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>
20 from nepi.util import guid
21 from nepi.util.parallel import ParallelRun
22 from nepi.util.timefuncs import tnow, tdiffsec, stabsformat, tsformat
23 from nepi.execution.resource import ResourceFactory, ResourceAction, \
24 ResourceState, ResourceState2str
25 from nepi.execution.scheduler import HeapScheduler, Task, TaskStatus
26 from nepi.execution.trace import TraceAttr
28 # TODO: use multiprocessing instead of threading
29 # TODO: Allow to reconnect to a running experiment instance! (reconnect mode vs deploy mode)
39 class FailureLevel(object):
40 """ Describes the system failure state """
45 class FailureManager(object):
46 """ The FailureManager is responsible for handling errors
47 and deciding whether an experiment should be aborted or not
51 def __init__(self, ec):
52 self._ec = weakref.ref(ec)
53 self._failure_level = FailureLevel.OK
57 """ Returns the ExperimentController associated to this FailureManager
65 if self._failure_level == FailureLevel.OK:
66 for guid in self.ec.resources:
67 state = self.ec.state(guid)
68 critical = self.ec.get(guid, "critical")
69 if state == ResourceState.FAILED and critical:
70 self._failure_level = FailureLevel.RM_FAILURE
71 self.ec.logger.debug("RM critical failure occurred on guid %d." \
72 " Setting EC FAILURE LEVEL to RM_FAILURE" % guid)
75 return self._failure_level != FailureLevel.OK
77 def set_ec_failure(self):
78 self._failure_level = FailureLevel.EC_FAILURE
81 class ECState(object):
82 """ Possible states for an ExperimentController
89 class ExperimentController(object):
91 .. class:: Class Args :
93 :param exp_id: Human readable identifier for the experiment scenario.
98 An experiment, or scenario, is defined by a concrete set of resources,
99 behavior, configuration and interconnection of those resources.
100 The Experiment Description (ED) is a detailed representation of a
101 single experiment. It contains all the necessary information to
102 allow repeating the experiment. NEPI allows to describe
103 experiments by registering components (resources), configuring them
104 and interconnecting them.
106 A same experiment (scenario) can be executed many times, generating
107 different results. We call an experiment execution (instance) a 'run'.
109 The ExperimentController (EC), is the entity responsible of
110 managing an experiment run. The same scenario can be
111 recreated (and re-run) by instantiating an EC and recreating
112 the same experiment description.
114 In NEPI, an experiment is represented as a graph of interconnected
115 resources. A resource is a generic concept in the sense that any
116 component taking part of an experiment, whether physical of
117 virtual, is considered a resource. A resources could be a host,
118 a virtual machine, an application, a simulator, a IP address.
120 A ResourceManager (RM), is the entity responsible for managing a
121 single resource. ResourceManagers are specific to a resource
122 type (i.e. An RM to control a Linux application will not be
123 the same as the RM used to control a ns-3 simulation).
124 To support a new type of resource in NEPI, a new RM must be
125 implemented. NEPI already provides a variety of
126 RMs to control basic resources, and new can be extended from
129 Through the EC interface the user can create ResourceManagers (RMs),
130 configure them and interconnect them, to describe an experiment.
131 Describing an experiment through the EC does not run the experiment.
132 Only when the 'deploy()' method is invoked on the EC, the EC will take
133 actions to transform the 'described' experiment into a 'running' experiment.
135 While the experiment is running, it is possible to continue to
136 create/configure/connect RMs, and to deploy them to involve new
137 resources in the experiment (this is known as 'interactive' deployment).
139 An experiments in NEPI is identified by a string id,
140 which is either given by the user, or automatically generated by NEPI.
141 The purpose of this identifier is to separate files and results that
142 belong to different experiment scenarios.
143 However, since a same 'experiment' can be run many times, the experiment
144 id is not enough to identify an experiment instance (run).
145 For this reason, the ExperimentController has two identifier, the
146 exp_id, which can be re-used in different ExperimentController,
147 and the run_id, which is unique to one ExperimentController instance, and
148 is automatically generated by NEPI.
