add HWM to TaskScheduler

This option allows users to limit the level of greedy task assignment.

IPython/config/default/ipcontroller_config.py

@@ -1,180 +1,180 b''
 from IPython.config.loader import Config
 
 c = get_config()
 
 #-----------------------------------------------------------------------------
 # Global configuration
 #-----------------------------------------------------------------------------
 
 # Basic Global config attributes
 
 # Start up messages are logged to stdout using the logging module.
 # These all happen before the twisted reactor is started and are
 # useful for debugging purposes. Can be (10=DEBUG,20=INFO,30=WARN,40=CRITICAL) 
 # and smaller is more verbose.
 # c.Global.log_level = 20
 
 # Log to a file in cluster_dir/log, otherwise just log to sys.stdout.
 # c.Global.log_to_file = False
 
 # Remove old logs from cluster_dir/log before starting.
 # c.Global.clean_logs = True
 
 # A list of Python statements that will be run before starting the 
 # controller. This is provided because occasionally certain things need to 
 # be imported in the controller for pickling to work.
 # c.Global.import_statements = ['import math']
 
 # Reuse the controller's JSON files. If False, JSON files are regenerated
 # each time the controller is run. If True, they will be reused, *but*, you
 # also must set the network ports by hand. If set, this will override the
 # values set for the client and engine connections below.
 # c.Global.reuse_files = True
 
 # Enable exec_key authentication on all messages. Default is True
 # c.Global.secure = True
 
 # The working directory for the process. The application will use os.chdir
 # to change to this directory before starting.
 # c.Global.work_dir = os.getcwd()
 
 # The log url for logging to an `iploggerz` application.  This will override
 # log-to-file.
 # c.Global.log_url = 'tcp://127.0.0.1:20202'
 
 # The specific external IP that is used to disambiguate multi-interface URLs.
 # The default behavior is to guess from external IPs gleaned from `socket`.
 # c.Global.location = '192.168.1.123'
 
 # The ssh server remote clients should use to connect to this controller.
 # It must be a machine that can see the interface specified in client_ip.
 # The default for client_ip is localhost, in which case the sshserver must
 # be an external IP of the controller machine.
 # c.Global.sshserver = 'controller.example.com'
 
 # the url to use for registration.  If set, this overrides engine-ip,
 # engine-transport client-ip,client-transport, and regport.
 # c.RegistrationFactory.url = 'tcp://*:12345'
 
 # the port to use for registration.  Clients and Engines both use this
 # port for registration.
 # c.RegistrationFactory.regport = 10101
 
 #-----------------------------------------------------------------------------
 # Configure the Task Scheduler
 #-----------------------------------------------------------------------------
 
 # The routing scheme. 'pure' will use the pure-ZMQ scheduler. Any other 
 # value will use a Python scheduler with various routing schemes.
 # python schemes are: lru, weighted, random, twobin. Default is 'weighted'.
 # Note that the pure ZMQ scheduler does not support many features, such as
 # dying engines, dependencies, or engine-subset load-balancing.
 # c.ControllerFactory.scheme = 'pure'
 
-# The pure ZMQ scheduler can limit the number of outstanding tasks per engine
-# by using the ZMQ HWM option.  This allows engines with long-running tasks
+# The Python scheduler can limit the number of outstanding tasks per engine
+# by using an HWM option.  This allows engines with long-running tasks
 # to not steal too many tasks from other engines. The default is 0, which
 # means agressively distribute messages, never waiting for them to finish.
-# c.ControllerFactory.hwm = 1
+# c.TaskScheduler.hwm = 0
 
 # Whether to use Threads or Processes to start the Schedulers.  Threads will
 # use less resources, but potentially reduce throughput. Default is to 
 # use processes.  Note that the a Python scheduler will always be in a Process.
 # c.ControllerFactory.usethreads
 
 #-----------------------------------------------------------------------------
 # Configure the Hub
 #-----------------------------------------------------------------------------
 
 # Which class to use for the db backend.  Currently supported are DictDB (the
 # default), and MongoDB. Uncomment this line to enable MongoDB, which will
 # slow-down the Hub's responsiveness, but also reduce its memory footprint.
 # c.HubFactory.db_class = 'IPython.parallel.controller.mongodb.MongoDB'
 
 # The heartbeat ping frequency.  This is the frequency (in ms) at which the
 # Hub pings engines for heartbeats.  This determines how quickly the Hub
 # will react to engines coming and going.  A lower number means faster response
 # time, but more network activity.  The default is 100ms
 # c.HubFactory.ping = 100
 
