"""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 import time 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, Integer, Enum, CBytes from IPython.parallel import error from IPython.parallel.factory import SessionFactory from IPython.parallel.util import connect_logger, local_logger, asbytes 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 Job(object): """Simple container for a job""" def __init__(self, msg_id, raw_msg, idents, msg, header, targets, after, follow, timeout): self.msg_id = msg_id self.raw_msg = raw_msg self.idents = idents self.msg = msg self.header = header self.targets = targets self.after = after self.follow = follow self.timeout = timeout self.timestamp = time.time() self.blacklist = set() @property def dependents(self): return self.follow.union(self.after) 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 = Integer(1, config=True, 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. The default (1) means that only one task can be outstanding on each engine. Setting TaskScheduler.hwm=0 means there is no limit, and the engines continue to be assigned tasks while they are working, effectively hiding network latency behind computation, but can result in an imbalance of work when submitting many heterogenous tasks all at once. Any positive value greater than one is a compromise between the two. """ ) scheme_name = Enum(('leastload', 'pure', 'lru', 'plainrandom', 'weighted', 'twobin'), 'leastload', config=True, 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 Jobs 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 auditor = Instance('zmq.eventloop.ioloop.PeriodicCallback') ident = CBytes() # ZMQ identity. This should just be self.session.session # but ensure Bytes def _ident_default(self): return self.session.bsession def start(self): self.engine_stream.on_recv(self.dispatch_result, copy=False) self.client_stream.on_recv(self.dispatch_submission, 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.unserialize(msg) except ValueError: self.log.warn("task::Unauthorized message from: %r"%idents) return msg_type = msg['header']['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(asbytes(msg['content']['queue'])) except Exception: self.log.error("task::Invalid notification msg: %r", msg, exc_info=True) 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] = {} # 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 pass # 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) 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() # build fake header header = dict( status='error', engine=engine, date=datetime.now(), ) msg = self.session.msg('apply_reply', content, parent=parent, subheader=header) 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 #----------------------------------------------------------------------- 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.unserialize(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([b'intask']+raw_msg, copy=False) header = msg['header'] msg_id = header['msg_id'] self.all_ids.add(msg_id) # get targets as a set of bytes objects # from a list of unicode objects targets = header.get('targets', []) targets = map(asbytes, targets) targets = set(targets) retries = header.get('retries', 0) self.retries[msg_id] = retries # time dependencies after = header.get('after', None) if after: after = Dependency(after) if after.all: if after.success: after = Dependency(after.difference(self.all_completed), success=after.success, failure=after.failure, all=after.all, ) if after.failure: after = Dependency(after.difference(self.all_failed), success=after.success, failure=after.failure, all=after.all, ) if after.check(self.all_completed, self.all_failed): # recast as empty set, if `after` already met, # to prevent unnecessary set comparisons after = MET else: after = MET # location dependencies follow = Dependency(header.get('follow', [])) # turn timeouts into datetime objects: timeout = header.get('timeout', None) if timeout: # cast to float, because jsonlib returns floats as decimal.Decimal, # which timedelta does not accept timeout = datetime.now() + timedelta(0,float(timeout),0) job = Job(msg_id=msg_id, raw_msg=raw_msg, idents=idents, msg=msg, header=header, targets=targets, after=after, follow=follow, timeout=timeout, ) # validate and reduce dependencies: for dep in after,follow: if not dep: # empty dependency continue # check valid: if msg_id in dep or dep.difference(self.all_ids): self.depending[msg_id] = job return self.fail_unreachable(msg_id, error.InvalidDependency) # check if unreachable: if dep.unreachable(self.all_completed, self.all_failed): self.depending[msg_id] = job 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(job): # can't run yet if msg_id not in self.all_failed: # could have failed as unreachable self.save_unmet(job) else: