allow true single-threaded Controller

Previously --usethreads only put MonitoredQueue devices in threads, and still launched a subprocess for the Python scheduler.  This adds support for the Python Scheduler in a thread.

IPython/parallel/apps/ipcontrollerapp.py

@@ -1,408 +1,414 b''
 #!/usr/bin/env python
 # encoding: utf-8
 """
 The IPython controller application.
 
 Authors:
 
 * Brian Granger
 * MinRK
 
 """
 
 #-----------------------------------------------------------------------------
 #  Copyright (C) 2008-2011  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
 #-----------------------------------------------------------------------------
 
 from __future__ import with_statement
 
 import os
 import socket
 import stat
 import sys
 import uuid
 
 from multiprocessing import Process
 
 import zmq
 from zmq.devices import ProcessMonitoredQueue
 from zmq.log.handlers import PUBHandler
 from zmq.utils import jsonapi as json
 
 from IPython.config.application import boolean_flag
 from IPython.core.profiledir import ProfileDir
 
 from IPython.parallel.apps.baseapp import (
     BaseParallelApplication,
     base_flags
 )
 from IPython.utils.importstring import import_item
 from IPython.utils.traitlets import Instance, Unicode, Bool, List, Dict
 
 # from IPython.parallel.controller.controller import ControllerFactory
 from IPython.zmq.session import Session
 from IPython.parallel.controller.heartmonitor import HeartMonitor
 from IPython.parallel.controller.hub import HubFactory
 from IPython.parallel.controller.scheduler import TaskScheduler,launch_scheduler
 from IPython.parallel.controller.sqlitedb import SQLiteDB
 
 from IPython.parallel.util import signal_children, split_url
 
 # conditional import of MongoDB backend class
 
 try:
     from IPython.parallel.controller.mongodb import MongoDB
 except ImportError:
     maybe_mongo = []
 else:
     maybe_mongo = [MongoDB]
 
 
 #-----------------------------------------------------------------------------
 # Module level variables
 #-----------------------------------------------------------------------------
 
 
 #: The default config file name for this application
 default_config_file_name = u'ipcontroller_config.py'
 
 
 _description = """Start the IPython controller for parallel computing.
 
 The IPython controller provides a gateway between the IPython engines and
 clients. The controller needs to be started before the engines and can be
 configured using command line options or using a cluster directory. Cluster
 directories contain config, log and security files and are usually located in
 your ipython directory and named as "profile_name". See the `profile`
 and `profile_dir` options for details.
 """
 
 
 
 
 #-----------------------------------------------------------------------------
 # The main application
 #-----------------------------------------------------------------------------
 flags = {}
 flags.update(base_flags)
 flags.update({
     'usethreads' : ( {'IPControllerApp' : {'use_threads' : True}},
                     'Use threads instead of processes for the schedulers'),
     'sqlitedb' : ({'HubFactory' : {'db_class' : 'IPython.parallel.controller.sqlitedb.SQLiteDB'}},
                     'use the SQLiteDB backend'),
     'mongodb' : ({'HubFactory' : {'db_class' : 'IPython.parallel.controller.mongodb.MongoDB'}},
                     'use the MongoDB backend'),
     'dictdb' : ({'HubFactory' : {'db_class' : 'IPython.parallel.controller.dictdb.DictDB'}},
                     'use the in-memory DictDB backend'),
     'reuse' : ({'IPControllerApp' : {'reuse_files' : True}},
                     'reuse existing json connection files')
 })
 
 flags.update(boolean_flag('secure', 'IPControllerApp.secure',
     "Use HMAC digests for authentication of messages.",
     "Don't authenticate messages."
 ))
 
 class IPControllerApp(BaseParallelApplication):
 
     name = u'ipcontroller'
     description = _description
     config_file_name = Unicode(default_config_file_name)
     classes = [ProfileDir, Session, HubFactory, TaskScheduler, HeartMonitor, SQLiteDB] + maybe_mongo
     
     # change default to True
     auto_create = Bool(True, config=True,
         help="""Whether to create profile dir if it doesn't exist.""")
     
     reuse_files = Bool(False, config=True,
         help='Whether to reuse existing json connection files.'
     )
     secure = Bool(True, config=True,
         help='Whether to use HMAC digests for extra message authentication.'
     )
     ssh_server = Unicode(u'', config=True,
         help="""ssh url for clients to use when connecting to the Controller
         processes. It should be of the form: [user@]server[:port]. The
         Controller's listening addresses must be accessible from the ssh server""",
     )
     location = Unicode(u'', config=True,
         help="""The external IP or domain name of the Controller, used for disambiguating
         engine and client connections.""",
     )
     import_statements = List([], config=True,
         help="import statements to be run at startup.  Necessary in some environments"
     )
 
     use_threads = Bool(False, config=True,
         help='Use threads instead of processes for the schedulers',
         )
 