 # HubFactory queue port pairs, to set by name: mux, iopub, control, task.  Set
 # each as a tuple of length 2 of ints.  The default is to find random
 # available ports
 # c.HubFactory.mux = (10102,10112)
 
 #-----------------------------------------------------------------------------
 # Configure the client connections
 #-----------------------------------------------------------------------------
 
 # Basic client connection config attributes
 
 # The network interface the controller will listen on for client connections.
 # This should be an IP address or interface on the controller. An asterisk
 # means listen on all interfaces. The transport can be any transport
 # supported by zeromq (tcp,epgm,pgm,ib,ipc):
 # c.HubFactory.client_ip = '*'
 # c.HubFactory.client_transport = 'tcp'
 
 # individual client ports to configure by name: query_port, notifier_port
 # c.HubFactory.query_port = 12345
 
 #-----------------------------------------------------------------------------
 # Configure the engine connections
 #-----------------------------------------------------------------------------
 
 # Basic config attributes for the engine connections.
 
 # The network interface the controller will listen on for engine connections.
 # This should be an IP address or interface on the controller. An asterisk
 # means listen on all interfaces. The transport can be any transport
 # supported by zeromq (tcp,epgm,pgm,ib,ipc):
 # c.HubFactory.engine_ip = '*'
 # c.HubFactory.engine_transport = 'tcp'
 
 # set the engine heartbeat ports to use:
 # c.HubFactory.hb = (10303,10313)
 
 #-----------------------------------------------------------------------------
 # Configure the TaskRecord database backend
 #-----------------------------------------------------------------------------
 
 # For memory/persistance reasons, tasks can be stored out-of-memory in a database.
 # Currently, only sqlite and mongodb are supported as backends, but the interface
 # is fairly simple, so advanced developers could write their own backend.
 
 # ----- in-memory configuration --------
 # this line restores the default behavior: in-memory storage of all results.
 # c.HubFactory.db_class = 'IPython.parallel.controller.dictdb.DictDB'
 
 # ----- sqlite configuration --------
 # use this line to activate sqlite:
 # c.HubFactory.db_class = 'IPython.parallel.controller.sqlitedb.SQLiteDB'
 
 # You can specify the name of the db-file.  By default, this will be located
 # in the active cluster_dir, e.g. ~/.ipython/clusterz_default/tasks.db
 # c.SQLiteDB.filename = 'tasks.db'
 
 # You can also specify the location of the db-file, if you want it to be somewhere
 # other than the cluster_dir.
 # c.SQLiteDB.location = '/scratch/'
 
 # This will specify the name of the table for the controller to use.  The default
 # behavior is to use the session ID of the SessionFactory object (a uuid). Overriding
 # this will result in results persisting for multiple sessions.
 # c.SQLiteDB.table = 'results'
 
 # ----- mongodb configuration --------
 # use this line to activate mongodb:
 # c.HubFactory.db_class = 'IPython.parallel.controller.mongodb.MongoDB'
 
 # You can specify the args and kwargs pymongo will use when creating the Connection.
 # For more information on what these options might be, see pymongo documentation.
 # c.MongoDB.connection_kwargs = {}
 # c.MongoDB.connection_args = []
 
 # This will specify the name of the mongo database for the controller to use.  The default
 # behavior is to use the session ID of the SessionFactory object (a uuid). Overriding
 # this will result in task results persisting through multiple sessions.
 # c.MongoDB.database = 'ipythondb'
 
 

IPython/parallel/controller/scheduler.py

@@ -1,596 +1,621 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.
 """
 #-----------------------------------------------------------------------------
 #  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.loader import Config
-from IPython.utils.traitlets import Instance, Dict, List, Set
+from IPython.utils.traitlets import Instance, Dict, List, Set, Int
 
 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) # limit number of outstanding tasks
+    
     # input arguments:
     scheme = Instance(FunctionType, default=leastload) # function for determining the destination
     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 ]
+    # 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...%r"%self)
     
     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."""
         idents,msg = self.session.feed_identities(msg)
         msg = self.session.unpack_message(msg)
         msg_type = msg['msg_type']
         handler = self._notification_handlers.get(msg_type, None)
         if handler is None:
             raise Exception("Unhandled message type: %s"%msg_type)
         else:
             try:
                 handler(str(msg['content']['queue']))
             except KeyError:
                 self.log.error("task::Invalid notification msg: %s"%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()
 
