map` functions allows a function to be applied to a
 sequence element-by-element. This type of code is typically trivial to
 parallelize. In fact, since IPython's interface is all about functions anyway,
 you can just use the builtin :func:`map` with a :class:`RemoteFunction`, or a 
 DirectView's :meth:`map` method:
 
 .. sourcecode:: ipython
 
     In [62]: serial_result = map(lambda x:x**10, range(32))
     
     In [63]: parallel_result = dview.map_sync(lambda x: x**10, range(32))
 
     In [67]: serial_result==parallel_result
     Out[67]: True
 
 
 .. note::
 
     The :class:`DirectView`'s version of :meth:`map` does
     not do dynamic load balancing. For a load balanced version, use a
     :class:`LoadBalancedView`.
 
 .. seealso::
     
     :meth:`map` is implemented via :class:`ParallelFunction`.
 
 Remote function decorators
 --------------------------
 
 Remote functions are just like normal functions, but when they are called,
 they execute on one or more engines, rather than locally. IPython provides
 two decorators:
 
 .. sourcecode:: ipython
 
     In [10]: @dview.remote(block=True)
        ....: def getpid():
        ....:     import os
        ....:     return os.getpid()
        ....: 
 
     In [11]: getpid()
     Out[11]: [12345, 12346, 12347, 12348]
 
 The ``@parallel`` decorator creates parallel functions, that break up an element-wise
 operations and distribute them, reconstructing the result.
 
 .. sourcecode:: ipython
 
     In [12]: import numpy as np
     
     In [13]: A = np.random.random((64,48))
     
     In [14]: @dview.parallel(block=True)
        ....: def pmul(A,B):
        ....:     return A*B
     
     In [15]: C_local = A*A
     
     In [16]: C_remote = pmul(A,A)
     
     In [17]: (C_local == C_remote).all()
     Out[17]: True
 
 Calling a ``@parallel`` function *does not* correspond to map. It is used for splitting
 element-wise operations that operate on a sequence or array.  For ``map`` behavior,
 parallel functions do have a map method.
 
 ====================    ============================    =============================
 call                    pfunc(seq)                      pfunc.map(seq)
 ====================    ============================    =============================
 # of tasks              # of engines (1 per engine)     # of engines (1 per engine)
 # of remote calls       # of engines (1 per engine)     ``len(seq)``
 argument to remote      ``seq[i:j]`` (sub-sequence)     ``seq[i]`` (single element)
 ====================    ============================    =============================
 
 A quick example to illustrate the difference in arguments for the two modes:
 
 .. sourcecode:: ipython
 
     In [16]: @dview.parallel(block=True)
        ....: def echo(x):
        ....:     return str(x)
        ....: 
 
     In [17]: echo(range(5))
     Out[17]: ['[0, 1]', '[2]', '[3]', '[4]']
 
     In [18]: echo.map(range(5))
     Out[18]: ['0', '1', '2', '3', '4']
 
 
 .. seealso::
 
     See the :func:`~.remotefunction.parallel` and :func:`~.remotefunction.remote`
     decorators for options.
 
 Calling Python functions
 ========================
 
 The most basic type of operation that can be performed on the engines is to
 execute Python code or call Python functions. Executing Python code can be
 done in blocking or non-blocking mode (non-blocking is default) using the
 :meth:`.View.execute` method, and calling functions can be done via the
 :meth:`.View.apply` method.
 
 apply
 -----
 
 The main method for doing remote execution (in fact, all methods that
 communicate with the engines are built on top of it), is :meth:`View.apply`.
 
 We strive to provide the cleanest interface we can, so `apply` has the following
 signature:
 
 .. sourcecode:: python
 
     view.apply(f, *args, **kwargs)
 
 There are various ways to call functions with IPython, and these flags are set as
 attributes of the View.  The ``DirectView`` has just two of these flags:
 
 dv.block : bool
     whether to wait for the result, or return an :class:`AsyncResult` object
     immediately
 dv.track : bool
     whether to instruct pyzmq to track when zeromq is done sending the message.
     This is primarily useful for non-copying sends of numpy arrays that you plan to
     edit in-place.  You need to know when it becomes safe to edit the buffer
     without corrupting the message.
 dv.targets : int, list of ints
     which targets this view is associated with.
 
 
 Creating a view is simple: index-access on a client creates a :class:`.DirectView`.
 
 .. sourcecode:: ipython
 
     In [4]: view = rc[1:3]
     Out[4]: <DirectView [1, 2]>
 
     In [5]: view.apply<tab>
     view.apply  view.apply_async  view.apply_sync
 
 For convenience, you can set block temporarily for a single call with the extra sync/async methods.
 
 Blocking execution
 ------------------
 
 In blocking mode, the :class:`.DirectView` object (called ``dview`` in
 these examples) submits the command to the controller, which places the
 command in the engines' queues for execution. The :meth:`apply` call then
 blocks until the engines are done executing the command:
 
 .. sourcecode:: ipython
 
     In [2]: dview = rc[:] # A DirectView of all engines
     In [3]: dview.block=True
     In [4]: dview['a'] = 5
 
     In [5]: dview['b'] = 10
 
     In [6]: dview.apply(lambda x: a+b+x, 27)
     Out[6]: [42, 42, 42, 42]
 
 You can also select blocking execution on a call-by-call basis with the :meth:`apply_sync`
 method:
 
     In [7]: dview.block=False
 
     In [8]: dview.apply_sync(lambda x: a+b+x, 27)
     Out[8]: [42, 42, 42, 42]
 
 Python commands can be executed as strings on specific engines by using a View's ``execute``
 method:
 
 .. sourcecode:: ipython
 
     In [6]: rc[::2].execute('c=a+b')
 
     In [7]: rc[1::2].execute('c=a-b')
 
     In [8]: dview['c'] # shorthand for dview.pull('c', block=True)
     Out[8]: [15, -5, 15, -5]
 
 
 Non-blocking execution
 ----------------------
 
 In non-blocking mode, :meth:`apply` submits the command to be executed and
 then returns a :class:`AsyncResult` object immediately. The
 :class:`AsyncResult` object gives you a way of getting a result at a later
 time through its :meth:`get` method.
 