152 def __init__(self, exp_id = None):
153 super(ExperimentController, self).__init__()
156 self._logger = logging.getLogger("ExperimentController")
158 # Run identifier. It identifies a concrete execution instance (run)
160 # Since a same experiment (same configuration) can be executed many
161 # times, this run_id permits to separate result files generated on
162 # different experiment executions
163 self._run_id = tsformat()
165 # Experiment identifier. Usually assigned by the user
166 # Identifies the experiment scenario (i.e. configuration,
167 # resources used, etc)
168 self._exp_id = exp_id or "exp-%s" % os.urandom(8).encode('hex')
170 # generator of globally unique ids
171 self._guid_generator = guid.GuidGenerator()
174 self._resources = dict()
176 # Scheduler. It a queue that holds tasks scheduled for
177 # execution, and yields the next task to be executed
178 # ordered by execution and arrival time
179 self._scheduler = HeapScheduler()
184 # RM groups (for deployment)
185 self._groups = dict()
187 # generator of globally unique id for groups
188 self._group_id_generator = guid.GuidGenerator()
190 # Flag to stop processing thread
193 # Entity in charge of managing system failures
194 self._fm = FailureManager(self)
197 self._state = ECState.RUNNING
199 # Blacklist file for PL nodes
200 nepi_home = os.path.join(os.path.expanduser("~"), ".nepi")
201 plblacklist_file = os.path.join(nepi_home, "plblacklist.txt")
202 if not os.path.exists(plblacklist_file):
203 if os.path.isdir(nepi_home):
204 open(plblacklist_file, 'w').close()
206 os.makedirs(nepi_home)
207 open(plblacklist_file, 'w').close()
209 # The runner is a pool of threads used to parallelize
211 nthreads = int(os.environ.get("NEPI_NTHREADS", "50"))
212 self._runner = ParallelRun(maxthreads = nthreads)
214 # Event processing thread
215 self._cond = threading.Condition()
216 self._thread = threading.Thread(target = self._process)
217 self._thread.setDaemon(True)
222 """ Returns the logger instance of the Experiment Controller
229 """ Returns the state of the Experiment Controller
236 """ Returns the experiment id assigned by the user
243 """ Returns the experiment instance (run) identifier (automatically
251 """ Returns True if the experiment has failed and should be interrupted,
255 return self._fm.abort
257 def wait_finished(self, guids):
258 """ Blocking method that waits until all RMs in the 'guids' list
259 have reached a state >= STOPPED (i.e. STOPPED, FAILED or
260 RELEASED ), or until a failure in the experiment occurs
263 :param guids: List of guids
271 return self.wait(guids, state = ResourceState.STOPPED,
274 def wait_started(self, guids):
275 """ Blocking method that waits until all RMs in the 'guids' list
276 have reached a state >= STARTED, or until a failure in the
277 experiment occurs (i.e. abort == True)
279 :param guids: List of guids
287 return self.wait(guids, state = ResourceState.STARTED,
290 def wait_released(self, guids):
291 """ Blocking method that waits until all RMs in the 'guids' list
292 have reached a state == RELEASED, or until the EC fails
294 :param guids: List of guids
300 return self._state == ECState.FAILED
302 return self.wait(guids, state = ResourceState.RELEASED,
305 def wait_deployed(self, guids):
306 """ Blocking method that waits until all RMs in the 'guids' list
307 have reached a state >= READY, or until a failure in the
308 experiment occurs (i.e. abort == True)
310 :param guids: List of guids
318 return self.wait(guids, state = ResourceState.READY,
321 def wait(self, guids, state, quit):
322 """ Blocking method that waits until all RMs in the 'guids' list
323 have reached a state >= 'state', or until the 'quit' callback
326 :param guids: List of guids
331 if isinstance(guids, int):
334 # Make a copy to avoid modifying the original guids list
338 # If there are no more guids to wait for
339 # or the quit function returns True, exit the loop
340 if len(guids) == 0 or quit():
343 # If a guid reached one of the target states, remove it from list
345 rstate = self.state(guid)
347 hrrstate = ResourceState2str.get(rstate)
348 hrstate = ResourceState2str.get(state)
352 rm = self.get_resource(guid)
353 self.logger.debug(" %s guid %d DONE - state is %s, required is >= %s " % (
354 rm.get_rtype(), guid, hrrstate, hrstate))
357 self.logger.debug(" WAITING FOR guid %d - state is %s, required is >= %s " % (