     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()
         
         self.completed.pop(uid)
         self.failed.pop(uid)
         # don't pop destinations, because it might be used later
         # map(self.destinations.pop, self.completed.pop(uid))
         # map(self.destinations.pop, self.failed.pop(uid))
-        
         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()
     
     @logged
     def handle_stranded_tasks(self, engine):
         """Deal with jobs resident in an engine that died."""
         lost = self.pending.pop(engine)
         
         for msg_id, (raw_msg, targets, MET, follow, timeout) in lost.iteritems():
             self.all_failed.add(msg_id)
             self.all_done.add(msg_id)
             idents,msg = self.session.feed_identities(raw_msg, copy=False)
             msg = self.session.unpack_message(msg, copy=False, content=False)
             parent = msg['header']
             idents = [idents[0],engine]+idents[1:]
             # print (idents)
             try:
                 raise error.EngineError("Engine %r died while running task %r"%(engine, msg_id))
             except:
                 content = error.wrap_exception()
             msg = self.session.send(self.client_stream, 'apply_reply', content, 
                                                     parent=parent, ident=idents)
             self.session.send(self.mon_stream, msg, ident=['outtask']+idents)
             self.update_graph(msg_id)
     
     
     #-----------------------------------------------------------------------
     # 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:
+        except Exception:
             self.log.error("task::Invaid task: %s"%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', []))
         
         # 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
                 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, timeout=True)
                 
     @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)
         msg = self.session.unpack_message(msg, copy=False, content=False)
         header = msg['header']
         
         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:
+        if follow or targets or blacklist or self.hwm:
             # we need a can_run filter
             def can_run(idx):
-                target = self.targets[idx]
-                # check targets
-                if targets and target not in targets:
+                # check hwm
+                if 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.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.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)
-        except:
+            engine = idents[0]
+            idx = self.targets.index(engine)
+            self.finish_job(idx)
+        except Exception:
             self.log.error("task::Invaid result: %s"%raw_msg, exc_info=True)
             return
-        
+
         header = msg['header']
         if header.get('dependencies_met', True):
             success = (header['status'] == 'ok')
             self.handle_result(idents, msg['parent_header'], raw_msg, success)
             # send to Hub monitor
             self.mon_stream.send_multipart(['outtask']+raw_msg, copy=False)
         else:
             self.handle_unmet_dependency(idents, msg['parent_header'])
         
     @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
-            return self.fail_unreachable(msg_id)
-        
+            self.fail_unreachable(msg_id)
         elif not self.maybe_run(msg_id, *args):
             # resubmit failed, put it back in our dependency tree
             self.save_unmet(msg_id, *args)
         
+        if self.hwm:
+            idx = self.targets.index(engine)
+            if self.loads[idx] == self.hwm-1:
+                self.update_graph(None)
+        
+        
     
     @logged
-    def update_graph(self, dep_id, success=True):
+    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."""
+        graph and submit any jobs that just became runable.
+        
+        Called with dep_id=None to update 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")
-        if dep_id not in self.graph:
-            return
-        jobs = self.graph.pop(dep_id)
+        # update any jobs that depended on the dependency
+        jobs = self.graph.pop(dep_id, [])
+        # if we have HWM and an engine just become no longer full
+        # recheck *all* jobs:
+        if 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='ZMQ', 
                             log_addr=None, loglevel=logging.DEBUG, scheme='lru',
                             identity=b'task'):
     from zmq.eventloop import ioloop
     from zmq.eventloop.zmqstream import ZMQStream
     
     if config:
         # unwrap dict back into Config
         config = Config(config)
 
     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.connect(mon_addr)
     nots = ZMQStream(ctx.socket(zmq.SUB),loop)
     nots.setsockopt(zmq.SUBSCRIBE, '')
     nots.connect(not_addr)
     
     scheme = globals().get(scheme, None)
     # setup logging
     if log_addr:
         connect_logger(logname, ctx, log_addr, root="scheduler", loglevel=loglevel)
     else:
         local_logger(logname, loglevel)
     
     scheduler = TaskScheduler(client_stream=ins, engine_stream=outs,
                             mon_stream=mons, notifier_stream=nots,
                             scheme=scheme, loop=loop, logname=logname,
                             config=config)
     scheduler.start()
     try:
         loop.start()
     except KeyboardInterrupt:
         print ("interrupted, exiting...", file=sys.__stderr__)