-.. Note::
-
-    The :class:`AsyncResult` object provides a superset of the interface in 
-    :py:class:`multiprocessing.pool.AsyncResult`.  See the 
-    `official Python documentation <http://docs.python.org/library/multiprocessing#multiprocessing.pool.AsyncResult>`_
-    for more.
+.. seealso::
 
+    Docs on the :ref:`AsyncResult <parallel_asyncresult>` object.
 
 This allows you to quickly submit long running commands without blocking your
 local Python/IPython session:
 
 .. sourcecode:: ipython
     
     # define our function
     In [6]: def wait(t):
       ....:     import time
       ....:     tic = time.time()
       ....:     time.sleep(t)
       ....:     return time.time()-tic
     
     # In non-blocking mode
     In [7]: ar = dview.apply_async(wait, 2)
 
     # Now block for the result
     In [8]: ar.get()
     Out[8]: [2.0006198883056641, 1.9997570514678955, 1.9996809959411621, 2.0003249645233154]
 
     # Again in non-blocking mode
     In [9]: ar = dview.apply_async(wait, 10)
 
     # Poll to see if the result is ready
     In [10]: ar.ready()
     Out[10]: False
     
     # ask for the result, but wait a maximum of 1 second:
     In [45]: ar.get(1)
     ---------------------------------------------------------------------------
     TimeoutError                              Traceback (most recent call last)
     /home/you/<ipython-input-45-7cd858bbb8e0> in <module>()
     ----> 1 ar.get(1)
 
     /path/to/site-packages/IPython/parallel/asyncresult.pyc in get(self, timeout)
          62                 raise self._exception
          63         else:
     ---> 64             raise error.TimeoutError("Result not ready.")
          65 
          66     def ready(self):
 
     TimeoutError: Result not ready.
 
 .. Note::
 
     Note the import inside the function. This is a common model, to ensure
     that the appropriate modules are imported where the task is run. You can
     also manually import modules into the engine(s) namespace(s) via 
     :meth:`view.execute('import numpy')`.
 
 Often, it is desirable to wait until a set of :class:`AsyncResult` objects
 are done. For this, there is a the method :meth:`wait`. This method takes a
 tuple of :class:`AsyncResult` objects (or `msg_ids` or indices to the client's History),
 and blocks until all of the associated results are ready:
 
 .. sourcecode:: ipython
 
     In [72]: dview.block=False
 
     # A trivial list of AsyncResults objects
     In [73]: pr_list = [dview.apply_async(wait, 3) for i in range(10)]
 
     # Wait until all of them are done
     In [74]: dview.wait(pr_list)
 
     # Then, their results are ready using get() or the `.r` attribute
     In [75]: pr_list[0].get()
     Out[75]: [2.9982571601867676, 2.9982588291168213, 2.9987530708312988, 2.9990990161895752]
 
 
 
 The ``block`` and ``targets`` keyword arguments and attributes
 --------------------------------------------------------------
 
 Most DirectView methods (excluding :meth:`apply`) accept ``block`` and
 ``targets`` as keyword arguments. As we have seen above, these keyword arguments control the
 blocking mode and which engines the command is applied to. The :class:`View` class also has
 :attr:`block` and :attr:`targets` attributes that control the default behavior when the keyword
 arguments are not provided. Thus the following logic is used for :attr:`block` and :attr:`targets`:
 
 * If no keyword argument is provided, the instance attributes are used.
 * Keyword argument, if provided override the instance attributes for
   the duration of a single call.
   
 The following examples demonstrate how to use the instance attributes:
 
 .. sourcecode:: ipython
 
     In [16]: dview.targets = [0,2]
     
     In [17]: dview.block = False
 
     In [18]: ar = dview.apply(lambda : 10)
 
     In [19]: ar.get()
     Out[19]: [10, 10]
 
     In [16]: dview.targets = v.client.ids # all engines (4)
     
     In [21]: dview.block = True
 
     In [22]: dview.apply(lambda : 42)
     Out[22]: [42, 42, 42, 42]
 
 The :attr:`block` and :attr:`targets` instance attributes of the
 :class:`.DirectView` also determine the behavior of the parallel magic commands.
 
 Parallel magic commands
 -----------------------
 
 We provide a few IPython magic commands (``%px``, ``%autopx`` and ``%result``)
 that make it more pleasant to execute Python commands on the engines
 interactively. These are simply shortcuts to :meth:`execute` and
 :meth:`get_result` of the :class:`DirectView`. The ``%px`` magic executes a single 
 Python command on the engines specified by the :attr:`targets` attribute of the
 :class:`DirectView` instance:
 
 .. sourcecode:: ipython
 
     # Create a DirectView for all targets
     In [22]: dv = rc[:]
     
     # Make this DirectView active for parallel magic commands
     In [23]: dv.activate()
 
     In [24]: dv.block=True
 
     # import numpy here and everywhere
     In [25]: with dv.sync_imports():
        ....:    import numpy
     importing numpy on engine(s)
 
     In [27]: %px a = numpy.random.rand(2,2)
     Parallel execution on engines: [0, 1, 2, 3]
 