358 guid, hrrstate, hrstate))
361 def get_task(self, tid):
362 """ Returns a task by its id
364 :param tid: Id of the task
370 return self._tasks.get(tid)
372 def get_resource(self, guid):
373 """ Returns a registered ResourceManager by its guid
375 :param guid: Id of the task
378 :rtype: ResourceManager
381 return self._resources.get(guid)
385 """ Returns the set() of guids of all the ResourceManager
387 :return: Set of all RM guids
391 return self._resources.keys()
393 def register_resource(self, rtype, guid = None):
394 """ Registers a new ResourceManager of type 'rtype' in the experiment
396 This method will assign a new 'guid' for the RM, if no guid
399 :param rtype: Type of the RM
402 :return: Guid of the RM
406 # Get next available guid
407 guid = self._guid_generator.next(guid)
410 rm = ResourceFactory.create(rtype, self, guid)
413 self._resources[guid] = rm
417 def get_attributes(self, guid):
418 """ Returns all the attributes of the RM with guid 'guid'
420 :param guid: Guid of the RM
423 :return: List of attributes
427 rm = self.get_resource(guid)
428 return rm.get_attributes()
430 def get_attribute(self, guid, name):
431 """ Returns the attribute 'name' of the RM with guid 'guid'
433 :param guid: Guid of the RM
436 :param name: Name of the attribute
439 :return: The attribute with name 'name'
443 rm = self.get_resource(guid)
444 return rm.get_attribute(name)
446 def register_connection(self, guid1, guid2):
447 """ Registers a connection between a RM with guid 'guid1'
448 and another RM with guid 'guid2'.
450 The order of the in which the two guids are provided is not
451 important, since the connection relationship is symmetric.
453 :param guid1: First guid to connect
454 :type guid1: ResourceManager
456 :param guid2: Second guid to connect
457 :type guid: ResourceManager
460 rm1 = self.get_resource(guid1)
461 rm2 = self.get_resource(guid2)
463 rm1.register_connection(guid2)
464 rm2.register_connection(guid1)
466 def register_condition(self, guids1, action, guids2, state,
468 """ Registers an action START, STOP or DEPLOY for all RM on list
469 guids1 to occur at time 'time' after all elements in list guids2
470 have reached state 'state'.
472 :param guids1: List of guids of RMs subjected to action
475 :param action: Action to perform (either START, STOP or DEPLOY)
476 :type action: ResourceAction
478 :param guids2: List of guids of RMs to we waited for
481 :param state: State to wait for on RMs of list guids2 (STARTED,
483 :type state: ResourceState
485 :param time: Time to wait after guids2 has reached status
489 if isinstance(guids1, int):
491 if isinstance(guids2, int):
495 rm = self.get_resource(guid1)
496 rm.register_condition(action, guids2, state, time)
498 def enable_trace(self, guid, name):
499 """ Enables a trace to be collected during the experiment run
501 :param name: Name of the trace
505 rm = self.get_resource(guid)
506 rm.enable_trace(name)
508 def trace_enabled(self, guid, name):
509 """ Returns True if the trace of name 'name' is enabled
511 :param name: Name of the trace
515 rm = self.get_resource(guid)
516 return rm.trace_enabled(name)
518 def trace(self, guid, name, attr = TraceAttr.ALL, block = 512, offset = 0):
519 """ Returns information on a collected trace, the trace stream or
520 blocks (chunks) of the trace stream
522 :param name: Name of the trace
525 :param attr: Can be one of:
526 - TraceAttr.ALL (complete trace content),
527 - TraceAttr.STREAM (block in bytes to read starting
529 - TraceAttr.PATH (full path to the trace file),
530 - TraceAttr.SIZE (size of trace file).
533 :param block: Number of bytes to retrieve from trace, when attr is
537 :param offset: Number of 'blocks' to skip, when attr is TraceAttr.STREAM
543 rm = self.get_resource(guid)
544 return rm.trace(name, attr, block, offset)
546 def get_traces(self, guid):
547 """ Returns the list of the trace names of the RM with guid 'guid'
549 :param guid: Guid of the RM
552 :return: List of trace names
556 rm = self.get_resource(guid)
557 return rm.get_traces()
560 def discover(self, guid):
561 """ Discovers an available resource matching the criteria defined
562 by the RM with guid 'guid', and associates that resource to the RM
564 Not all RM types require (or are capable of) performing resource
565 discovery. For the RM types which are not capable of doing so,
566 invoking this method does not have any consequences.