     # internal
     children = List()
     mq_class = Unicode('zmq.devices.ProcessMonitoredQueue')
 
     def _use_threads_changed(self, name, old, new):
         self.mq_class = 'zmq.devices.%sMonitoredQueue'%('Thread' if new else 'Process')
 
     aliases = Dict(dict(
         log_level = 'IPControllerApp.log_level',
         log_url = 'IPControllerApp.log_url',
         reuse_files = 'IPControllerApp.reuse_files',
         secure = 'IPControllerApp.secure',
         ssh = 'IPControllerApp.ssh_server',
         use_threads = 'IPControllerApp.use_threads',
         import_statements = 'IPControllerApp.import_statements',
         location = 'IPControllerApp.location',
 
         ident = 'Session.session',
         user = 'Session.username',
         exec_key = 'Session.keyfile',
 
         url = 'HubFactory.url',
         ip = 'HubFactory.ip',
         transport = 'HubFactory.transport',
         port = 'HubFactory.regport',
 
         ping = 'HeartMonitor.period',
 
         scheme = 'TaskScheduler.scheme_name',
         hwm = 'TaskScheduler.hwm',
 
 
         profile = "BaseIPythonApplication.profile",
         profile_dir = 'ProfileDir.location',
         
     ))
     flags = Dict(flags)
     
 
     def save_connection_dict(self, fname, cdict):
         """save a connection dict to json file."""
         c = self.config
         url = cdict['url']
         location = cdict['location']
         if not location:
             try:
                 proto,ip,port = split_url(url)
             except AssertionError:
                 pass
             else:
                 location = socket.gethostbyname_ex(socket.gethostname())[2][-1]
             cdict['location'] = location
         fname = os.path.join(self.profile_dir.security_dir, fname)
         with open(fname, 'w') as f:
             f.write(json.dumps(cdict, indent=2))
         os.chmod(fname, stat.S_IRUSR|stat.S_IWUSR)
     
     def load_config_from_json(self):
         """load config from existing json connector files."""
         c = self.config
         # load from engine config
         with open(os.path.join(self.profile_dir.security_dir, 'ipcontroller-engine.json')) as f:
             cfg = json.loads(f.read())
         key = c.Session.key = cfg['exec_key']
         xport,addr = cfg['url'].split('://')
         c.HubFactory.engine_transport = xport
         ip,ports = addr.split(':')
         c.HubFactory.engine_ip = ip
         c.HubFactory.regport = int(ports)
         self.location = cfg['location']
         
         # load client config
         with open(os.path.join(self.profile_dir.security_dir, 'ipcontroller-client.json')) as f:
             cfg = json.loads(f.read())
         assert key == cfg['exec_key'], "exec_key mismatch between engine and client keys"
         xport,addr = cfg['url'].split('://')
         c.HubFactory.client_transport = xport
         ip,ports = addr.split(':')
         c.HubFactory.client_ip = ip
         self.ssh_server = cfg['ssh']
         assert int(ports) == c.HubFactory.regport, "regport mismatch"
     
     def init_hub(self):
         c = self.config
         
         self.do_import_statements()
         reusing = self.reuse_files
         if reusing:
             try:
                 self.load_config_from_json()
             except (AssertionError,IOError):
                 reusing=False
         # check again, because reusing may have failed:
         if reusing:
             pass
         elif self.secure:
             key = str(uuid.uuid4())
             # keyfile = os.path.join(self.profile_dir.security_dir, self.exec_key)
             # with open(keyfile, 'w') as f:
             #     f.write(key)
             # os.chmod(keyfile, stat.S_IRUSR|stat.S_IWUSR)
             c.Session.key = key
         else:
             key = c.Session.key = ''
         
         try:
             self.factory = HubFactory(config=c, log=self.log)
             # self.start_logging()
             self.factory.init_hub()
         except:
             self.log.error("Couldn't construct the Controller", exc_info=True)
             self.exit(1)
         
         if not reusing:
             # save to new json config files
             f = self.factory
             cdict = {'exec_key' : key,
                     'ssh' : self.ssh_server,
                     'url' : "%s://%s:%s"%(f.client_transport, f.client_ip, f.regport),
                     'location' : self.location
                     }
             self.save_connection_dict('ipcontroller-client.json', cdict)
             edict = cdict
             edict['url']="%s://%s:%s"%((f.client_transport, f.client_ip, f.regport))
             self.save_connection_dict('ipcontroller-engine.json', edict)
 
     #
     def init_schedulers(self):
         children = self.children
         mq = import_item(str(self.mq_class))
         
         hub = self.factory
         # maybe_inproc = 'inproc://monitor' if self.use_threads else self.monitor_url
         # IOPub relay (in a Process)
         q = mq(zmq.PUB, zmq.SUB, zmq.PUB, 'N/A','iopub')
         q.bind_in(hub.client_info['iopub'])
         q.bind_out(hub.engine_info['iopub'])
         q.setsockopt_out(zmq.SUBSCRIBE, '')
         q.connect_mon(hub.monitor_url)
         q.daemon=True
         children.append(q)
 