     In [28]: %px ev = numpy.linalg.eigvals(a)
     Parallel execution on engines: [0, 1, 2, 3]
 
     In [28]: dv['ev']
     Out[28]: [ array([ 1.09522024, -0.09645227]), 
        ....:   array([ 1.21435496, -0.35546712]), 
        ....:   array([ 0.72180653,  0.07133042]), 
        ....:   array([  1.46384341,   1.04353244e-04])
        ....: ]
 
 The ``%result`` magic gets the most recent result, or takes an argument
 specifying the index of the result to be requested. It is simply a shortcut to the
 :meth:`get_result` method:
 
 .. sourcecode:: ipython
     
     In [29]: dv.apply_async(lambda : ev)
     
     In [30]: %result
     Out[30]: [ [ 1.28167017  0.14197338],
        ....:    [-0.14093616  1.27877273],
        ....:    [-0.37023573  1.06779409],
        ....:    [ 0.83664764 -0.25602658] ]
 
 The ``%autopx`` magic switches to a mode where everything you type is executed
 on the engines given by the :attr:`targets` attribute:
 
 .. sourcecode:: ipython
 
     In [30]: dv.block=False
 
     In [31]: %autopx
     Auto Parallel Enabled
     Type %autopx to disable
 
     In [32]: max_evals = []
     <IPython.parallel.AsyncResult object at 0x17b8a70>
 
     In [33]: for i in range(100):
        ....:     a = numpy.random.rand(10,10)
        ....:     a = a+a.transpose()
        ....:     evals = numpy.linalg.eigvals(a)
        ....:     max_evals.append(evals[0].real)
        ....:
        ....:
     <IPython.parallel.AsyncResult object at 0x17af8f0>
 
     In [34]: %autopx
     Auto Parallel Disabled
 
     In [35]: dv.block=True
 
     In [36]: px ans= "Average max eigenvalue is: %f"%(sum(max_evals)/len(max_evals))
     Parallel execution on engines: [0, 1, 2, 3]
     
     In [37]: dv['ans']
     Out[37]: [ 'Average max eigenvalue is:  10.1387247332',
        ....:   'Average max eigenvalue is:  10.2076902286',
        ....:   'Average max eigenvalue is:  10.1891484655',
        ....:   'Average max eigenvalue is:  10.1158837784',]
 
 
 Moving Python objects around
 ============================
 
 In addition to calling functions and executing code on engines, you can
 transfer Python objects to and from your IPython session and the engines. In
 IPython, these operations are called :meth:`push` (sending an object to the
 engines) and :meth:`pull` (getting an object from the engines).
 
 Basic push and pull
 -------------------
 
 Here are some examples of how you use :meth:`push` and :meth:`pull`:
 
 .. sourcecode:: ipython
 
     In [38]: dview.push(dict(a=1.03234,b=3453))
     Out[38]: [None,None,None,None]
 
     In [39]: dview.pull('a')
     Out[39]: [ 1.03234, 1.03234, 1.03234, 1.03234]
 
     In [40]: dview.pull('b', targets=0)
     Out[40]: 3453
 
     In [41]: dview.pull(('a','b'))
     Out[41]: [ [1.03234, 3453], [1.03234, 3453], [1.03234, 3453], [1.03234, 3453] ]
     
     In [43]: dview.push(dict(c='speed'))
     Out[43]: [None,None,None,None]
 
 In non-blocking mode :meth:`push` and :meth:`pull` also return
 :class:`AsyncResult` objects:
 
 .. sourcecode:: ipython
 
     In [48]: ar = dview.pull('a', block=False)
 
     In [49]: ar.get()
     Out[49]: [1.03234, 1.03234, 1.03234, 1.03234]
 
 
 Dictionary interface
 --------------------
 
 Since a Python namespace is just 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`:
 
 .. sourcecode:: 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, pyzmq, or some other direct interconnect should be used.
 
 .. sourcecode:: 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]
 
 Other things to look at
 =======================
 
 How to do parallel list comprehensions
 --------------------------------------
 
 In many cases list comprehensions are nicer than using the map function. While
 we don't have fully parallel list comprehensions, it is simple to get the
 basic effect using :meth:`scatter` and :meth:`gather`:
 
 .. sourcecode:: ipython
 
     In [66]: dview.scatter('x',range(64))
 
     In [67]: %px y = [i**10 for i in x]
     Parallel execution on engines: [0, 1, 2, 3]
     Out[67]:
 
     In [68]: y = dview.gather('y')
 
     In [69]: print y
     [0, 1, 1024, 59049, 1048576, 9765625, 60466176, 282475249, 1073741824,...]
 
 Remote imports
 --------------
 
 Sometimes you will want to import packages both in your interactive session
 and on your remote engines.  This can be done with the :class:`ContextManager`
 created by a DirectView's :meth:`sync_imports` method:
 
 .. sourcecode:: ipython
 
     In [69]: with dview.sync_imports():
        ....:     import numpy
     importing numpy on engine(s)
 
 Any imports made inside the block will also be performed on the view's engines.
 sync_imports also takes a `local` boolean flag that defaults to True, which specifies
 whether the local imports should also be performed.  However, support for `local=False`
 has not been implemented, so only packages that can be imported locally will work
 this way.
 
 You can also specify imports via the ``@require`` decorator.  This is a decorator
 designed for use in Dependencies, but can be used to handle remote imports as well.
 Modules or module names passed to ``@require`` will be imported before the decorated
 function is called.  If they cannot be imported, the decorated function will never
 execution, and will fail with an UnmetDependencyError.
 
