2 # NEPI, a framework to manage network experiments
3 # Copyright (C) 2013 INRIA
5 # This program is free software: you can redistribute it and/or modify
6 # it under the terms of the GNU General Public License as published by
7 # the Free Software Foundation, either version 3 of the License, or
8 # (at your option) any later version.
10 # This program is distributed in the hope that it will be useful,
11 # but WITHOUT ANY WARRANTY; without even the implied warranty of
12 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 # GNU General Public License for more details.
15 # You should have received a copy of the GNU General Public License
16 # along with this program. If not, see <http://www.gnu.org/licenses/>.
18 # Author: Alina Quereilhac <alina.quereilhac@inria.fr>
28 from nepi.util import guid
29 from nepi.util.parallel import ParallelRun
30 from nepi.util.timefuncs import tnow, tdiffsec, stabsformat, tsformat
31 from nepi.execution.resource import ResourceFactory, ResourceAction, \
32 ResourceState, ResourceState2str
33 from nepi.execution.scheduler import HeapScheduler, Task, TaskStatus
34 from nepi.execution.trace import TraceAttr
36 # TODO: use multiprocessing instead of threading
37 # TODO: When a failure occurs during deployment, scp and ssh processes are left running behind!!
38 # TODO: Allow to reconnect to a running experiment instance! (reconnect mode vs deploy mode)
40 class ECState(object):
41 """ State of the Experiment Controller
48 class ExperimentController(object):
50 .. class:: Class Args :
52 :param exp_id: Human readable identifier for the experiment scenario.
53 It will be used in the name of the directory
54 where experiment related information is stored
59 An experiment, or scenario, is defined by a concrete use, behavior,
60 configuration and interconnection of resources that describe a single
61 experiment case (We call this the experiment description).
62 A same experiment (scenario) can be run many times.
64 The ExperimentController (EC), is the entity responsible for
65 managing an experiment instance (run). The same scenario can be
66 recreated (and re-run) by instantiating an EC and recreating
67 the same experiment description.
69 In NEPI, an experiment is represented as a graph of interconnected
70 resources. A resource is a generic concept in the sense that any
71 component taking part of an experiment, whether physical of
72 virtual, is considered a resource. A resources could be a host,
73 a virtual machine, an application, a simulator, a IP address.
75 A ResourceManager (RM), is the entity responsible for managing a
76 single resource. ResourceManagers are specific to a resource
77 type (i.e. An RM to control a Linux application will not be
78 the same as the RM used to control a ns-3 simulation).
79 In order for a new type of resource to be supported in NEPI
80 a new RM must be implemented. NEPI already provides different
81 RMs to control basic resources, and new can be extended from
84 Through the EC interface the user can create ResourceManagers (RMs),
85 configure them and interconnect them, in order to describe an experiment.
86 Describing an experiment through the EC does not run the experiment.
87 Only when the 'deploy()' method is invoked on the EC, will the EC take
88 actions to transform the 'described' experiment into a 'running' experiment.
90 While the experiment is running, it is possible to continue to
91 create/configure/connect RMs, and to deploy them to involve new
92 resources in the experiment (this is known as 'interactive' deployment).
94 An experiments in NEPI is identified by a string id,
95 which is either given by the user, or automatically generated by NEPI.
96 The purpose of this identifier is to separate files and results that
97 belong to different experiment scenarios.
98 However, since a same 'experiment' can be run many times, the experiment
99 id is not enough to identify an experiment instance (run).
100 For this reason, the ExperimentController has two identifier, the
101 exp_id, which can be re-used by different ExperimentController instances,
102 and the run_id, which unique to a ExperimentController instance, and
103 is automatically generated by NEPI.
107 def __init__(self, exp_id = None):
108 super(ExperimentController, self).__init__()
109 # root directory to store files
111 # Run identifier. It identifies a concrete instance (run) of an experiment.