         # Multiplexer Queue (in a Process)
         q = mq(zmq.XREP, zmq.XREP, zmq.PUB, 'in', 'out')
         q.bind_in(hub.client_info['mux'])
         q.setsockopt_in(zmq.IDENTITY, 'mux')
         q.bind_out(hub.engine_info['mux'])
         q.connect_mon(hub.monitor_url)
         q.daemon=True
         children.append(q)
 
         # Control Queue (in a Process)
         q = mq(zmq.XREP, zmq.XREP, zmq.PUB, 'incontrol', 'outcontrol')
         q.bind_in(hub.client_info['control'])
         q.setsockopt_in(zmq.IDENTITY, 'control')
         q.bind_out(hub.engine_info['control'])
         q.connect_mon(hub.monitor_url)
         q.daemon=True
         children.append(q)
         try:
             scheme = self.config.TaskScheduler.scheme_name
         except AttributeError:
             scheme = TaskScheduler.scheme_name.get_default_value()
         # Task Queue (in a Process)
         if scheme == 'pure':
             self.log.warn("task::using pure XREQ Task scheduler")
             q = mq(zmq.XREP, zmq.XREQ, zmq.PUB, 'intask', 'outtask')
             # q.setsockopt_out(zmq.HWM, hub.hwm)
             q.bind_in(hub.client_info['task'][1])
             q.setsockopt_in(zmq.IDENTITY, 'task')
             q.bind_out(hub.engine_info['task'])
             q.connect_mon(hub.monitor_url)
             q.daemon=True
             children.append(q)
         elif scheme == 'none':
             self.log.warn("task::using no Task scheduler")
 
         else:
             self.log.info("task::using Python %s Task scheduler"%scheme)
             sargs = (hub.client_info['task'][1], hub.engine_info['task'],
                                 hub.monitor_url, hub.client_info['notification'])
             kwargs = dict(logname='scheduler', loglevel=self.log_level,
                             log_url = self.log_url, config=dict(self.config))
-            q = Process(target=launch_scheduler, args=sargs, kwargs=kwargs)
-            q.daemon=True
-            children.append(q)
+            if 'Process' in self.mq_class:
+                # run the Python scheduler in a Process
+                q = Process(target=launch_scheduler, args=sargs, kwargs=kwargs)
+                q.daemon=True
+                children.append(q)
+            else:
+                # single-threaded Controller
+                kwargs['in_thread'] = True
+                launch_scheduler(*sargs, **kwargs)
 
     
     def save_urls(self):
         """save the registration urls to files."""
         c = self.config
         
         sec_dir = self.profile_dir.security_dir
         cf = self.factory
         
         with open(os.path.join(sec_dir, 'ipcontroller-engine.url'), 'w') as f:
             f.write("%s://%s:%s"%(cf.engine_transport, cf.engine_ip, cf.regport))
         
         with open(os.path.join(sec_dir, 'ipcontroller-client.url'), 'w') as f:
             f.write("%s://%s:%s"%(cf.client_transport, cf.client_ip, cf.regport))
         
     
     def do_import_statements(self):
         statements = self.import_statements
         for s in statements:
             try:
                 self.log.msg("Executing statement: '%s'" % s)
                 exec s in globals(), locals()
             except:
                 self.log.msg("Error running statement: %s" % s)
 
     def forward_logging(self):
         if self.log_url:
             self.log.info("Forwarding logging to %s"%self.log_url)
             context = zmq.Context.instance()
             lsock = context.socket(zmq.PUB)
             lsock.connect(self.log_url)
             handler = PUBHandler(lsock)
             self.log.removeHandler(self._log_handler)
             handler.root_topic = 'controller'
             handler.setLevel(self.log_level)
             self.log.addHandler(handler)
             self._log_handler = handler
     # #
     
     def initialize(self, argv=None):
         super(IPControllerApp, self).initialize(argv)
         self.forward_logging()
         self.init_hub()
         self.init_schedulers()
     
     def start(self):
         # Start the subprocesses:
         self.factory.start()
         child_procs = []
         for child in self.children:
             child.start()
             if isinstance(child, ProcessMonitoredQueue):
                 child_procs.append(child.launcher)
             elif isinstance(child, Process):
                 child_procs.append(child)
         if child_procs:
             signal_children(child_procs)
 
         self.write_pid_file(overwrite=True)
 
         try:
             self.factory.loop.start()
         except KeyboardInterrupt:
             self.log.critical("Interrupted, Exiting...\n")
             
 
 
 def launch_new_instance():
     """Create and run the IPython controller"""
     app = IPControllerApp.instance()
     app.initialize()
     app.start()
 
 
 if __name__ == '__main__':
     launch_new_instance()

IPython/parallel/controller/scheduler.py

@@ -1,692 +1,703 b''
 """The Python scheduler for rich scheduling.
 