 .. sourcecode:: ipython
 
     In [69]: from IPython.parallel import require
 
     In [70]: @require('re'):
        ....: def findall(pat, x):
        ....:     # re is guaranteed to be available
        ....:     return re.findall(pat, x)
           
     # you can also pass modules themselves, that you already have locally:
     In [71]: @require(time):
        ....: def wait(t):
        ....:     time.sleep(t)
        ....:     return t
 
 .. _parallel_exceptions:
 
 Parallel exceptions
 -------------------
 
 In the multiengine interface, parallel commands can raise Python exceptions,
 just like serial commands. But, it is a little subtle, because a single
 parallel command can actually raise multiple exceptions (one for each engine
 the command was run on). To express this idea, we have a
 :exc:`CompositeError` exception class that will be raised in most cases. The
 :exc:`CompositeError` class is a special type of exception that wraps one or
 more other types of exceptions. Here is how it works:
 
 .. sourcecode:: ipython
 
     In [76]: dview.block=True
 
     In [77]: dview.execute('1/0')
     ---------------------------------------------------------------------------
     CompositeError                            Traceback (most recent call last)
     /home/user/<ipython-input-10-5d56b303a66c> in <module>()
     ----> 1 dview.execute('1/0')
 
     /path/to/site-packages/IPython/parallel/client/view.pyc in execute(self, code, targets, block)
         591                 default: self.block
         592         """
     --> 593         return self._really_apply(util._execute, args=(code,), block=block, targets=targets)
         594 
         595     def run(self, filename, targets=None, block=None):
 
     /home/user/<string> in _really_apply(self, f, args, kwargs, targets, block, track)
 
     /path/to/site-packages/IPython/parallel/client/view.pyc in sync_results(f, self, *args, **kwargs)
          55 def sync_results(f, self, *args, **kwargs):
          56     """sync relevant results from self.client to our results attribute."""
     ---> 57     ret = f(self, *args, **kwargs)
          58     delta = self.outstanding.difference(self.client.outstanding)
          59     completed = self.outstanding.intersection(delta)
 
     /home/user/<string> in _really_apply(self, f, args, kwargs, targets, block, track)
 
     /path/to/site-packages/IPython/parallel/client/view.pyc in save_ids(f, self, *args, **kwargs)
          44     n_previous = len(self.client.history)
          45     try:
     ---> 46         ret = f(self, *args, **kwargs)
          47     finally:
          48         nmsgs = len(self.client.history) - n_previous
 
     /path/to/site-packages/IPython/parallel/client/view.pyc in _really_apply(self, f, args, kwargs, targets, block, track)
         529         if block:
         530             try:
     --> 531                 return ar.get()
         532             except KeyboardInterrupt:
         533                 pass
 
     /path/to/site-packages/IPython/parallel/client/asyncresult.pyc in get(self, timeout)
         101                 return self._result
         102             else:
     --> 103                 raise self._exception
         104         else:
         105             raise error.TimeoutError("Result not ready.")
 
     CompositeError: one or more exceptions from call to method: _execute
     [0:apply]: ZeroDivisionError: integer division or modulo by zero
     [1:apply]: ZeroDivisionError: integer division or modulo by zero
     [2:apply]: ZeroDivisionError: integer division or modulo by zero
     [3:apply]: ZeroDivisionError: integer division or modulo by zero
 
 Notice how the error message printed when :exc:`CompositeError` is raised has
 information about the individual exceptions that were raised on each engine.
 If you want, you can even raise one of these original exceptions:
 
 .. sourcecode:: ipython
 
     In [80]: try:
        ....:     dview.execute('1/0')
        ....: except parallel.error.CompositeError, e:
        ....:     e.raise_exception()
        ....:     
        ....:     
     ---------------------------------------------------------------------------
     RemoteError                               Traceback (most recent call last)
     /home/user/<ipython-input-17-8597e7e39858> in <module>()
           2     dview.execute('1/0')
           3 except CompositeError as e:
     ----> 4     e.raise_exception()
 
     /path/to/site-packages/IPython/parallel/error.pyc in raise_exception(self, excid)
         266             raise IndexError("an exception with index %i does not exist"%excid)
         267         else:
     --> 268             raise RemoteError(en, ev, etb, ei)
         269 
         270 
 
     RemoteError: ZeroDivisionError(integer division or modulo by zero)
     Traceback (most recent call last):
       File "/path/to/site-packages/IPython/parallel/engine/streamkernel.py", line 330, in apply_request
         exec code in working,working
       File "<string>", line 1, in <module>
       File "/path/to/site-packages/IPython/parallel/util.py", line 354, in _execute
         exec code in globals()
       File "<string>", line 1, in <module>
     ZeroDivisionError: integer division or modulo by zero
 
 If you are working in IPython, you can simple type ``%debug`` after one of
 these :exc:`CompositeError` exceptions is raised, and inspect the exception
 instance:
 
 .. sourcecode:: ipython
 
     In [81]: dview.execute('1/0')
     ---------------------------------------------------------------------------
     CompositeError                            Traceback (most recent call last)
     /home/user/<ipython-input-10-5d56b303a66c> in <module>()
     ----> 1 dview.execute('1/0')
 
     /path/to/site-packages/IPython/parallel/client/view.pyc in execute(self, code, targets, block)
         591                 default: self.block
         592         """
     --> 593         return self._really_apply(util._execute, args=(code,), block=block, targets=targets)
         594 
         595     def run(self, filename, targets=None, block=None):
 
     /home/user/<string> in _really_apply(self, f, args, kwargs, targets, block, track)
 
     /path/to/site-packages/IPython/parallel/client/view.pyc in sync_results(f, self, *args, **kwargs)
          55 def sync_results(f, self, *args, **kwargs):
          56     """sync relevant results from self.client to our results attribute."""
     ---> 57     ret = f(self, *args, **kwargs)
          58     delta = self.outstanding.difference(self.client.outstanding)
          59     completed = self.outstanding.intersection(delta)
 
