enable non-copying sends in push

zero-copy of numpy arrays only applies to first-level in args/kwargs, but push took a dict as arg.  This simply changes push to pass the same ns as kwargs, so numpy arrays will be discovered.

IPython/parallel/client/view.py

@@ -1,1069 +1,1069 b''
 """Views of remote engines.
 
 Authors:
 
 * Min RK
 """
 #-----------------------------------------------------------------------------
 #  Copyright (C) 2010-2011  The IPython Development Team
 #
 #  Distributed under the terms of the BSD License.  The full license is in
 #  the file COPYING, distributed as part of this software.
 #-----------------------------------------------------------------------------
 
 #-----------------------------------------------------------------------------
 # Imports
 #-----------------------------------------------------------------------------
 
 import imp
 import sys
 import warnings
 from contextlib import contextmanager
 from types import ModuleType
 
 import zmq
 
 from IPython.testing.skipdoctest import skip_doctest
 from IPython.utils.traitlets import (
     HasTraits, Any, Bool, List, Dict, Set, Instance, CFloat, Integer
 )
 from IPython.external.decorator import decorator
 
 from IPython.parallel import util
 from IPython.parallel.controller.dependency import Dependency, dependent
 
 from . import map as Map
 from .asyncresult import AsyncResult, AsyncMapResult
 from .remotefunction import ParallelFunction, parallel, remote, getname
 
 #-----------------------------------------------------------------------------
 # Decorators
 #-----------------------------------------------------------------------------
 
 @decorator
 def save_ids(f, self, *args, **kwargs):
     """Keep our history and outstanding attributes up to date after a method call."""
     n_previous = len(self.client.history)
     try:
         ret = f(self, *args, **kwargs)
     finally:
         nmsgs = len(self.client.history) - n_previous
         msg_ids = self.client.history[-nmsgs:]
         self.history.extend(msg_ids)
         map(self.outstanding.add, msg_ids)
     return ret
 
 @decorator
 def sync_results(f, self, *args, **kwargs):
     """sync relevant results from self.client to our results attribute."""
     ret = f(self, *args, **kwargs)
     delta = self.outstanding.difference(self.client.outstanding)
     completed = self.outstanding.intersection(delta)
     self.outstanding = self.outstanding.difference(completed)
     for msg_id in completed:
         self.results[msg_id] = self.client.results[msg_id]
     return ret
 
 @decorator
 def spin_after(f, self, *args, **kwargs):
     """call spin after the method."""
     ret = f(self, *args, **kwargs)
     self.spin()
     return ret
 
 #-----------------------------------------------------------------------------
 # Classes
 #-----------------------------------------------------------------------------
 
 @skip_doctest
 class View(HasTraits):
     """Base View class for more convenint apply(f,*args,**kwargs) syntax via attributes.
 
     Don't use this class, use subclasses.
 
     Methods
     -------
 
     spin
         flushes incoming results and registration state changes
         control methods spin, and requesting `ids` also ensures up to date
 
     wait
         wait on one or more msg_ids
 
     execution methods
         apply
         legacy: execute, run
 
     data movement
         push, pull, scatter, gather
 
     query methods
         get_result, queue_status, purge_results, result_status
 
     control methods
         abort, shutdown
 
     """
     # flags
     block=Bool(False)
     track=Bool(True)
     targets = Any()
 
     history=List()
     outstanding = Set()
     results = Dict()
     client = Instance('IPython.parallel.Client')
 
     _socket = Instance('zmq.Socket')
     _flag_names = List(['targets', 'block', 'track'])
     _targets = Any()
     _idents = Any()
 
     def __init__(self, client=None, socket=None, **flags):
         super(View, self).__init__(client=client, _socket=socket)
         self.block = client.block
 
         self.set_flags(**flags)
 
         assert not self.__class__ is View, "Don't use base View objects, use subclasses"
 
 
     def __repr__(self):
         strtargets = str(self.targets)
         if len(strtargets) > 16:
             strtargets = strtargets[:12]+'...]'
         return "<%s %s>"%(self.__class__.__name__, strtargets)
 
     def set_flags(self, **kwargs):
         """set my attribute flags by keyword.
 
         Views determine behavior with a few attributes (`block`, `track`, etc.).
         These attributes can be set all at once by name with this method.
 
         Parameters
         ----------
 
         block : bool
             whether to wait for results
         track : bool
             whether to create a MessageTracker to allow the user to
             safely edit after arrays and buffers during non-copying
             sends.
         """
         for name, value in kwargs.iteritems():
             if name not in self._flag_names:
                 raise KeyError("Invalid name: %r"%name)
             else:
                 setattr(self, name, value)
 
     @contextmanager
     def temp_flags(self, **kwargs):
         """temporarily set flags, for use in `with` statements.
 
         See set_flags for permanent setting of flags
 
         Examples
         --------
 
         >>> view.track=False
         ...
         >>> with view.temp_flags(track=True):
         ...    ar = view.apply(dostuff, my_big_array)
         ...    ar.tracker.wait() # wait for send to finish
         >>> view.track
         False
 
         """
         # preflight: save flags, and set temporaries
         saved_flags = {}
         for f in self._flag_names:
             saved_flags[f] = getattr(self, f)
         self.set_flags(**kwargs)
         # yield to the with-statement block
         try:
             yield
         finally:
             # postflight: restore saved flags
             self.set_flags(**saved_flags)
 
 
     #----------------------------------------------------------------
     # apply
     #----------------------------------------------------------------
 
     @sync_results
     @save_ids
     def _really_apply(self, f, args, kwargs, block=None, **options):
         """wrapper for client.send_apply_message"""
         raise NotImplementedError("Implement in subclasses")
 
     def apply(self, f, *args, **kwargs):
         """calls f(*args, **kwargs) on remote engines, returning the result.
 