 The Pure ZMQ scheduler does not allow routing schemes other than LRU,
 nor does it check msg_id DAG dependencies. For those, a slightly slower
 Python Scheduler exists.
 
 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
 #----------------------------------------------------------------------
 
 from __future__ import print_function
 
 import logging
 import sys
 
 from datetime import datetime, timedelta
 from random import randint, random
 from types import FunctionType
 
 try:
     import numpy
 except ImportError:
     numpy = None
 
 import zmq
 from zmq.eventloop import ioloop, zmqstream
 
 # local imports
 from IPython.external.decorator import decorator
+from IPython.config.application import Application
 from IPython.config.loader import Config
 from IPython.utils.traitlets import Instance, Dict, List, Set, Int, Enum
 
 from IPython.parallel import error
 from IPython.parallel.factory import SessionFactory
 from IPython.parallel.util import connect_logger, local_logger
 
 from .dependency import Dependency
 
 @decorator
 def logged(f,self,*args,**kwargs):
     # print ("#--------------------")
     self.log.debug("scheduler::%s(*%s,**%s)"%(f.func_name, args, kwargs))
     # print ("#--")
     return f(self,*args, **kwargs)
 
 #----------------------------------------------------------------------
 # Chooser functions
 #----------------------------------------------------------------------
 
 def plainrandom(loads):
     """Plain random pick."""
     n = len(loads)
     return randint(0,n-1)
 
 def lru(loads):
     """Always pick the front of the line.
     
     The content of `loads` is ignored.
     
     Assumes LRU ordering of loads, with oldest first.
     """
     return 0
 
 def twobin(loads):
     """Pick two at random, use the LRU of the two.
     
     The content of loads is ignored.
     
     Assumes LRU ordering of loads, with oldest first.
     """
     n = len(loads)
     a = randint(0,n-1)
     b = randint(0,n-1)
     return min(a,b)
 
 def weighted(loads):
     """Pick two at random using inverse load as weight.
     
     Return the less loaded of the two.
     """
     # weight 0 a million times more than 1:
     weights = 1./(1e-6+numpy.array(loads))
     sums = weights.cumsum()
     t = sums[-1]
     x = random()*t
     y = random()*t
     idx = 0
     idy = 0
     while sums[idx] < x:
         idx += 1
     while sums[idy] < y:
         idy += 1
     if weights[idy] > weights[idx]:
         return idy
     else:
         return idx
 
 def leastload(loads):
     """Always choose the lowest load.
     
     If the lowest load occurs more than once, the first
     occurance will be used.  If loads has LRU ordering, this means
     the LRU of those with the lowest load is chosen.
     """
     return loads.index(min(loads))
 
 #---------------------------------------------------------------------
 # Classes
 #---------------------------------------------------------------------
 # store empty default dependency:
 MET = Dependency([])
 
 class TaskScheduler(SessionFactory):
     """Python TaskScheduler object.
     
     This is the simplest object that supports msg_id based
     DAG dependencies. *Only* task msg_ids are checked, not
     msg_ids of jobs submitted via the MUX queue.
     
     """
     
     hwm = Int(0, config=True, shortname='hwm',
         help="""specify the High Water Mark (HWM) for the downstream
         socket in the Task scheduler. This is the maximum number
         of allowed outstanding tasks on each engine."""
     )
     scheme_name = Enum(('leastload', 'pure', 'lru', 'plainrandom', 'weighted', 'twobin'),
         'leastload', config=True, shortname='scheme', allow_none=False,
         help="""select the task scheduler scheme  [default: Python LRU]
         Options are: 'pure', 'lru', 'plainrandom', 'weighted', 'twobin','leastload'"""
     )
     def _scheme_name_changed(self, old, new):
         self.log.debug("Using scheme %r"%new)
         self.scheme = globals()[new]
     
     # input arguments:
     scheme = Instance(FunctionType) # function for determining the destination
     def _scheme_default(self):
         return leastload
     client_stream = Instance(zmqstream.ZMQStream) # client-facing stream
     engine_stream = Instance(zmqstream.ZMQStream) # engine-facing stream
     notifier_stream = Instance(zmqstream.ZMQStream) # hub-facing sub stream
     mon_stream = Instance(zmqstream.ZMQStream) # hub-facing pub stream
     