112 # Since a same experiment (same configuration) can be run many times,
113 # this id permits to identify concrete exoeriment run
114 self._run_id = tsformat()
116 # Experiment identifier. Usually assigned by the user
117 self._exp_id = exp_id or "exp-%s" % os.urandom(8).encode('hex')
119 # generator of globally unique ids
120 self._guid_generator = guid.GuidGenerator()
123 self._resources = dict()
126 self._scheduler = HeapScheduler()
131 # Event processing thread
132 self._cond = threading.Condition()
133 self._thread = threading.Thread(target = self._process)
134 self._thread.setDaemon(True)
138 self._state = ECState.RUNNING
141 self._logger = logging.getLogger("ExperimentController")
145 """ Return the logger of the Experiment Controller
152 """ Return the state of the Experiment Controller
159 """ Return the experiment id assigned by the user
166 """ Return the experiment instance (run) identifier
173 """ Put the state of the Experiment Controller into a final state :
174 Either TERMINATED or FAILED
177 return self.ecstate in [ECState.FAILED, ECState.TERMINATED]
179 def wait_finished(self, guids):
180 """ Blocking method that wait until all the RM from the 'guid' list
181 reached the state FINISHED
183 :param guids: List of guids
186 return self.wait(guids)
188 def wait_started(self, guids):
189 """ Blocking method that wait until all the RM from the 'guid' list
190 reached the state STARTED
192 :param guids: List of guids
195 return self.wait(guids, states = [ResourceState.STARTED,
196 ResourceState.STOPPED,
197 ResourceState.FAILED,
198 ResourceState.FINISHED])
200 def wait(self, guids, states = [ResourceState.FINISHED,
201 ResourceState.FAILED,
202 ResourceState.STOPPED]):
203 """ Blocking method that waits until all the RM from the 'guid' list
204 reached state 'state' or until a failure occurs
206 :param guids: List of guids
209 if isinstance(guids, int):
212 # we randomly alter the order of the guids to avoid ordering
213 # dependencies (e.g. LinuxApplication RMs runing on the same
214 # linux host will be synchronized by the LinuxNode SSH lock)
215 random.shuffle(guids)
218 # If no more guids to wait for or an error occured, then exit
219 if len(guids) == 0 or self.finished:
222 # If a guid reached one of the target states, remove it from list
224 state = self.state(guid)
230 self.logger.debug(" WAITING FOR %g - state %s " % (guid,
231 self.state(guid, hr = True)))
233 # Take the opportunity to 'refresh' the states of the RMs.
234 # Query only the first up to N guids (not to overwhelm
237 lim = n if len(guids) > n else ( len(guids) -1 )
238 nguids = guids[0: lim]
240 # schedule state request for all guids (take advantage of
241 # scheduler multi threading).
243 callback = functools.partial(self.state, guid)
244 self.schedule("0s", callback)
246 # If the guid is not in one of the target states, wait and
247 # continue quering. We keep the sleep big to decrease the
248 # number of RM state queries
251 def get_task(self, tid):
252 """ Get a specific task
254 :param tid: Id of the task
258 return self._tasks.get(tid)
260 def get_resource(self, guid):
261 """ Get a specific Resource Manager
263 :param guid: Id of the task
265 :rtype: ResourceManager
267 return self._resources.get(guid)
271 """ Returns the list of all the Resource Manager Id
276 return self._resources.keys()
278 def register_resource(self, rtype, guid = None):
279 """ Register a Resource Manager. It creates a new 'guid', if it is not specified,
280 for the RM of type 'rtype' and add it to the list of Resources.
282 :param rtype: Type of the RM
284 :return: Id of the RM
287 # Get next available guid
288 guid = self._guid_generator.next(guid)
291 rm = ResourceFactory.create(rtype, self, guid)
294 self._resources[guid] = rm
298 def get_attributes(self, guid):
299 """ Return all the attibutes of a specific RM
301 :param guid: Guid of the RM
303 :return: List of attributes
306 rm = self.get_resource(guid)
307 return rm.get_attributes()
309 def register_connection(self, guid1, guid2):
310 """ Registers a guid1 with a guid2.
311 The declaration order is not important
313 :param guid1: First guid to connect
314 :type guid1: ResourceManager
316 :param guid2: Second guid to connect
317 :type guid: ResourceManager
319 rm1 = self.get_resource(guid1)
320 rm2 = self.get_resource(guid2)
322 rm1.register_connection(guid2)
323 rm2.register_connection(guid1)
325 def register_condition(self, group1, action, group2, state,
327 """ Registers an action START or STOP for all RM on group1 to occur
328 time 'time' after all elements in group2 reached state 'state'.
330 :param group1: List of guids of RMs subjected to action
333 :param action: Action to register (either START or STOP)
334 :type action: ResourceAction
336 :param group2: List of guids of RMs to we waited for
339 :param state: State to wait for on RMs (STARTED, STOPPED, etc)
340 :type state: ResourceState
342 :param time: Time to wait after group2 has reached status
346 if isinstance(group1, int):
348 if isinstance(group2, int):
352 rm = self.get_resource(guid1)
353 rm.register_condition(action, group2, state, time)
355 def register_trace(self, guid, name):
358 :param name: Name of the trace
361 rm = self.get_resource(guid)
362 rm.register_trace(name)
364 def trace(self, guid, name, attr = TraceAttr.ALL, block = 512, offset = 0):
365 """ Get information on collected trace
367 :param name: Name of the trace
370 :param attr: Can be one of:
371 - TraceAttr.ALL (complete trace content),
372 - TraceAttr.STREAM (block in bytes to read starting at offset),
373 - TraceAttr.PATH (full path to the trace file),
374 - TraceAttr.SIZE (size of trace file).