         This method sets all apply flags via this View's attributes.
 
         if self.block is False:
             returns AsyncResult
         else:
             returns actual result of f(*args, **kwargs)
         """
         return self._really_apply(f, args, kwargs)
 
     def apply_async(self, f, *args, **kwargs):
         """calls f(*args, **kwargs) on remote engines in a nonblocking manner.
 
         returns AsyncResult
         """
         return self._really_apply(f, args, kwargs, block=False)
 
     @spin_after
     def apply_sync(self, f, *args, **kwargs):
         """calls f(*args, **kwargs) on remote engines in a blocking manner,
          returning the result.
 
         returns: actual result of f(*args, **kwargs)
         """
         return self._really_apply(f, args, kwargs, block=True)
 
     #----------------------------------------------------------------
     # wrappers for client and control methods
     #----------------------------------------------------------------
     @sync_results
     def spin(self):
         """spin the client, and sync"""
         self.client.spin()
 
     @sync_results
     def wait(self, jobs=None, timeout=-1):
         """waits on one or more `jobs`, for up to `timeout` seconds.
 
         Parameters
         ----------
 
         jobs : int, str, or list of ints and/or strs, or one or more AsyncResult objects
                 ints are indices to self.history
                 strs are msg_ids
                 default: wait on all outstanding messages
         timeout : float
                 a time in seconds, after which to give up.
                 default is -1, which means no timeout
 
         Returns
         -------
 
         True : when all msg_ids are done
         False : timeout reached, some msg_ids still outstanding
         """
         if jobs is None:
             jobs = self.history
         return self.client.wait(jobs, timeout)
 
     def abort(self, jobs=None, targets=None, block=None):
         """Abort jobs on my engines.
 
         Parameters
         ----------
 
         jobs : None, str, list of strs, optional
             if None: abort all jobs.
             else: abort specific msg_id(s).
         """
         block = block if block is not None else self.block
         targets = targets if targets is not None else self.targets
         jobs = jobs if jobs is not None else list(self.outstanding)
         
         return self.client.abort(jobs=jobs, targets=targets, block=block)
 
     def queue_status(self, targets=None, verbose=False):
         """Fetch the Queue status of my engines"""
         targets = targets if targets is not None else self.targets
         return self.client.queue_status(targets=targets, verbose=verbose)
 
     def purge_results(self, jobs=[], targets=[]):
         """Instruct the controller to forget specific results."""
         if targets is None or targets == 'all':
             targets = self.targets
         return self.client.purge_results(jobs=jobs, targets=targets)
 
     def shutdown(self, targets=None, restart=False, hub=False, block=None):
         """Terminates one or more engine processes, optionally including the hub.
         """
         block = self.block if block is None else block
         if targets is None or targets == 'all':
             targets = self.targets
         return self.client.shutdown(targets=targets, restart=restart, hub=hub, block=block)
 
     @spin_after
     def get_result(self, indices_or_msg_ids=None):
         """return one or more results, specified by history index or msg_id.
 
         See client.get_result for details.
 
         """
 
         if indices_or_msg_ids is None:
             indices_or_msg_ids = -1
         if isinstance(indices_or_msg_ids, int):
             indices_or_msg_ids = self.history[indices_or_msg_ids]
         elif isinstance(indices_or_msg_ids, (list,tuple,set)):
             indices_or_msg_ids = list(indices_or_msg_ids)
             for i,index in enumerate(indices_or_msg_ids):
                 if isinstance(index, int):
                     indices_or_msg_ids[i] = self.history[index]
         return self.client.get_result(indices_or_msg_ids)
 
     #-------------------------------------------------------------------
     # Map
     #-------------------------------------------------------------------
 
     def map(self, f, *sequences, **kwargs):
         """override in subclasses"""
         raise NotImplementedError
 
     def map_async(self, f, *sequences, **kwargs):
         """Parallel version of builtin `map`, using this view's engines.
 
         This is equivalent to map(...block=False)
 
         See `self.map` for details.
         """
         if 'block' in kwargs:
             raise TypeError("map_async doesn't take a `block` keyword argument.")
         kwargs['block'] = False
         return self.map(f,*sequences,**kwargs)
 
     def map_sync(self, f, *sequences, **kwargs):
         """Parallel version of builtin `map`, using this view's engines.
 
         This is equivalent to map(...block=True)
 
         See `self.map` for details.
         """
         if 'block' in kwargs:
             raise TypeError("map_sync doesn't take a `block` keyword argument.")
         kwargs['block'] = True
         return self.map(f,*sequences,**kwargs)
 
     def imap(self, f, *sequences, **kwargs):
         """Parallel version of `itertools.imap`.
 
         See `self.map` for details.
 
         """
 
         return iter(self.map_async(f,*sequences, **kwargs))
 
     #-------------------------------------------------------------------
     # Decorators
     #-------------------------------------------------------------------
 
     def remote(self, block=True, **flags):
         """Decorator for making a RemoteFunction"""
         block = self.block if block is None else block
         return remote(self, block=block, **flags)
 
     def parallel(self, dist='b', block=None, **flags):
         """Decorator for making a ParallelFunction"""
         block = self.block if block is None else block
         return parallel(self, dist=dist, block=block, **flags)
 
 @skip_doctest
 class DirectView(View):
     """Direct Multiplexer View of one or more engines.
 