     # internals:
     graph = Dict() # dict by msg_id of [ msg_ids that depend on key ]
     retries = Dict() # dict by msg_id of retries remaining (non-neg ints)
     # waiting = List() # list of msg_ids ready to run, but haven't due to HWM
     depending = Dict() # dict by msg_id of (msg_id, raw_msg, after, follow)
     pending = Dict() # dict by engine_uuid of submitted tasks
     completed = Dict() # dict by engine_uuid of completed tasks
     failed = Dict() # dict by engine_uuid of failed tasks
     destinations = Dict() # dict by msg_id of engine_uuids where jobs ran (reverse of completed+failed)
     clients = Dict() # dict by msg_id for who submitted the task
     targets = List() # list of target IDENTs
     loads = List() # list of engine loads
     # full = Set() # set of IDENTs that have HWM outstanding tasks
     all_completed = Set() # set of all completed tasks
     all_failed = Set() # set of all failed tasks
     all_done = Set() # set of all finished tasks=union(completed,failed)
     all_ids = Set() # set of all submitted task IDs
     blacklist = Dict() # dict by msg_id of locations where a job has encountered UnmetDependency
     auditor = Instance('zmq.eventloop.ioloop.PeriodicCallback')
     
     
     def start(self):
         self.engine_stream.on_recv(self.dispatch_result, copy=False)
         self._notification_handlers = dict(
             registration_notification = self._register_engine,
             unregistration_notification = self._unregister_engine
         )
         self.notifier_stream.on_recv(self.dispatch_notification)
         self.auditor = ioloop.PeriodicCallback(self.audit_timeouts, 2e3, self.loop) # 1 Hz
         self.auditor.start()
         self.log.info("Scheduler started [%s]"%self.scheme_name)
     
     def resume_receiving(self):
         """Resume accepting jobs."""
         self.client_stream.on_recv(self.dispatch_submission, copy=False)
     
     def stop_receiving(self):
         """Stop accepting jobs while there are no engines.
         Leave them in the ZMQ queue."""
         self.client_stream.on_recv(None)
     
     #-----------------------------------------------------------------------
     # [Un]Registration Handling
     #-----------------------------------------------------------------------
     
     def dispatch_notification(self, msg):
         """dispatch register/unregister events."""
         try:
             idents,msg = self.session.feed_identities(msg)
         except ValueError:
             self.log.warn("task::Invalid Message: %r"%msg)
             return
         try:
             msg = self.session.unpack_message(msg)
         except ValueError:
             self.log.warn("task::Unauthorized message from: %r"%idents)
             return
         
         msg_type = msg['msg_type']
         
         handler = self._notification_handlers.get(msg_type, None)
         if handler is None:
             self.log.error("Unhandled message type: %r"%msg_type)
         else:
             try:
                 handler(str(msg['content']['queue']))
             except KeyError:
                 self.log.error("task::Invalid notification msg: %r"%msg)
     
     @logged
     def _register_engine(self, uid):
         """New engine with ident `uid` became available."""
         # head of the line:
         self.targets.insert(0,uid)
         self.loads.insert(0,0)
         # initialize sets
         self.completed[uid] = set()
         self.failed[uid] = set()
         self.pending[uid] = {}
         if len(self.targets) == 1:
             self.resume_receiving()
         # rescan the graph:
         self.update_graph(None)
 
     def _unregister_engine(self, uid):
         """Existing engine with ident `uid` became unavailable."""
         if len(self.targets) == 1:
             # this was our only engine
             self.stop_receiving()
         
         # handle any potentially finished tasks:
         self.engine_stream.flush()
         
         # don't pop destinations, because they might be used later
         # map(self.destinations.pop, self.completed.pop(uid))
         # map(self.destinations.pop, self.failed.pop(uid))
 
         # prevent this engine from receiving work
         idx = self.targets.index(uid)
         self.targets.pop(idx)
         self.loads.pop(idx)
         
         # wait 5 seconds before cleaning up pending jobs, since the results might
         # still be incoming
         if self.pending[uid]:
             dc = ioloop.DelayedCallback(lambda : self.handle_stranded_tasks(uid), 5000, self.loop)
             dc.start()
         else:
             self.completed.pop(uid)
             self.failed.pop(uid)
 
     
     @logged
     def handle_stranded_tasks(self, engine):
         """Deal with jobs resident in an engine that died."""
         lost = self.pending[engine]
         for msg_id in lost.keys():
             if msg_id not in self.pending[engine]:
                 # prevent double-handling of messages
                 continue
 
             raw_msg = lost[msg_id][0]
             idents,msg = self.session.feed_identities(raw_msg, copy=False)
             parent = self.session.unpack(msg[1].bytes)
             idents = [engine, idents[0]]
 
             # build fake error reply
             try:
                 raise error.EngineError("Engine %r died while running task %r"%(engine, msg_id))
             except:
                 content = error.wrap_exception()
             msg = self.session.msg('apply_reply', content, parent=parent, subheader={'status':'error'})
             raw_reply = map(zmq.Message, self.session.serialize(msg, ident=idents))
             # and dispatch it
             self.dispatch_result(raw_reply)
 