     /home/user/<string> in _really_apply(self, f, args, kwargs, targets, block, track)
 
     /path/to/site-packages/IPython/parallel/client/view.pyc in save_ids(f, self, *args, **kwargs)
          44     n_previous = len(self.client.history)
          45     try:
     ---> 46         ret = f(self, *args, **kwargs)
          47     finally:
          48         nmsgs = len(self.client.history) - n_previous
 
     /path/to/site-packages/IPython/parallel/client/view.pyc in _really_apply(self, f, args, kwargs, targets, block, track)
         529         if block:
         530             try:
     --> 531                 return ar.get()
         532             except KeyboardInterrupt:
         533                 pass
 
     /path/to/site-packages/IPython/parallel/client/asyncresult.pyc in get(self, timeout)
         101                 return self._result
         102             else:
     --> 103                 raise self._exception
         104         else:
         105             raise error.TimeoutError("Result not ready.")
 
     CompositeError: one or more exceptions from call to method: _execute
     [0:apply]: ZeroDivisionError: integer division or modulo by zero
     [1:apply]: ZeroDivisionError: integer division or modulo by zero
     [2:apply]: ZeroDivisionError: integer division or modulo by zero
     [3:apply]: ZeroDivisionError: integer division or modulo by zero
     
     In [82]: %debug
     > /path/to/site-packages/IPython/parallel/client/asyncresult.py(103)get()
         102             else:
     --> 103                 raise self._exception
         104         else:
     
     # With the debugger running, self._exception is the exceptions instance.  We can tab complete
     # on it and see the extra methods that are available.
     ipdb> self._exception.<tab>
     e.__class__         e.__getitem__       e.__new__           e.__setstate__      e.args
     e.__delattr__       e.__getslice__      e.__reduce__        e.__str__           e.elist
     e.__dict__          e.__hash__          e.__reduce_ex__     e.__weakref__       e.message
     e.__doc__           e.__init__          e.__repr__          e._get_engine_str   e.print_tracebacks
     e.__getattribute__  e.__module__        e.__setattr__       e._get_traceback    e.raise_exception
     ipdb> self._exception.print_tracebacks()
     [0:apply]: 
     Traceback (most recent call last):
       File "/path/to/site-packages/IPython/parallel/engine/streamkernel.py", line 330, in apply_request
         exec code in working,working
       File "<string>", line 1, in <module>
       File "/path/to/site-packages/IPython/parallel/util.py", line 354, in _execute
         exec code in globals()
       File "<string>", line 1, in <module>
     ZeroDivisionError: integer division or modulo by zero
 
 
     [1:apply]: 
     Traceback (most recent call last):
       File "/path/to/site-packages/IPython/parallel/engine/streamkernel.py", line 330, in apply_request
         exec code in working,working
       File "<string>", line 1, in <module>
       File "/path/to/site-packages/IPython/parallel/util.py", line 354, in _execute
         exec code in globals()
       File "<string>", line 1, in <module>
     ZeroDivisionError: integer division or modulo by zero
 
 
     [2:apply]: 
     Traceback (most recent call last):
       File "/path/to/site-packages/IPython/parallel/engine/streamkernel.py", line 330, in apply_request
         exec code in working,working
       File "<string>", line 1, in <module>
       File "/path/to/site-packages/IPython/parallel/util.py", line 354, in _execute
         exec code in globals()
       File "<string>", line 1, in <module>
     ZeroDivisionError: integer division or modulo by zero
 
 
     [3:apply]: 
     Traceback (most recent call last):
       File "/path/to/site-packages/IPython/parallel/engine/streamkernel.py", line 330, in apply_request
         exec code in working,working
       File "<string>", line 1, in <module>
       File "/path/to/site-packages/IPython/parallel/util.py", line 354, in _execute
         exec code in globals()
       File "<string>", line 1, in <module>
     ZeroDivisionError: integer division or modulo by zero
 
 
 All of this same error handling magic even works in non-blocking mode:
 
 .. sourcecode:: ipython
 
     In [83]: dview.block=False
 
     In [84]: ar = dview.execute('1/0')
 
     In [85]: ar.get()
     ---------------------------------------------------------------------------
     CompositeError                            Traceback (most recent call last)
     /home/user/<ipython-input-21-8531eb3d26fb> in <module>()
     ----> 1 ar.get()
 
     /path/to/site-packages/IPython/parallel/client/asyncresult.pyc in get(self, timeout)
         101                 return self._result
         102             else:
     --> 103                 raise self._exception
         104         else:
         105             raise error.TimeoutError("Result not ready.")
 
     CompositeError: one or more exceptions from call to method: _execute
     [0:apply]: ZeroDivisionError: integer division or modulo by zero
     [1:apply]: ZeroDivisionError: integer division or modulo by zero
     [2:apply]: ZeroDivisionError: integer division or modulo by zero
     [3:apply]: ZeroDivisionError: integer division or modulo by zero
 

docs/source/parallel/parallel_task.txt

@@ -1,492 +1,471 b''
 .. _parallel_task:
 
 ==========================
 The IPython task interface
 ==========================
 
 The task interface to the cluster presents the engines as a fault tolerant,
 dynamic load-balanced system of workers. Unlike the multiengine interface, in
 the task interface the user have no direct access to individual engines. By
 allowing the IPython scheduler to assign work, this interface is simultaneously
 simpler and more powerful.
 
 Best of all, the user can use both of these interfaces running at the same time
 to take advantage of their respective strengths. When the user can break up
 the user's work into segments that do not depend on previous execution, the
 task interface is ideal. But it also has more power and flexibility, allowing
 the user to guide the distribution of jobs, without having to assign tasks to
 engines explicitly.
 