self.save_unmet(job) 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: job = self.depending[msg_id] if job.timeout and job.timeout < now: self.fail_unreachable(msg_id, error.TaskTimeout) 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 job = self.depending.pop(msg_id) for mid in job.dependents: if mid in self.graph: self.graph[mid].remove(msg_id) 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=job.header, ident=job.idents) self.session.send(self.mon_stream, msg, ident=[b'outtask']+job.idents) self.update_graph(msg_id, success=False) def maybe_run(self, job): """check location dependencies, and run if they are met.""" msg_id = job.msg_id self.log.debug("Attempting to assign task %s", msg_id) if not self.targets: # no engines, definitely can't run return False if job.follow or job.targets or job.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 job.blacklist: return False # check targets if job.targets and target not in job.targets: return False # check follow return job.follow.check(self.completed[target], self.failed[target]) indices = filter(can_run, range(len(self.targets))) if not indices: # couldn't run if job.follow.all: # check follow for impossibility dests = set() relevant = set() if job.follow.success: relevant = self.all_completed if job.follow.failure: relevant = relevant.union(self.all_failed) for m in job.follow.intersection(relevant): dests.add(self.destinations[m]) if len(dests) > 1: self.depending[msg_id] = job self.fail_unreachable(msg_id) return False if job.targets: # check blacklist+targets for impossibility job.targets.difference_update(job.blacklist) if not job.targets or not job.targets.intersection(self.targets): self.depending[msg_id] = job self.fail_unreachable(msg_id) return False return False else: indices = None self.submit_task(job, indices) return True def save_unmet(self, job): """Save a message for later submission when its dependencies are met.""" msg_id = job.msg_id self.depending[msg_id] = job # track the ids in follow or after, but not those already finished for dep_id in job.after.union(job.follow).difference(self.all_done): if dep_id not in self.graph: self.graph[dep_id] = set() self.graph[dep_id].add(msg_id) def submit_task(self, job, 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(job.raw_msg, copy=False) # update load self.add_job(idx) self.pending[target][job.msg_id] = job # notify Hub content = dict(msg_id=job.msg_id, engine_id=target.decode('ascii')) self.session.send(self.mon_stream, 'task_destination', content=content, ident=[b'tracktask',self.ident]) #----------------------------------------------------------------------- # Result Handling #----------------------------------------------------------------------- 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.unserialize(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([b'outtask']+raw_msg, copy=False) else: self.handle_unmet_dependency(idents, parent) 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.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) def handle_unmet_dependency(self, idents, parent): """handle an unmet dependency""" engine = idents[0] msg_id = parent['msg_id'] job = self.pending[engine].pop(msg_id) job.blacklist.add(engine) if job.blacklist == job.targets: self.depending[msg_id] = job self.fail_unreachable(msg_id) elif not self.maybe_run(job): # resubmit failed if msg_id not in self.all_failed: # put it back in our dependency tree self.save_unmet(job) 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) 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 sorted(jobs, key=lambda msg_id: self.depending[msg_id].timestamp): job = self.depending[msg_id] if job.after.unreachable(self.all_completed, self.all_failed)\ or job.follow.unreachable(self.all_completed, self.all_failed): self.fail_unreachable(msg_id) elif job.after.check(self.all_completed, self.all_failed): # time deps met, maybe run if self.maybe_run(job): self.depending.pop(msg_id) for mid in job.dependents: 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', in_thread=False): ZMQStream = zmqstream.ZMQStream if config: # unwrap dict back into Config config = Config(config) 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.ROUTER),loop) ins.setsockopt(zmq.IDENTITY, identity) ins.bind(in_addr) outs = ZMQStream(ctx.socket(zmq.ROUTER),loop) outs.setsockopt(zmq.IDENTITY, identity) outs.bind(out_addr) mons = zmqstream.ZMQStream(ctx.socket(zmq.PUB),loop) mons.connect(mon_addr) nots = zmqstream.ZMQStream(ctx.socket(zmq.SUB),loop) nots.setsockopt(zmq.SUBSCRIBE, b'') nots.connect(not_addr) # setup logging. if in_thread: log = Application.instance().log else: 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() if not in_thread: try: loop.start() except KeyboardInterrupt: scheduler.log.critical("Interrupted, exiting...")