     These are created via indexed access to a client:
 
     >>> dv_1 = client[1]
     >>> dv_all = client[:]
     >>> dv_even = client[::2]
     >>> dv_some = client[1:3]
 
     This object provides dictionary access to engine namespaces:
 
     # push a=5:
     >>> dv['a'] = 5
     # pull 'foo':
     >>> db['foo']
 
     """
 
     def __init__(self, client=None, socket=None, targets=None):
         super(DirectView, self).__init__(client=client, socket=socket, targets=targets)
 
     @property
     def importer(self):
         """sync_imports(local=True) as a property.
 
         See sync_imports for details.
 
         """
         return self.sync_imports(True)
 
     @contextmanager
     def sync_imports(self, local=True, quiet=False):
         """Context Manager for performing simultaneous local and remote imports.
 
         'import x as y' will *not* work.  The 'as y' part will simply be ignored.
         
         If `local=True`, then the package will also be imported locally.
         
         If `quiet=True`, no output will be produced when attempting remote 
         imports. 
         
         Note that remote-only (`local=False`) imports have not been implemented.
         
         >>> with view.sync_imports():
         ...    from numpy import recarray
         importing recarray from numpy on engine(s)
 
         """
         import __builtin__
         local_import = __builtin__.__import__
         modules = set()
         results = []
         @util.interactive
         def remote_import(name, fromlist, level):
             """the function to be passed to apply, that actually performs the import
             on the engine, and loads up the user namespace.
             """
             import sys
             user_ns = globals()
             mod = __import__(name, fromlist=fromlist, level=level)
             if fromlist:
                 for key in fromlist:
                     user_ns[key] = getattr(mod, key)
             else:
                 user_ns[name] = sys.modules[name]
 
         def view_import(name, globals={}, locals={}, fromlist=[], level=-1):
             """the drop-in replacement for __import__, that optionally imports
             locally as well.
             """
             # don't override nested imports
             save_import = __builtin__.__import__
             __builtin__.__import__ = local_import
 
             if imp.lock_held():
                 # this is a side-effect import, don't do it remotely, or even
                 # ignore the local effects
                 return local_import(name, globals, locals, fromlist, level)
 
             imp.acquire_lock()
             if local:
                 mod = local_import(name, globals, locals, fromlist, level)
             else:
                 raise NotImplementedError("remote-only imports not yet implemented")
             imp.release_lock()
 
             key = name+':'+','.join(fromlist or [])
             if level == -1 and key not in modules:
                 modules.add(key)
                 if not quiet:
                     if fromlist:
                         print "importing %s from %s on engine(s)"%(','.join(fromlist), name)
                     else:
                         print "importing %s on engine(s)"%name
                 results.append(self.apply_async(remote_import, name, fromlist, level))
             # restore override
             __builtin__.__import__ = save_import
 
             return mod
 
         # override __import__
         __builtin__.__import__ = view_import
         try:
             # enter the block
             yield
         except ImportError:
             if local:
                 raise
             else:
                 # ignore import errors if not doing local imports
                 pass
         finally:
             # always restore __import__
             __builtin__.__import__ = local_import
 
         for r in results:
             # raise possible remote ImportErrors here
             r.get()
 
 
     @sync_results
     @save_ids
     def _really_apply(self, f, args=None, kwargs=None, targets=None, block=None, track=None):
         """calls f(*args, **kwargs) on remote engines, returning the result.
 
         This method sets all of `apply`'s flags via this View's attributes.
 
         Parameters
         ----------
 
         f : callable
 
         args : list [default: empty]
 
         kwargs : dict [default: empty]
 
         targets : target list [default: self.targets]
             where to run
         block : bool [default: self.block]
             whether to block
         track : bool [default: self.track]
             whether to ask zmq to track the message, for safe non-copying sends
 
         Returns
         -------
 
         if self.block is False:
             returns AsyncResult
         else:
             returns actual result of f(*args, **kwargs) on the engine(s)
             This will be a list of self.targets is also a list (even length 1), or
             the single result if self.targets is an integer engine id
         """
         args = [] if args is None else args
         kwargs = {} if kwargs is None else kwargs
         block = self.block if block is None else block
         track = self.track if track is None else track
         targets = self.targets if targets is None else targets
 
         _idents = self.client._build_targets(targets)[0]
         msg_ids = []
         trackers = []
         for ident in _idents:
             msg = self.client.send_apply_message(self._socket, f, args, kwargs, track=track,
                                     ident=ident)
             if track:
                 trackers.append(msg['tracker'])
             msg_ids.append(msg['header']['msg_id'])
         tracker = None if track is False else zmq.MessageTracker(*trackers)
         ar = AsyncResult(self.client, msg_ids, fname=getname(f), targets=targets, tracker=tracker)
         if block:
             try:
                 return ar.get()
             except KeyboardInterrupt:
                 pass
         return ar
 
     @spin_after
     def map(self, f, *sequences, **kwargs):
         """view.map(f, *sequences, block=self.block) => list|AsyncMapResult
 
         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.
 
         Parameters
         ----------
 
         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
             whether to wait for the result or not [default self.block]
 
         Returns
         -------
 
         if block=False:
             AsyncMapResult
                 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:
             list
                 the result of map(f,*sequences)
         """
 
         block = kwargs.pop('block', self.block)
         for k in kwargs.keys():
             if k not in ['block', 'track']:
                 raise TypeError("invalid keyword arg, %r"%k)
 
         assert len(sequences) > 0, "must have some sequences to map onto!"
         pf = ParallelFunction(self, f, block=block, **kwargs)
         return pf.map(*sequences)
 
     def execute(self, code, targets=None, block=None):
         """Executes `code` on `targets` in blocking or nonblocking manner.
 