 Starting the IPython controller and engines
 ===========================================
 
 To follow along with this tutorial, you will need to start the IPython
 controller and four IPython engines. The simplest way of doing this is to use
 the :command:`ipcluster` command::
 
 	$ ipcluster start -n 4
 	
 For more detailed information about starting the controller and engines, see
 our :ref:`introduction <parallel_overview>` to using IPython for parallel computing.
 
 Creating a ``LoadBalancedView`` instance
 ========================================
 
 The first step is to import the IPython :mod:`IPython.parallel`
 module and then create a :class:`.Client` instance, and we will also be using
 a :class:`LoadBalancedView`, here called `lview`:
 
 .. sourcecode:: ipython
 
     In [1]: from IPython.parallel import Client
 
     In [2]: rc = Client()
     
 
 This form assumes that the controller was started on localhost with default
 configuration. If not, the location of the controller must be given as an
 argument to the constructor:
 
 .. sourcecode:: ipython
 
     # for a visible LAN controller listening on an external port:
     In [2]: rc = Client('tcp://192.168.1.16:10101')
     # or to connect with a specific profile you have set up:
     In [3]: rc = Client(profile='mpi')
 
 For load-balanced execution, we will make use of a :class:`LoadBalancedView` object, which can
 be constructed via the client's :meth:`load_balanced_view` method:
 
 .. sourcecode:: ipython
 
     In [4]: lview = rc.load_balanced_view() # default load-balanced view
 
 .. seealso::
 
     For more information, see the in-depth explanation of :ref:`Views <parallel_details>`.
 
 
 Quick and easy parallelism
 ==========================
 
 In many cases, you simply want to apply a Python function to a sequence of
 objects, but *in parallel*. Like the multiengine interface, these can be
 implemented via the task interface. The exact same tools can perform these
 actions in load-balanced ways as well as multiplexed ways: a parallel version
 of :func:`map` and :func:`@parallel` function decorator. If one specifies the
 argument `balanced=True`, then they are dynamically load balanced. Thus, if the
 execution time per item varies significantly, you should use the versions in
 the task interface.
 
 Parallel map
 ------------
 
 To load-balance :meth:`map`,simply use a LoadBalancedView:
 
 .. sourcecode:: ipython
     
     In [62]: lview.block = True
     
     In [63]: serial_result = map(lambda x:x**10, range(32))
 
     In [64]: parallel_result = lview.map(lambda x:x**10, range(32))
 
     In [65]: serial_result==parallel_result
     Out[65]: True
 
 Parallel function decorator
 ---------------------------
 
 Parallel functions are just like normal function, but they can be called on
 sequences and *in parallel*. The multiengine interface provides a decorator
 that turns any Python function into a parallel function:
 
 .. sourcecode:: ipython
 
     In [10]: @lview.parallel()
        ....: def f(x):
        ....:     return 10.0*x**4
        ....:
 
     In [11]: f.map(range(32))    # this is done in parallel
     Out[11]: [0.0,10.0,160.0,...]
 
-.. _parallel_taskmap:
-
-Map results are iterable!
--------------------------
-
-When an AsyncResult object actually maps multiple results (e.g. the :class:`~AsyncMapResult`
-object), you can actually iterate through them, and act on the results as they arrive:
-
-.. literalinclude:: ../../examples/parallel/itermapresult.py
-    :language: python
-    :lines: 9-34
-
-.. seealso::
-
-    When AsyncResult or the AsyncMapResult don't provide what you need (for instance,
-    handling individual results as they arrive, but with metadata), you can always
-    just split the original result's ``msg_ids`` attribute, and handle them as you like.
-    
-    For an example of this, see :file:`docs/examples/parallel/customresult.py`
-    
-
 .. _parallel_dependencies:
 
 Dependencies
 ============
 
 Often, pure atomic load-balancing is too primitive for your work. In these cases, you
 may want to associate some kind of `Dependency` that describes when, where, or whether
 a task can be run.  In IPython, we provide two types of dependencies: 
 `Functional Dependencies`_ and `Graph Dependencies`_
 
 .. note::
 
     It is important to note that the pure ZeroMQ scheduler does not support dependencies,
     and you will see errors or warnings if you try to use dependencies with the pure
     scheduler.
 
 Functional Dependencies
 -----------------------
 
 Functional dependencies are used to determine whether a given engine is capable of running
 a particular task.  This is implemented via a special :class:`Exception` class, 
 :class:`UnmetDependency`, found in `IPython.parallel.error`.  Its use is very simple: 
 if a task fails with an UnmetDependency exception, then the scheduler, instead of relaying
 the error up to the client like any other error, catches the error, and submits the task
 to a different engine.  This will repeat indefinitely, and a task will never be submitted
 to a given engine a second time.
 
 You can manually raise the :class:`UnmetDependency` yourself, but IPython has provided 
 some decorators for facilitating this behavior.
 
 There are two decorators and a class used for functional dependencies:
 
 .. sourcecode:: ipython
 
     In [9]: from IPython.parallel import depend, require, dependent
 
 @require
 ********
 
 The simplest sort of dependency is requiring that a Python module is available. The
 ``@require`` decorator lets you define a function that will only run on engines where names
 you specify are importable:
 
 .. sourcecode:: ipython
 
     In [10]: @require('numpy', 'zmq')
        ....: def myfunc():
        ....:     return dostuff()
 
 Now, any time you apply :func:`myfunc`, the task will only run on a machine that has
 numpy and pyzmq available, and when :func:`myfunc` is called, numpy and zmq will be imported.
 