         # finally scrub completed/failed lists
         self.completed.pop(engine)
         self.failed.pop(engine)
     
     
     #-----------------------------------------------------------------------
     # Job Submission
     #-----------------------------------------------------------------------
     @logged
     def dispatch_submission(self, raw_msg):
         """Dispatch job submission to appropriate handlers."""
         # ensure targets up to date:
         self.notifier_stream.flush()
         try:
             idents, msg = self.session.feed_identities(raw_msg, copy=False)
             msg = self.session.unpack_message(msg, content=False, copy=False)
         except Exception:
             self.log.error("task::Invaid task msg: %r"%raw_msg, exc_info=True)
             return
         
         
         # send to monitor
         self.mon_stream.send_multipart(['intask']+raw_msg, copy=False)
         
         header = msg['header']
         msg_id = header['msg_id']
         self.all_ids.add(msg_id)
         
         # targets
         targets = set(header.get('targets', []))
         retries = header.get('retries', 0)
         self.retries[msg_id] = retries
         
         # time dependencies
         after = Dependency(header.get('after', []))
         if after.all:
             if after.success:
                 after.difference_update(self.all_completed)
             if after.failure:
                 after.difference_update(self.all_failed)
         if after.check(self.all_completed, self.all_failed):
             # recast as empty set, if `after` already met,
             # to prevent unnecessary set comparisons
             after = MET
         
         # location dependencies
         follow = Dependency(header.get('follow', []))
         
         # turn timeouts into datetime objects:
         timeout = header.get('timeout', None)
         if timeout:
             timeout = datetime.now() + timedelta(0,timeout,0)
         
         args = [raw_msg, targets, after, follow, timeout]
         
         # validate and reduce dependencies:
         for dep in after,follow:
             # check valid:
             if msg_id in dep or dep.difference(self.all_ids):
                 self.depending[msg_id] = args
                 return self.fail_unreachable(msg_id, error.InvalidDependency)
             # check if unreachable:
             if dep.unreachable(self.all_completed, self.all_failed):
                 self.depending[msg_id] = args
                 return self.fail_unreachable(msg_id)
         
         if after.check(self.all_completed, self.all_failed):
             # time deps already met, try to run
             if not self.maybe_run(msg_id, *args):
                 # can't run yet
                 if msg_id not in self.all_failed:
                     # could have failed as unreachable
                     self.save_unmet(msg_id, *args)
         else:
             self.save_unmet(msg_id, *args)
     
     # @logged
     def audit_timeouts(self):
         """Audit all waiting tasks for expired timeouts."""
         now = datetime.now()
         for msg_id in self.depending.keys():
             # must recheck, in case one failure cascaded to another:
             if msg_id in self.depending:
                 raw,after,targets,follow,timeout = self.depending[msg_id]
                 if timeout and timeout < now:
                     self.fail_unreachable(msg_id, error.TaskTimeout)
                 
     @logged
     def fail_unreachable(self, msg_id, why=error.ImpossibleDependency):
         """a task has become unreachable, send a reply with an ImpossibleDependency
         error."""
         if msg_id not in self.depending:
             self.log.error("msg %r already failed!"%msg_id)
             return
         raw_msg,targets,after,follow,timeout = self.depending.pop(msg_id)
         for mid in follow.union(after):
             if mid in self.graph:
                 self.graph[mid].remove(msg_id)
         
         # FIXME: unpacking a message I've already unpacked, but didn't save:
         idents,msg = self.session.feed_identities(raw_msg, copy=False)
         header = self.session.unpack(msg[1].bytes)
         
         try:
             raise why()
         except:
             content = error.wrap_exception()
         
         self.all_done.add(msg_id)
         self.all_failed.add(msg_id)
         
         msg = self.session.send(self.client_stream, 'apply_reply', content, 
                                                 parent=header, ident=idents)
         self.session.send(self.mon_stream, msg, ident=['outtask']+idents)
         
         self.update_graph(msg_id, success=False)
     
     @logged
     def maybe_run(self, msg_id, raw_msg, targets, after, follow, timeout):
         """check location dependencies, and run if they are met."""
         blacklist = self.blacklist.setdefault(msg_id, set())
         if follow or targets or blacklist or self.hwm:
             # we need a can_run filter
             def can_run(idx):
                 # check hwm
                 if self.hwm and self.loads[idx] == self.hwm:
                     return False
                 target = self.targets[idx]
                 # check blacklist
                 if target in blacklist:
                     return False
                 # check targets
                 if targets and target not in targets:
                     return False
                 # check follow
                 return follow.check(self.completed[target], self.failed[target])
             
             indices = filter(can_run, range(len(self.targets)))
 
             if not indices:
                 # couldn't run
                 if follow.all:
                     # check follow for impossibility
                     dests = set()
                     relevant = set()
                     if follow.success:
                         relevant = self.all_completed
                     if follow.failure:
                         relevant = relevant.union(self.all_failed)
                     for m in follow.intersection(relevant):
                         dests.add(self.destinations[m])
                     if len(dests) > 1:
                         self.depending[msg_id] = (raw_msg, targets, after, follow, timeout)
                         self.fail_unreachable(msg_id)
                         return False
                 if targets:
                     # check blacklist+targets for impossibility
                     targets.difference_update(blacklist)
                     if not targets or not targets.intersection(self.targets):
                         self.depending[msg_id] = (raw_msg, targets, after, follow, timeout)
                         self.fail_unreachable(msg_id)
                         return False
                 return False
         else:
             indices = None
             
         self.submit_task(msg_id, raw_msg, targets, follow, timeout, indices)
         return True
             