         ``execute`` is always `bound` (affects engine namespace)
 
         Parameters
         ----------
 
         code : str
                 the code string to be executed
         block : bool
                 whether or not to wait until done to return
                 default: self.block
         """
         return self._really_apply(util._execute, args=(code,), block=block, targets=targets)
 
     def run(self, filename, targets=None, block=None):
         """Execute contents of `filename` on my engine(s).
 
         This simply reads the contents of the file and calls `execute`.
 
         Parameters
         ----------
 
         filename : str
                 The path to the file
         targets : int/str/list of ints/strs
                 the engines on which to execute
                 default : all
         block : bool
                 whether or not to wait until done
                 default: self.block
 
         """
         with open(filename, 'r') as f:
             # add newline in case of trailing indented whitespace
             # which will cause SyntaxError
             code = f.read()+'\n'
         return self.execute(code, block=block, targets=targets)
 
     def update(self, ns):
         """update remote namespace with dict `ns`
 
         See `push` for details.
         """
         return self.push(ns, block=self.block, track=self.track)
 
     def push(self, ns, targets=None, block=None, track=None):
         """update remote namespace with dict `ns`
 
         Parameters
         ----------
 
         ns : dict
             dict of keys with which to update engine namespace(s)
         block : bool [default : self.block]
             whether to wait to be notified of engine receipt
 
         """
 
         block = block if block is not None else self.block
         track = track if track is not None else self.track
         targets = targets if targets is not None else self.targets
         # applier = self.apply_sync if block else self.apply_async
         if not isinstance(ns, dict):
             raise TypeError("Must be a dict, not %s"%type(ns))
-        return self._really_apply(util._push, (ns,), block=block, track=track, targets=targets)
+        return self._really_apply(util._push, kwargs=ns, block=block, track=track, targets=targets)
 
     def get(self, key_s):
         """get object(s) by `key_s` from remote namespace
 
         see `pull` for details.
         """
         # block = block if block is not None else self.block
         return self.pull(key_s, block=True)
 
     def pull(self, names, targets=None, block=None):
         """get object(s) by `name` from remote namespace
 
         will return one object if it is a key.
         can also take a list of keys, in which case it will return a list of objects.
         """
         block = block if block is not None else self.block
         targets = targets if targets is not None else self.targets
         applier = self.apply_sync if block else self.apply_async
         if isinstance(names, basestring):
             pass
         elif isinstance(names, (list,tuple,set)):
             for key in names:
                 if not isinstance(key, basestring):
                     raise TypeError("keys must be str, not type %r"%type(key))
         else:
             raise TypeError("names must be strs, not %r"%names)
         return self._really_apply(util._pull, (names,), block=block, targets=targets)
 
     def scatter(self, key, seq, dist='b', flatten=False, targets=None, block=None, track=None):
         """
         Partition a Python sequence and send the partitions to a set of engines.
         """
         block = block if block is not None else self.block
         track = track if track is not None else self.track
         targets = targets if targets is not None else self.targets
 
         mapObject = Map.dists[dist]()
         nparts = len(targets)
         msg_ids = []
         trackers = []
         for index, engineid in enumerate(targets):
             partition = mapObject.getPartition(seq, index, nparts)
             if flatten and len(partition) == 1:
                 ns = {key: partition[0]}
             else:
                 ns = {key: partition}
             r = self.push(ns, block=False, track=track, targets=engineid)
             msg_ids.extend(r.msg_ids)
             if track:
                 trackers.append(r._tracker)
 
         if track:
             tracker = zmq.MessageTracker(*trackers)
         else:
             tracker = None
 
         r = AsyncResult(self.client, msg_ids, fname='scatter', targets=targets, tracker=tracker)
         if block:
             r.wait()
         else:
             return r
 
     @sync_results
     @save_ids
     def gather(self, key, dist='b', targets=None, block=None):
         """
         Gather a partitioned sequence on a set of engines as a single local seq.
         """
         block = block if block is not None else self.block
         targets = targets if targets is not None else self.targets
         mapObject = Map.dists[dist]()
         msg_ids = []
 
         for index, engineid in enumerate(targets):
             msg_ids.extend(self.pull(key, block=False, targets=engineid).msg_ids)
 
         r = AsyncMapResult(self.client, msg_ids, mapObject, fname='gather')
 
         if block:
             try:
                 return r.get()
             except KeyboardInterrupt:
                 pass
         return r
 
     def __getitem__(self, key):
         return self.get(key)
 
     def __setitem__(self,key, value):
         self.update({key:value})
 
     def clear(self, targets=None, block=False):
         """Clear the remote namespaces on my engines."""
         block = block if block is not None else self.block
         targets = targets if targets is not None else self.targets
         return self.client.clear(targets=targets, block=block)
 
     def kill(self, targets=None, block=True):
         """Kill my engines."""
         block = block if block is not None else self.block
         targets = targets if targets is not None else self.targets
         return self.client.kill(targets=targets, block=block)
 
     #----------------------------------------
     # activate for %px,%autopx magics
     #----------------------------------------
     def activate(self):
         """Make this `View` active for parallel magic commands.
 
         IPython has a magic command syntax to work with `MultiEngineClient` objects.
         In a given IPython session there is a single active one.  While
         there can be many `Views` created and used by the user,
         there is only one active one.  The active `View` is used whenever
         the magic commands %px and %autopx are used.
 
         The activate() method is called on a given `View` to make it
         active.  Once this has been done, the magic commands can be used.
         """
 
         try:
             # This is injected into __builtins__.
             ip = get_ipython()
         except NameError:
             print "The IPython parallel magics (%result, %px, %autopx) only work within IPython."
         else:
             pmagic = ip.plugin_manager.get_plugin('parallelmagic')
             if pmagic is None:
                 ip.magic_load_ext('parallelmagic')
                 pmagic = ip.plugin_manager.get_plugin('parallelmagic')
 
             pmagic.active_view = self
 
 
 @skip_doctest
 class LoadBalancedView(View):
     """An load-balancing View that only executes via the Task scheduler.
 