 @depend
 *******
 
 The ``@depend`` decorator lets you decorate any function with any *other* function to
 evaluate the dependency. The dependency function will be called at the start of the task,
 and if it returns ``False``, then the dependency will be considered unmet, and the task
 will be assigned to another engine. If the dependency returns *anything other than
 ``False``*, the rest of the task will continue.
 
 .. sourcecode:: ipython
 
     In [10]: def platform_specific(plat):
        ....:    import sys
        ....:    return sys.platform == plat
     
     In [11]: @depend(platform_specific, 'darwin')
        ....: def mactask():
        ....:    do_mac_stuff()
 
     In [12]: @depend(platform_specific, 'nt')
        ....: def wintask():
        ....:    do_windows_stuff()
     
 In this case, any time you apply ``mytask``, it will only run on an OSX machine.
 ``@depend`` is just like ``apply``, in that it has a ``@depend(f,*args,**kwargs)``
 signature.
 
 dependents
 **********
 
 You don't have to use the decorators on your tasks, if for instance you may want
 to run tasks with a single function but varying dependencies, you can directly construct
 the :class:`dependent` object that the decorators use:
 
 .. sourcecode::ipython
 
     In [13]: def mytask(*args):
        ....:    dostuff()
     
     In [14]: mactask = dependent(mytask, platform_specific, 'darwin')
     # this is the same as decorating the declaration of mytask with @depend
     # but you can do it again:
     
     In [15]: wintask = dependent(mytask, platform_specific, 'nt')
     
     # in general:
     In [16]: t = dependent(f, g, *dargs, **dkwargs)
     
     # is equivalent to:
     In [17]: @depend(g, *dargs, **dkwargs)
        ....: def t(a,b,c):
        ....:     # contents of f
 
 Graph Dependencies
 ------------------
 
 Sometimes you want to restrict the time and/or location to run a given task as a function
 of the time and/or location of other tasks. This is implemented via a subclass of
 :class:`set`, called a :class:`Dependency`. A Dependency is just a set of `msg_ids`
 corresponding to tasks, and a few attributes to guide how to decide when the Dependency
 has been met.
 
 The switches we provide for interpreting whether a given dependency set has been met:
 
 any|all
     Whether the dependency is considered met if *any* of the dependencies are done, or
     only after *all* of them have finished.  This is set by a Dependency's :attr:`all`
     boolean attribute, which defaults to ``True``.
 
 success [default: True]
     Whether to consider tasks that succeeded as fulfilling dependencies.
 
 failure [default : False]
     Whether to consider tasks that failed as fulfilling dependencies.
     using `failure=True,success=False` is useful for setting up cleanup tasks, to be run
     only when tasks have failed.
 
 Sometimes you want to run a task after another, but only if that task succeeded. In this case,
 ``success`` should be ``True`` and ``failure`` should be ``False``. However sometimes you may
 not care whether the task succeeds, and always want the second task to run, in which case you
 should use `success=failure=True`. The default behavior is to only use successes.
 
 There are other switches for interpretation that are made at the *task* level.  These are
 specified via keyword arguments to the client's :meth:`apply` method.
 
 after,follow
     You may want to run a task *after* a given set of dependencies have been run and/or
     run it *where* another set of dependencies are met. To support this, every task has an
     `after` dependency to restrict time, and a `follow` dependency to restrict
     destination.
 
 timeout
     You may also want to set a time-limit for how long the scheduler should wait before a
     task's dependencies are met. This is done via a `timeout`, which defaults to 0, which
     indicates that the task should never timeout. If the timeout is reached, and the
     scheduler still hasn't been able to assign the task to an engine, the task will fail
     with a :class:`DependencyTimeout`.
 
 .. note::
 
     Dependencies only work within the task scheduler. You cannot instruct a load-balanced
     task to run after a job submitted via the MUX interface.
 
 The simplest form of Dependencies is with `all=True,success=True,failure=False`. In these cases,
 you can skip using Dependency objects, and just pass msg_ids or AsyncResult objects as the
 `follow` and `after` keywords to :meth:`client.apply`:
 
 .. sourcecode:: ipython
 
     In [14]: client.block=False
 
     In [15]: ar = lview.apply(f, args, kwargs)
 
     In [16]: ar2 = lview.apply(f2)
 
     In [17]: with lview.temp_flags(after=[ar,ar2]):
        ....:    ar3 = lview.apply(f3)
 
     In [18]: with lview.temp_flags(follow=[ar], timeout=2.5)
        ....:    ar4 = lview.apply(f3)
 
 .. seealso::
 
     Some parallel workloads can be described as a `Directed Acyclic Graph
     <http://en.wikipedia.org/wiki/Directed_acyclic_graph>`_, or DAG. See :ref:`DAG
     Dependencies <dag_dependencies>` for an example demonstrating how to use map a NetworkX DAG
     onto task dependencies.
 
 
 Impossible Dependencies
 ***********************
 
 The schedulers do perform some analysis on graph dependencies to determine whether they
 are not possible to be met. If the scheduler does discover that a dependency cannot be
 met, then the task will fail with an :class:`ImpossibleDependency` error. This way, if the
 scheduler realized that a task can never be run, it won't sit indefinitely in the
 scheduler clogging the pipeline.
 
 The basic cases that are checked:
 
 * depending on nonexistent messages
 * `follow` dependencies were run on more than one machine and `all=True`
 * any dependencies failed and `all=True,success=True,failures=False`
 * all dependencies failed and `all=False,success=True,failure=False`
 
 .. warning::
 
     This analysis has not been proven to be rigorous, so it is likely possible for tasks
     to become impossible to run in obscure situations, so a timeout may be a good choice.
 