     @logged
     def save_unmet(self, msg_id, raw_msg, targets, after, follow, timeout):
         """Save a message for later submission when its dependencies are met."""
         self.depending[msg_id] = [raw_msg,targets,after,follow,timeout]
         # track the ids in follow or after, but not those already finished
         for dep_id in after.union(follow).difference(self.all_done):
             if dep_id not in self.graph:
                 self.graph[dep_id] = set()
             self.graph[dep_id].add(msg_id)
     
     @logged
     def submit_task(self, msg_id, raw_msg, targets, follow, timeout, indices=None):
         """Submit a task to any of a subset of our targets."""
         if indices:
             loads = [self.loads[i] for i in indices]
         else:
             loads = self.loads
         idx = self.scheme(loads)
         if indices:
             idx = indices[idx]
         target = self.targets[idx]
         # print (target, map(str, msg[:3]))
         # send job to the engine
         self.engine_stream.send(target, flags=zmq.SNDMORE, copy=False)
         self.engine_stream.send_multipart(raw_msg, copy=False)
         # update load
         self.add_job(idx)
         self.pending[target][msg_id] = (raw_msg, targets, MET, follow, timeout)
         # notify Hub
         content = dict(msg_id=msg_id, engine_id=target)
         self.session.send(self.mon_stream, 'task_destination', content=content, 
                         ident=['tracktask',self.session.session])
         
     
     #-----------------------------------------------------------------------
     # Result Handling
     #-----------------------------------------------------------------------
     @logged
     def dispatch_result(self, raw_msg):
         """dispatch method for result replies"""
         try:
             idents,msg = self.session.feed_identities(raw_msg, copy=False)
             msg = self.session.unpack_message(msg, content=False, copy=False)
             engine = idents[0]
             try:
                 idx = self.targets.index(engine)
             except ValueError:
                 pass # skip load-update for dead engines
             else:
                 self.finish_job(idx)
         except Exception:
             self.log.error("task::Invaid result: %r"%raw_msg, exc_info=True)
             return
 
         header = msg['header']
         parent = msg['parent_header']
         if header.get('dependencies_met', True):
             success = (header['status'] == 'ok')
             msg_id = parent['msg_id']
             retries = self.retries[msg_id]
             if not success and retries > 0:
                 # failed
                 self.retries[msg_id] = retries - 1
                 self.handle_unmet_dependency(idents, parent)
             else:
                 del self.retries[msg_id]
                 # relay to client and update graph
                 self.handle_result(idents, parent, raw_msg, success)
                 # send to Hub monitor
                 self.mon_stream.send_multipart(['outtask']+raw_msg, copy=False)
         else:
             self.handle_unmet_dependency(idents, parent)
         
     @logged
     def handle_result(self, idents, parent, raw_msg, success=True):
         """handle a real task result, either success or failure"""
         # first, relay result to client
         engine = idents[0]
         client = idents[1]
         # swap_ids for XREP-XREP mirror
         raw_msg[:2] = [client,engine]
         # print (map(str, raw_msg[:4]))
         self.client_stream.send_multipart(raw_msg, copy=False)
         # now, update our data structures
         msg_id = parent['msg_id']
         self.blacklist.pop(msg_id, None)
         self.pending[engine].pop(msg_id)
         if success:
             self.completed[engine].add(msg_id)
             self.all_completed.add(msg_id)
         else:
             self.failed[engine].add(msg_id)
             self.all_failed.add(msg_id)
         self.all_done.add(msg_id)
         self.destinations[msg_id] = engine
         
         self.update_graph(msg_id, success)
         
     @logged
     def handle_unmet_dependency(self, idents, parent):
         """handle an unmet dependency"""
         engine = idents[0]
         msg_id = parent['msg_id']
         
         if msg_id not in self.blacklist:
             self.blacklist[msg_id] = set()
         self.blacklist[msg_id].add(engine)
         
         args = self.pending[engine].pop(msg_id)
         raw,targets,after,follow,timeout = args
         
         if self.blacklist[msg_id] == targets:
             self.depending[msg_id] = args
             self.fail_unreachable(msg_id)
         elif not self.maybe_run(msg_id, *args):
             # resubmit failed
             if msg_id not in self.all_failed:
                 # put it back in our dependency tree
                 self.save_unmet(msg_id, *args)
         
         if self.hwm:
             try:
                 idx = self.targets.index(engine)
             except ValueError:
                 pass # skip load-update for dead engines
             else:
                 if self.loads[idx] == self.hwm-1:
                     self.update_graph(None)
         
         
     
     @logged
     def update_graph(self, dep_id=None, success=True):
         """dep_id just finished. Update our dependency
         graph and submit any jobs that just became runable.
         