     Load-balanced views can be created with the client's `view` method:
 
     >>> v = client.load_balanced_view()
 
     or targets can be specified, to restrict the potential destinations:
 
     >>> v = client.client.load_balanced_view([1,3])
 
     which would restrict loadbalancing to between engines 1 and 3.
 
     """
 
     follow=Any()
     after=Any()
     timeout=CFloat()
     retries = Integer(0)
 
     _task_scheme = Any()
     _flag_names = List(['targets', 'block', 'track', 'follow', 'after', 'timeout', 'retries'])
 
     def __init__(self, client=None, socket=None, **flags):
         super(LoadBalancedView, self).__init__(client=client, socket=socket, **flags)
         self._task_scheme=client._task_scheme
 
     def _validate_dependency(self, dep):
         """validate a dependency.
 
         For use in `set_flags`.
         """
         if dep is None or isinstance(dep, (basestring, AsyncResult, Dependency)):
             return True
         elif isinstance(dep, (list,set, tuple)):
             for d in dep:
                 if not isinstance(d, (basestring, AsyncResult)):
                     return False
         elif isinstance(dep, dict):
             if set(dep.keys()) != set(Dependency().as_dict().keys()):
                 return False
             if not isinstance(dep['msg_ids'], list):
                 return False
             for d in dep['msg_ids']:
                 if not isinstance(d, basestring):
                     return False
         else:
             return False
 
         return True
 
     def _render_dependency(self, dep):
         """helper for building jsonable dependencies from various input forms."""
         if isinstance(dep, Dependency):
             return dep.as_dict()
         elif isinstance(dep, AsyncResult):
             return dep.msg_ids
         elif dep is None:
             return []
         else:
             # pass to Dependency constructor
             return list(Dependency(dep))
 
     def set_flags(self, **kwargs):
         """set my attribute flags by keyword.
 
         A View is a wrapper for the Client's apply method, but with attributes
         that specify keyword arguments, those attributes can be set by keyword
         argument with this method.
 
         Parameters
         ----------
 
         block : bool
             whether to wait for results
         track : bool
             whether to create a MessageTracker to allow the user to
             safely edit after arrays and buffers during non-copying
             sends.
 
         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.
 
         retries : int
             Number of times a task will be retried on failure.
         """
 
         super(LoadBalancedView, self).set_flags(**kwargs)
         for name in ('follow', 'after'):
             if name in kwargs:
                 value = kwargs[name]
                 if self._validate_dependency(value):
                     setattr(self, name, value)
                 else:
                     raise ValueError("Invalid dependency: %r"%value)
         if 'timeout' in kwargs:
             t = kwargs['timeout']
             if not isinstance(t, (int, long, float, type(None))):
                 raise TypeError("Invalid type for timeout: %r"%type(t))
             if t is not None:
                 if t < 0:
                     raise ValueError("Invalid timeout: %s"%t)
             self.timeout = t
 
     @sync_results
     @save_ids
     def _really_apply(self, f, args=None, kwargs=None, block=None, track=None,
                                         after=None, follow=None, timeout=None,
                                         targets=None, retries=None):
         """calls f(*args, **kwargs) on a remote engine, returning the result.
 
         This method temporarily sets all of `apply`'s flags for a single call.
 
         Parameters
         ----------
 
         f : callable
 
         args : list [default: empty]
 
         kwargs : dict [default: empty]
 
         block : bool [default: self.block]
             whether to block
         track : bool [default: self.track]
             whether to ask zmq to track the message, for safe non-copying sends
 
         !!!!!! TODO: THE REST HERE  !!!!
 
         Returns
         -------
 
         if self.block is False:
             returns AsyncResult
         else:
             returns actual result of f(*args, **kwargs) on the engine(s)
             This will be a list of self.targets is also a list (even length 1), or
             the single result if self.targets is an integer engine id
         """
 
         # validate whether we can run
         if self._socket.closed:
             msg = "Task farming is disabled"
             if self._task_scheme == 'pure':
                 msg += " because the pure ZMQ scheduler cannot handle"
                 msg += " disappearing engines."
             raise RuntimeError(msg)
 
         if self._task_scheme == 'pure':
             # pure zmq scheme doesn't support extra features
             msg = "Pure ZMQ scheduler doesn't support the following flags:"
             "follow, after, retries, targets, timeout"
             if (follow or after or retries or targets or timeout):
                 # hard fail on Scheduler flags
                 raise RuntimeError(msg)
             if isinstance(f, dependent):
                 # soft warn on functional dependencies
                 warnings.warn(msg, RuntimeWarning)
 
         # build args
         args = [] if args is None else args
         kwargs = {} if kwargs is None else kwargs
         block = self.block if block is None else block
         track = self.track if track is None else track
         after = self.after if after is None else after
         retries = self.retries if retries is None else retries
         follow = self.follow if follow is None else follow
         timeout = self.timeout if timeout is None else timeout
         targets = self.targets if targets is None else targets
 
         if not isinstance(retries, int):
             raise TypeError('retries must be int, not %r'%type(retries))
 
         if targets is None:
             idents = []
         else:
             idents = self.client._build_targets(targets)[0]
             # ensure *not* bytes
             idents = [ ident.decode() for ident in idents ]
 
         after = self._render_dependency(after)
         follow = self._render_dependency(follow)
         subheader = dict(after=after, follow=follow, timeout=timeout, targets=idents, retries=retries)
 
         msg = self.client.send_apply_message(self._socket, f, args, kwargs, track=track,
                                 subheader=subheader)
         tracker = None if track is False else msg['tracker']
 
         ar = AsyncResult(self.client, msg['header']['msg_id'], fname=getname(f), targets=None, tracker=tracker)
 
         if block:
             try:
                 return ar.get()
             except KeyboardInterrupt:
                 pass
         return ar
 
     @spin_after
     @save_ids
     def map(self, f, *sequences, **kwargs):
         """view.map(f, *sequences, block=self.block, chunksize=1, ordered=True) => list|AsyncMapResult
 
         Parallel version of builtin `map`, load-balanced by this View.
 