377 :param block: Number of bytes to retrieve from trace, when attr is TraceAttr.STREAM
380 :param offset: Number of 'blocks' to skip, when attr is TraceAttr.STREAM
385 rm = self.get_resource(guid)
386 return rm.trace(name, attr, block, offset)
388 def discover(self, guid):
389 """ Discover a specific RM defined by its 'guid'
391 :param guid: Guid of the RM
395 rm = self.get_resource(guid)
398 def provision(self, guid):
399 """ Provision a specific RM defined by its 'guid'
401 :param guid: Guid of the RM
405 rm = self.get_resource(guid)
406 return rm.provision()
408 def get(self, guid, name):
409 """ Get a specific attribute 'name' from the RM 'guid'
411 :param guid: Guid of the RM
414 :param name: attribute's name
418 rm = self.get_resource(guid)
421 def set(self, guid, name, value):
422 """ Set a specific attribute 'name' from the RM 'guid'
423 with the value 'value'
425 :param guid: Guid of the RM
428 :param name: attribute's name
431 :param value: attribute's value
434 rm = self.get_resource(guid)
435 return rm.set(name, value)
437 def state(self, guid, hr = False):
438 """ Returns the state of a resource
440 :param guid: Resource guid
443 :param hr: Human readable. Forces return of a
444 status string instead of a number
448 rm = self.get_resource(guid)
452 return ResourceState2str.get(state)
456 def stop(self, guid):
457 """ Stop a specific RM defined by its 'guid'
459 :param guid: Guid of the RM
463 rm = self.get_resource(guid)
466 def start(self, guid):
467 """ Start a specific RM defined by its 'guid'
469 :param guid: Guid of the RM
473 rm = self.get_resource(guid)
476 def set_with_conditions(self, name, value, group1, group2, state,
478 """ Set value 'value' on attribute with name 'name' on all RMs of
479 group1 when 'time' has elapsed since all elements in group2
480 have reached state 'state'.
482 :param name: Name of attribute to set in RM
485 :param value: Value of attribute to set in RM
488 :param group1: List of guids of RMs subjected to action
491 :param action: Action to register (either START or STOP)
492 :type action: ResourceAction
494 :param group2: List of guids of RMs to we waited for
497 :param state: State to wait for on RMs (STARTED, STOPPED, etc)
498 :type state: ResourceState
500 :param time: Time to wait after group2 has reached status
504 if isinstance(group1, int):
506 if isinstance(group2, int):
510 rm = self.get_resource(guid)
511 rm.set_with_conditions(name, value, group2, state, time)
513 def stop_with_conditions(self, guid):
514 """ Stop a specific RM defined by its 'guid' only if all the conditions are true
516 :param guid: Guid of the RM
520 rm = self.get_resource(guid)
521 return rm.stop_with_conditions()
523 def start_with_conditions(self, guid):
524 """ Start a specific RM defined by its 'guid' only if all the conditions are true
526 :param guid: Guid of the RM
530 rm = self.get_resource(guid)
531 return rm.start_with_condition()
533 def deploy(self, group = None, wait_all_ready = True):
534 """ Deploy all resource manager in group
536 :param group: List of guids of RMs to deploy
539 :param wait_all_ready: Wait until all RMs are ready in
540 order to start the RMs
544 self.logger.debug(" ------- DEPLOY START ------ ")
547 # By default, if not deployment group is indicated,
548 # all RMs that are undeployed will be deployed
550 for guid in self.resources:
551 if self.state(guid) == ResourceState.NEW:
554 if isinstance(group, int):
557 # Before starting deployment we disorder the group list with the
558 # purpose of speeding up the whole deployment process.
559 # It is likely that the user inserted in the 'group' list closely
560 # resources one after another (e.g. all applications
561 # connected to the same node can likely appear one after another).
562 # This can originate a slow down in the deployment since the N
563 # threads the parallel runner uses to processes tasks may all
564 # be taken up by the same family of resources waiting for the
565 # same conditions (e.g. LinuxApplications running on a same
566 # node share a single lock, so they will tend to be serialized).
567 # If we disorder the group list, this problem can be mitigated.
568 random.shuffle(group)
570 def wait_all_and_start(group):
573 if self.state(guid) < ResourceState.READY:
578 callback = functools.partial(wait_all_and_start, group)
579 self.schedule("1s", callback)
581 # If all resources are read, we schedule the start
583 rm = self.get_resource(guid)
584 self.schedule("0s", rm.start_with_conditions)
587 # Schedule the function that will check all resources are
588 # READY, and only then it will schedule the start.
589 # This is aimed to reduce the number of tasks looping in the scheduler.