         Called with dep_id=None to update entire graph for hwm, but without finishing
         a task.
         """
         # print ("\n\n***********")
         # pprint (dep_id)
         # pprint (self.graph)
         # pprint (self.depending)
         # pprint (self.all_completed)
         # pprint (self.all_failed)
         # print ("\n\n***********\n\n")
         # update any jobs that depended on the dependency
         jobs = self.graph.pop(dep_id, [])
 
         # recheck *all* jobs if
         # a) we have HWM and an engine just become no longer full
         # or b) dep_id was given as None
         if dep_id is None or self.hwm and any( [ load==self.hwm-1 for load in self.loads ]):
             jobs = self.depending.keys()
         
         for msg_id in jobs:
             raw_msg, targets, after, follow, timeout = self.depending[msg_id]
             
             if after.unreachable(self.all_completed, self.all_failed)\
                     or follow.unreachable(self.all_completed, self.all_failed):
                 self.fail_unreachable(msg_id)
             
             elif after.check(self.all_completed, self.all_failed): # time deps met, maybe run
                 if self.maybe_run(msg_id, raw_msg, targets, MET, follow, timeout):
                     
                     self.depending.pop(msg_id)
                     for mid in follow.union(after):
                         if mid in self.graph:
                             self.graph[mid].remove(msg_id)
     
     #----------------------------------------------------------------------
     # methods to be overridden by subclasses
     #----------------------------------------------------------------------
     
     def add_job(self, idx):
         """Called after self.targets[idx] just got the job with header.
         Override with subclasses.  The default ordering is simple LRU.
         The default loads are the number of outstanding jobs."""
         self.loads[idx] += 1
         for lis in (self.targets, self.loads):
             lis.append(lis.pop(idx))
             
     
     def finish_job(self, idx):
         """Called after self.targets[idx] just finished a job.
         Override with subclasses."""
         self.loads[idx] -= 1
     
 
 
 def launch_scheduler(in_addr, out_addr, mon_addr, not_addr, config=None,
                         logname='root', log_url=None, loglevel=logging.DEBUG,
-                        identity=b'task'):
-    from zmq.eventloop import ioloop
-    from zmq.eventloop.zmqstream import ZMQStream
+                        identity=b'task', in_thread=False):
+    
+    ZMQStream = zmqstream.ZMQStream
     
     if config:
         # unwrap dict back into Config
         config = Config(config)
 
-    ctx = zmq.Context()
-    loop = ioloop.IOLoop()
+    if in_thread:
+        # use instance() to get the same Context/Loop as our parent
+        ctx = zmq.Context.instance()
+        loop = ioloop.IOLoop.instance()
+    else:
+        # in a process, don't use instance() 
+        # for safety with multiprocessing
+        ctx = zmq.Context()
+        loop = ioloop.IOLoop()
     ins = ZMQStream(ctx.socket(zmq.XREP),loop)
     ins.setsockopt(zmq.IDENTITY, identity)
     ins.bind(in_addr)
     
     outs = ZMQStream(ctx.socket(zmq.XREP),loop)
     outs.setsockopt(zmq.IDENTITY, identity)
     outs.bind(out_addr)
-    mons = ZMQStream(ctx.socket(zmq.PUB),loop)
+    mons = zmqstream.ZMQStream(ctx.socket(zmq.PUB),loop)
     mons.connect(mon_addr)
-    nots = ZMQStream(ctx.socket(zmq.SUB),loop)
+    nots = zmqstream.ZMQStream(ctx.socket(zmq.SUB),loop)
     nots.setsockopt(zmq.SUBSCRIBE, '')
     nots.connect(not_addr)
     
-    # setup logging. Note that these will not work in-process, because they clobber
-    # existing loggers.
-    if log_url:
-        log = connect_logger(logname, ctx, log_url, root="scheduler", loglevel=loglevel)
+    # setup logging.
+    if in_thread:
+        log = Application.instance().log
     else:
-        log = local_logger(logname, loglevel)
+        if log_url:
+            log = connect_logger(logname, ctx, log_url, root="scheduler", loglevel=loglevel)
+        else:
+            log = local_logger(logname, loglevel)
     
     scheduler = TaskScheduler(client_stream=ins, engine_stream=outs,
                             mon_stream=mons, notifier_stream=nots,
                             loop=loop, log=log,
                             config=config)
     scheduler.start()
-    try:
-        loop.start()
-    except KeyboardInterrupt:
-        print ("interrupted, exiting...", file=sys.__stderr__)
+    if not in_thread:
+        try:
+            loop.start()
+        except KeyboardInterrupt:
+            print ("interrupted, exiting...", file=sys.__stderr__)