         `block`, and `chunksize` can be specified by keyword only.
 
         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.
 
         Parameters
         ----------
 
         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
             whether to create a MessageTracker to allow the user to
             safely edit after arrays and buffers during non-copying
             sends.
         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.
 
         Returns
         -------
 
         if block=False:
             AsyncMapResult
                 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:
                 the result of map(f,*sequences)
 
         """
 
         # default
         block = kwargs.get('block', self.block)
         chunksize = kwargs.get('chunksize', 1)
         ordered = kwargs.get('ordered', True)
 
         keyset = set(kwargs.keys())
         extra_keys = keyset.difference_update(set(['block', 'chunksize']))
         if extra_keys:
             raise TypeError("Invalid kwargs: %s"%list(extra_keys))
 
         assert len(sequences) > 0, "must have some sequences to map onto!"
 
         pf = ParallelFunction(self, f, block=block, chunksize=chunksize, ordered=ordered)
         return pf.map(*sequences)
 
 __all__ = ['LoadBalancedView', 'DirectView']

IPython/parallel/util.py

@@ -1,476 +1,476 b''
 """some generic utilities for dealing with classes, urls, and serialization
 
 Authors:
 
 * Min RK
 """
 #-----------------------------------------------------------------------------
 #  Copyright (C) 2010-2011  The IPython Development Team
 #
 #  Distributed under the terms of the BSD License.  The full license is in
 #  the file COPYING, distributed as part of this software.
 #-----------------------------------------------------------------------------
 
 #-----------------------------------------------------------------------------
 # Imports
 #-----------------------------------------------------------------------------
 
 # Standard library imports.
 import logging
 import os
 import re
 import stat
 import socket
 import sys
 from signal import signal, SIGINT, SIGABRT, SIGTERM
 try:
     from signal import SIGKILL
 except ImportError:
     SIGKILL=None
 
 try:
     import cPickle
     pickle = cPickle
 except:
     cPickle = None
     import pickle
 
 # System library imports
 import zmq
 from zmq.log import handlers
 
 # IPython imports
 from IPython.config.application import Application
 from IPython.utils.pickleutil import can, uncan, canSequence, uncanSequence
 from IPython.utils.newserialized import serialize, unserialize
 from IPython.zmq.log import EnginePUBHandler
 
 #-----------------------------------------------------------------------------
 # Classes
 #-----------------------------------------------------------------------------
 
 class Namespace(dict):
     """Subclass of dict for attribute access to keys."""
     
     def __getattr__(self, key):
         """getattr aliased to getitem"""
         if key in self.iterkeys():
             return self[key]
         else:
             raise NameError(key)
 
     def __setattr__(self, key, value):
         """setattr aliased to setitem, with strict"""
         if hasattr(dict, key):
             raise KeyError("Cannot override dict keys %r"%key)
         self[key] = value
     
 
 class ReverseDict(dict):
     """simple double-keyed subset of dict methods."""
     
     def __init__(self, *args, **kwargs):
         dict.__init__(self, *args, **kwargs)
         self._reverse = dict()
         for key, value in self.iteritems():
             self._reverse[value] = key
     
     def __getitem__(self, key):
         try:
             return dict.__getitem__(self, key)
         except KeyError:
             return self._reverse[key]
     
     def __setitem__(self, key, value):
         if key in self._reverse:
             raise KeyError("Can't have key %r on both sides!"%key)
         dict.__setitem__(self, key, value)
         self._reverse[value] = key
     
     def pop(self, key):
         value = dict.pop(self, key)
         self._reverse.pop(value)
         return value
     
     def get(self, key, default=None):
         try:
             return self[key]
         except KeyError:
             return default
 
 #-----------------------------------------------------------------------------
 # Functions
 #-----------------------------------------------------------------------------
 
 def asbytes(s):
     """ensure that an object is ascii bytes"""
     if isinstance(s, unicode):
         s = s.encode('ascii')
     return s
 
 def is_url(url):
     """boolean check for whether a string is a zmq url"""
     if '://' not in url:
         return False
     proto, addr = url.split('://', 1)
     if proto.lower() not in ['tcp','pgm','epgm','ipc','inproc']:
         return False
     return True
 
 def validate_url(url):
     """validate a url for zeromq"""
     if not isinstance(url, basestring):
         raise TypeError("url must be a string, not %r"%type(url))
     url = url.lower()
     
     proto_addr = url.split('://')
     assert len(proto_addr) == 2, 'Invalid url: %r'%url
     proto, addr = proto_addr
     assert proto in ['tcp','pgm','epgm','ipc','inproc'], "Invalid protocol: %r"%proto
     