590 # Intead of having N start tasks, we will have only one
591 callback = functools.partial(wait_all_and_start, group)
592 self.schedule("1s", callback)
595 rm = self.get_resource(guid)
596 self.schedule("0s", rm.deploy)
598 if not wait_all_ready:
599 self.schedule("1s", rm.start_with_conditions)
601 if rm.conditions.get(ResourceAction.STOP):
602 # Only if the RM has STOP conditions we
603 # schedule a stop. Otherwise the RM will stop immediately
604 self.schedule("2s", rm.stop_with_conditions)
606 def release(self, group = None):
607 """ Release the elements of the list 'group' or
608 all the resources if any group is specified
610 :param group: List of RM
615 group = self.resources
619 rm = self.get_resource(guid)
620 thread = threading.Thread(target=rm.release)
621 threads.append(thread)
622 thread.setDaemon(True)
625 while list(threads) and not self.finished:
627 # Time out after 5 seconds to check EC not terminated
629 if not thread.is_alive():
630 threads.remove(thread)
633 """ Shutdown the Experiment Controller.
634 Releases all the resources and stops task processing thread
639 # Mark the EC state as TERMINATED
640 self._state = ECState.TERMINATED
642 # Notify condition to wake up the processing thread
645 if self._thread.is_alive():
648 def schedule(self, date, callback, track = False):
649 """ Schedule a callback to be executed at time date.
651 :param date: string containing execution time for the task.
652 It can be expressed as an absolute time, using
653 timestamp format, or as a relative time matching
654 ^\d+.\d+(h|m|s|ms|us)$
656 :param callback: code to be executed for the task. Must be a
657 Python function, and receives args and kwargs
660 :param track: if set to True, the task will be retrivable with
661 the get_task() method
663 :return : The Id of the task
665 timestamp = stabsformat(date)
666 task = Task(timestamp, callback)
667 task = self._scheduler.schedule(task)
670 self._tasks[task.id] = task
672 # Notify condition to wake up the processing thread
678 """ Process scheduled tasks.
682 The _process method is executed in an independent thread held by the
683 ExperimentController for as long as the experiment is running.
685 Tasks are scheduled by invoking the schedule method with a target callback.
686 The schedule method is given a execution time which controls the
687 order in which tasks are processed.
689 Tasks are processed in parallel using multithreading.
690 The environmental variable NEPI_NTHREADS can be used to control
691 the number of threads used to process tasks. The default value is 50.
695 To execute tasks in parallel, an ParallelRunner (PR) object, holding
696 a pool of threads (workers), is used.
697 For each available thread in the PR, the next task popped from
698 the scheduler queue is 'put' in the PR.
699 Upon receiving a task to execute, each PR worker (thread) invokes the
700 _execute method of the EC, passing the task as argument.
701 This method, calls task.callback inside a try/except block. If an
702 exception is raised by the tasks.callback, it will be trapped by the
703 try block, logged to standard error (usually the console), and the EC
704 state will be set to ECState.FAILED.
705 The invocation of _notify immediately after, forces the processing
706 loop in the _process method, to wake up if it was blocked waiting for new
707 tasks to arrived, and to check the EC state.
708 As the EC is in FAILED state, the processing loop exits and the
709 'finally' block is invoked. In the 'finally' block, the 'sync' method
710 of the PR is invoked, which forces the PR to raise any unchecked errors
711 that might have been raised by the workers.
714 nthreads = int(os.environ.get("NEPI_NTHREADS", "50"))
716 runner = ParallelRun(maxthreads = nthreads)
720 while not self.finished:
723 task = self._scheduler.next()
726 # No task to execute. Wait for a new task to be scheduled.
729 # The task timestamp is in the future. Wait for timeout
730 # or until another task is scheduled.
732 if now < task.timestamp:
733 # Calculate timeout in seconds
734 timeout = tdiffsec(task.timestamp, now)
736 # Re-schedule task with the same timestamp
737 self._scheduler.schedule(task)
741 # Wait timeout or until a new task awakes the condition
742 self._cond.wait(timeout)
747 # Process tasks in parallel
748 runner.put(self._execute, task)
751 err = traceback.format_exc()
752 self.logger.error("Error while processing tasks in the EC: %s" % err)
754 self._state = ECState.FAILED
756 self.logger.debug("Exiting the task processing loop ... ")
759 def _execute(self, task):
760 """ Executes a single task.
762 :param task: Object containing the callback to execute
767 If the invokation of the task callback raises an
768 exception, the processing thread of the ExperimentController
769 will be stopped and the experiment will be aborted.
773 task.status = TaskStatus.DONE
776 task.result = task.callback()
779 err = traceback.format_exc()
781 task.status = TaskStatus.ERROR
783 self.logger.error("Error occurred while executing task: %s" % err)
785 # Set the EC to FAILED state (this will force to exit the task
787 self._state = ECState.FAILED
789 # Notify condition to wake up the processing thread
792 # Propage error to the ParallelRunner
796 """ Awakes the processing thread in case it is blocked waiting
797 for a new task to be scheduled.