 
 Retries and Resubmit
 ====================
 
 Retries
 -------
 
 Another flag for tasks is `retries`.  This is an integer, specifying how many times
 a task should be resubmitted after failure.  This is useful for tasks that should still run
 if their engine was shutdown, or may have some statistical chance of failing.  The default
 is to not retry tasks.
 
 Resubmit
 --------
 
 Sometimes you may want to re-run a task. This could be because it failed for some reason, and
 you have fixed the error, or because you want to restore the cluster to an interrupted state.
 For this, the :class:`Client` has a :meth:`rc.resubmit` method.  This simply takes one or more
 msg_ids, and returns an :class:`AsyncHubResult` for the result(s).  You cannot resubmit
 a task that is pending - only those that have finished, either successful or unsuccessful.
 
 .. _parallel_schedulers:
 
 Schedulers
 ==========
 
 There are a variety of valid ways to determine where jobs should be assigned in a 
 load-balancing situation.  In IPython, we support several standard schemes, and
 even make it easy to define your own.  The scheme can be selected via the ``scheme``
 argument to :command:`ipcontroller`, or in the :attr:`TaskScheduler.schemename` attribute
 of a controller config object.
 
 The built-in routing schemes:
 
 To select one of these schemes, simply do::
 
     $ ipcontroller --scheme=<schemename>
     for instance:
     $ ipcontroller --scheme=lru
 
 lru: Least Recently Used
 
     Always assign work to the least-recently-used engine.  A close relative of
     round-robin, it will be fair with respect to the number of tasks, agnostic 
     with respect to runtime of each task.
 
 plainrandom: Plain Random
 
     Randomly picks an engine on which to run.
 
 twobin: Two-Bin Random
 
     **Requires numpy**
     
     Pick two engines at random, and use the LRU of the two. This is known to be better
     than plain random in many cases, but requires a small amount of computation.
 
 leastload: Least Load
 
     **This is the default scheme**
 
     Always assign tasks to the engine with the fewest outstanding tasks (LRU breaks tie).
 
 weighted: Weighted Two-Bin Random
 
     **Requires numpy**
     
     Pick two engines at random using the number of outstanding tasks as inverse weights,
     and use the one with the lower load.
 
 Greedy Assignment
 -----------------
 
 Tasks are assigned greedily as they are submitted. If their dependencies are
 met, they will be assigned to an engine right away, and multiple tasks can be
 assigned to an engine at a given time. This limit is set with the
 ``TaskScheduler.hwm`` (high water mark) configurable:
 
 .. sourcecode:: python
 
     # the most common choices are:
     c.TaskSheduler.hwm = 0 # (minimal latency, default in IPython ≤ 0.12)
     # or
     c.TaskScheduler.hwm = 1 # (most-informed balancing, default in > 0.12)
 
 In IPython ≤ 0.12,the default is 0, or no-limit. That is, there is no limit to the number of
 tasks that can be outstanding on a given engine. This greatly benefits the
 latency of execution, because network traffic can be hidden behind computation.
 However, this means that workload is assigned without knowledge of how long
 each task might take, and can result in poor load-balancing, particularly for
 submitting a collection of heterogeneous tasks all at once. You can limit this
 effect by setting hwm to a positive integer, 1 being maximum load-balancing (a
 task will never be waiting if there is an idle engine), and any larger number
 being a compromise between load-balance and latency-hiding.
 
 In practice, some users have been confused by having this optimization on by
 default, and the default value has been changed to 1. This can be slower,
 but has more obvious behavior and won't result in assigning too many tasks to
 some engines in heterogeneous cases.
 
 
 Pure ZMQ Scheduler
 ------------------
 
 For maximum throughput, the 'pure' scheme is not Python at all, but a C-level
 :class:`MonitoredQueue` from PyZMQ, which uses a ZeroMQ ``DEALER`` socket to perform all
 load-balancing. This scheduler does not support any of the advanced features of the Python
 :class:`.Scheduler`.
 
 Disabled features when using the ZMQ Scheduler:
 
 * Engine unregistration
     Task farming will be disabled if an engine unregisters.  
     Further, if an engine is unregistered during computation, the scheduler may not recover.
 * Dependencies
     Since there is no Python logic inside the Scheduler, routing decisions cannot be made
     based on message content.
 * Early destination notification
     The Python schedulers know which engine gets which task, and notify the Hub.  This 
     allows graceful handling of Engines coming and going.  There is no way to know
     where ZeroMQ messages have gone, so there is no way to know what tasks are on which
     engine until they *finish*.  This makes recovery from engine shutdown very difficult.
 
 
 .. note::
 
     TODO: performance comparisons
 
 
 
 
 More details
 ============
 
 The :class:`LoadBalancedView` has many more powerful features that allow quite a bit
 of flexibility in how tasks are defined and run. The next places to look are
 in the following classes:
 
 * :class:`~IPython.parallel.client.view.LoadBalancedView`
 * :class:`~IPython.parallel.client.asyncresult.AsyncResult`
 * :meth:`~IPython.parallel.client.view.LoadBalancedView.apply`
 * :mod:`~IPython.parallel.controller.dependency`
 
 The following is an overview of how to use these classes together:
 
 1. Create a :class:`Client` and :class:`LoadBalancedView`
 2. Define some functions to be run as tasks
 3. Submit your tasks to using the :meth:`apply` method of your
    :class:`LoadBalancedView` instance.
 4. Use :meth:`.Client.get_result` to get the results of the
    tasks, or use the :meth:`AsyncResult.get` method of the results to wait
    for and then receive the results.
 
 .. seealso::
 
     A demo of :ref:`DAG Dependencies <dag_dependencies>` with NetworkX and IPython.