     # domain pattern adapted from http://www.regexlib.com/REDetails.aspx?regexp_id=391
     # author: Remi Sabourin
     pat = re.compile(r'^([\w\d]([\w\d\-]{0,61}[\w\d])?\.)*[\w\d]([\w\d\-]{0,61}[\w\d])?$')
     
     if proto == 'tcp':
         lis = addr.split(':')
         assert len(lis) == 2, 'Invalid url: %r'%url
         addr,s_port = lis
         try:
             port = int(s_port)
         except ValueError:
             raise AssertionError("Invalid port %r in url: %r"%(port, url))
         
         assert addr == '*' or pat.match(addr) is not None, 'Invalid url: %r'%url
         
     else:
         # only validate tcp urls currently
         pass
     
     return True
 
 
 def validate_url_container(container):
     """validate a potentially nested collection of urls."""
     if isinstance(container, basestring):
         url = container
         return validate_url(url)
     elif isinstance(container, dict):
         container = container.itervalues()
     
     for element in container:
         validate_url_container(element)
 
 
 def split_url(url):
     """split a zmq url (tcp://ip:port) into ('tcp','ip','port')."""
     proto_addr = url.split('://')
     assert len(proto_addr) == 2, 'Invalid url: %r'%url
     proto, addr = proto_addr
     lis = addr.split(':')
     assert len(lis) == 2, 'Invalid url: %r'%url
     addr,s_port = lis
     return proto,addr,s_port
     
 def disambiguate_ip_address(ip, location=None):
     """turn multi-ip interfaces '0.0.0.0' and '*' into connectable
     ones, based on the location (default interpretation of location is localhost)."""
     if ip in ('0.0.0.0', '*'):
         try:
             external_ips = socket.gethostbyname_ex(socket.gethostname())[2]
         except (socket.gaierror, IndexError):
             # couldn't identify this machine, assume localhost
             external_ips = []
         if location is None or location in external_ips or not external_ips:
             # If location is unspecified or cannot be determined, assume local
             ip='127.0.0.1'
         elif location:
             return location
     return ip
 
 def disambiguate_url(url, location=None):
     """turn multi-ip interfaces '0.0.0.0' and '*' into connectable
     ones, based on the location (default interpretation is localhost).
     
     This is for zeromq urls, such as tcp://*:10101."""
     try:
         proto,ip,port = split_url(url)
     except AssertionError:
         # probably not tcp url; could be ipc, etc.
         return url
     
     ip = disambiguate_ip_address(ip,location)
     
     return "%s://%s:%s"%(proto,ip,port)
 
 def serialize_object(obj, threshold=64e-6):
     """Serialize an object into a list of sendable buffers.
     
     Parameters
     ----------
     
     obj : object
         The object to be serialized
     threshold : float
         The threshold for not double-pickling the content.
         
     
     Returns
     -------
     ('pmd', [bufs]) :
         where pmd is the pickled metadata wrapper,
         bufs is a list of data buffers
     """
     databuffers = []
     if isinstance(obj, (list, tuple)):
         clist = canSequence(obj)
         slist = map(serialize, clist)
         for s in slist:
             if s.typeDescriptor in ('buffer', 'ndarray') or s.getDataSize() > threshold:
                 databuffers.append(s.getData())
                 s.data = None
         return pickle.dumps(slist,-1), databuffers
     elif isinstance(obj, dict):
         sobj = {}
         for k in sorted(obj.iterkeys()):
             s = serialize(can(obj[k]))
             if s.typeDescriptor in ('buffer', 'ndarray') or s.getDataSize() > threshold:
                 databuffers.append(s.getData())
                 s.data = None
             sobj[k] = s
         return pickle.dumps(sobj,-1),databuffers
     else:
         s = serialize(can(obj))
         if s.typeDescriptor in ('buffer', 'ndarray') or s.getDataSize() > threshold:
             databuffers.append(s.getData())
             s.data = None
         return pickle.dumps(s,-1),databuffers
             
         
 def unserialize_object(bufs):
     """reconstruct an object serialized by serialize_object from data buffers."""
     bufs = list(bufs)
     sobj = pickle.loads(bufs.pop(0))
     if isinstance(sobj, (list, tuple)):
         for s in sobj:
             if s.data is None:
                 s.data = bufs.pop(0)
         return uncanSequence(map(unserialize, sobj)), bufs
     elif isinstance(sobj, dict):
         newobj = {}
         for k in sorted(sobj.iterkeys()):
             s = sobj[k]
             if s.data is None:
                 s.data = bufs.pop(0)
             newobj[k] = uncan(unserialize(s))
         return newobj, bufs
     else:
         if sobj.data is None:
             sobj.data = bufs.pop(0)
         return uncan(unserialize(sobj)), bufs
 
 def pack_apply_message(f, args, kwargs, threshold=64e-6):
     """pack up a function, args, and kwargs to be sent over the wire
     as a series of buffers. Any object whose data is larger than `threshold`
     will not have their data copied (currently only numpy arrays support zero-copy)"""
     msg = [pickle.dumps(can(f),-1)]
     databuffers = [] # for large objects
     sargs, bufs = serialize_object(args,threshold)
     msg.append(sargs)
     databuffers.extend(bufs)
     skwargs, bufs = serialize_object(kwargs,threshold)
     msg.append(skwargs)
     databuffers.extend(bufs)
     msg.extend(databuffers)
     return msg
 
 def unpack_apply_message(bufs, g=None, copy=True):
     """unpack f,args,kwargs from buffers packed by pack_apply_message()
     Returns: original f,args,kwargs"""
     bufs = list(bufs) # allow us to pop
     assert len(bufs) >= 3, "not enough buffers!"
     if not copy:
         for i in range(3):
             bufs[i] = bufs[i].bytes
     cf = pickle.loads(bufs.pop(0))
     sargs = list(pickle.loads(bufs.pop(0)))
     skwargs = dict(pickle.loads(bufs.pop(0)))
     # print sargs, skwargs
     f = uncan(cf, g)
     for sa in sargs:
         if sa.data is None:
             m = bufs.pop(0)
             if sa.getTypeDescriptor() in ('buffer', 'ndarray'):
                 # always use a buffer, until memoryviews get sorted out
                 sa.data = buffer(m)
                 # disable memoryview support
                 # if copy:
                 #     sa.data = buffer(m)
                 # else:
                 #     sa.data = m.buffer
             else:
                 if copy:
                     sa.data = m
                 else:
                     sa.data = m.bytes
     
     args = uncanSequence(map(unserialize, sargs), g)
     kwargs = {}
     for k in sorted(skwargs.iterkeys()):
         sa = skwargs[k]
         if sa.data is None:
             m = bufs.pop(0)
             if sa.getTypeDescriptor() in ('buffer', 'ndarray'):
                 # always use a buffer, until memoryviews get sorted out
                 sa.data = buffer(m)
                 # disable memoryview support
                 # if copy:
                 #     sa.data = buffer(m)
                 # else:
                 #     sa.data = m.buffer
             else:
                 if copy:
                     sa.data = m
                 else:
                     sa.data = m.bytes
 
         kwargs[k] = uncan(unserialize(sa), g)
     
     return f,args,kwargs
 
 #--------------------------------------------------------------------------
 # helpers for implementing old MEC API via view.apply
 #--------------------------------------------------------------------------
 
 def interactive(f):
     """decorator for making functions appear as interactively defined.
     This results in the function being linked to the user_ns as globals()
     instead of the module globals().
     """
     f.__module__ = '__main__'
     return f
 
 @interactive
-def _push(ns):
+def _push(**ns):
     """helper method for implementing `client.push` via `client.apply`"""
     globals().update(ns)
 
 @interactive
 def _pull(keys):
     """helper method for implementing `client.pull` via `client.apply`"""
     user_ns = globals()
     if isinstance(keys, (list,tuple, set)):
         for key in keys:
             if not user_ns.has_key(key):
                 raise NameError("name '%s' is not defined"%key)
         return map(user_ns.get, keys)
     else:
         if not user_ns.has_key(keys):
             raise NameError("name '%s' is not defined"%keys)
         return user_ns.get(keys)
 
 @interactive
 def _execute(code):
     """helper method for implementing `client.execute` via `client.apply`"""
     exec code in globals()
 
 #--------------------------------------------------------------------------
 # extra process management utilities
 #--------------------------------------------------------------------------
 
 _random_ports = set()
 
 def select_random_ports(n):
     """Selects and return n random ports that are available."""
     ports = []
     for i in xrange(n):
         sock = socket.socket()
         sock.bind(('', 0))
         while sock.getsockname()[1] in _random_ports:
             sock.close()
             sock = socket.socket()
             sock.bind(('', 0))
         ports.append(sock)
     for i, sock in enumerate(ports):
         port = sock.getsockname()[1]
         sock.close()
         ports[i] = port
         _random_ports.add(port)
     return ports
 
 def signal_children(children):
     """Relay interupt/term signals to children, for more solid process cleanup."""
     def terminate_children(sig, frame):
         log = Application.instance().log
         log.critical("Got signal %i, terminating children..."%sig)
         for child in children:
             child.terminate()
         
         sys.exit(sig != SIGINT)
         # sys.exit(sig)
     for sig in (SIGINT, SIGABRT, SIGTERM):
         signal(sig, terminate_children)
 
 def generate_exec_key(keyfile):
     import uuid
     newkey = str(uuid.uuid4())
     with open(keyfile, 'w') as f:
         # f.write('ipython-key ')
         f.write(newkey+'\n')
     # set user-only RW permissions (0600)
     # this will have no effect on Windows
     os.chmod(keyfile, stat.S_IRUSR|stat.S_IWUSR)
 
 
 def integer_loglevel(loglevel):
     try:
         loglevel = int(loglevel)
     except ValueError:
         if isinstance(loglevel, str):
             loglevel = getattr(logging, loglevel)
     return loglevel
 
 def connect_logger(logname, context, iface, root="ip", loglevel=logging.DEBUG):
     logger = logging.getLogger(logname)
     if any([isinstance(h, handlers.PUBHandler) for h in logger.handlers]):
         # don't add a second PUBHandler
         return
     loglevel = integer_loglevel(loglevel)
     lsock = context.socket(zmq.PUB)
     lsock.connect(iface)
     handler = handlers.PUBHandler(lsock)
     handler.setLevel(loglevel)
     handler.root_topic = root
     logger.addHandler(handler)
     logger.setLevel(loglevel)
 
 def connect_engine_logger(context, iface, engine, loglevel=logging.DEBUG):
     logger = logging.getLogger()
     if any([isinstance(h, handlers.PUBHandler) for h in logger.handlers]):
         # don't add a second PUBHandler
         return
     loglevel = integer_loglevel(loglevel)
     lsock = context.socket(zmq.PUB)
     lsock.connect(iface)
     handler = EnginePUBHandler(engine, lsock)
     handler.setLevel(loglevel)
     logger.addHandler(handler)
     logger.setLevel(loglevel)
     return logger
 
 def local_logger(logname, loglevel=logging.DEBUG):
     loglevel = integer_loglevel(loglevel)
     logger = logging.getLogger(logname)
     if any([isinstance(h, logging.StreamHandler) for h in logger.handlers]):
         # don't add a second StreamHandler
         return
     handler = logging.StreamHandler()
     handler.setLevel(loglevel)
     formatter = logging.Formatter("%(asctime)s.%(msecs).03d [%(name)s] %(message)s",
                 datefmt="%Y-%m-%d %H:%M:%S")
     handler.setFormatter(formatter)
 
     logger.addHandler(handler)
     logger.setLevel(loglevel)
     return logger