568 :param guid: Guid of the RM
572 rm = self.get_resource(guid)
575 def provision(self, guid):
576 """ Provisions the resource associated to the RM with guid 'guid'.
578 Provisioning means making a resource 'accessible' to the user.
579 Not all RM types require (or are capable of) performing resource
580 provisioning. For the RM types which are not capable of doing so,
581 invoking this method does not have any consequences.
583 :param guid: Guid of the RM
587 rm = self.get_resource(guid)
588 return rm.provision()
590 def get(self, guid, name):
591 """ Returns the value of the attribute with name 'name' on the
594 :param guid: Guid of the RM
597 :param name: Name of the attribute
600 :return: The value of the attribute with name 'name'
603 rm = self.get_resource(guid)
606 def set(self, guid, name, value):
607 """ Modifies the value of the attribute with name 'name' on the
610 :param guid: Guid of the RM
613 :param name: Name of the attribute
616 :param value: Value of the attribute
619 rm = self.get_resource(guid)
622 def get_global(self, rtype, name):
623 """ Returns the value of the global attribute with name 'name' on the
624 RMs of rtype 'rtype'.
626 :param guid: Guid of the RM
629 :param name: Name of the attribute
632 :return: The value of the attribute with name 'name'
635 rclass = ResourceFactory.get_resource_type(rtype)
636 return rclass.get_global(name)
638 def set_global(self, rtype, name, value):
639 """ Modifies the value of the global attribute with name 'name' on the
640 RMs of with rtype 'rtype'.
642 :param guid: Guid of the RM
645 :param name: Name of the attribute
648 :param value: Value of the attribute
651 rclass = ResourceFactory.get_resource_type(rtype)
652 return rclass.set_global(name, value)
654 def state(self, guid, hr = False):
655 """ Returns the state of a resource
657 :param guid: Resource guid
660 :param hr: Human readable. Forces return of a
661 status string instead of a number
665 rm = self.get_resource(guid)
669 return ResourceState2str.get(state)
673 def stop(self, guid):
674 """ Stops the RM with guid 'guid'
676 Stopping a RM means that the resource it controls will
677 no longer take part of the experiment.
679 :param guid: Guid of the RM
683 rm = self.get_resource(guid)
686 def start(self, guid):
687 """ Starts the RM with guid 'guid'
689 Starting a RM means that the resource it controls will
690 begin taking part of the experiment.
692 :param guid: Guid of the RM
696 rm = self.get_resource(guid)
699 def set_with_conditions(self, name, value, guids1, guids2, state,
701 """ Modifies the value of attribute with name 'name' on all RMs
702 on the guids1 list when time 'time' has elapsed since all
703 elements in guids2 list have reached state 'state'.
705 :param name: Name of attribute to set in RM
708 :param value: Value of attribute to set in RM
711 :param guids1: List of guids of RMs subjected to action
714 :param action: Action to register (either START or STOP)
715 :type action: ResourceAction
717 :param guids2: List of guids of RMs to we waited for
720 :param state: State to wait for on RMs (STARTED, STOPPED, etc)
721 :type state: ResourceState
723 :param time: Time to wait after guids2 has reached status
727 if isinstance(guids1, int):
729 if isinstance(guids2, int):
733 rm = self.get_resource(guid)
734 rm.set_with_conditions(name, value, guids2, state, time)
736 def deploy(self, guids = None, wait_all_ready = True, group = None):
737 """ Deploys all ResourceManagers in the guids list.
739 If the argument 'guids' is not given, all RMs with state NEW
742 :param guids: List of guids of RMs to deploy
745 :param wait_all_ready: Wait until all RMs are ready in
746 order to start the RMs
749 :param group: Id of deployment group in which to deploy RMs
753 self.logger.debug(" ------- DEPLOY START ------ ")
756 # If no guids list was passed, all 'NEW' RMs will be deployed
758 for guid in self.resources:
759 if self.state(guid) == ResourceState.NEW:
762 if isinstance(guids, int):
765 # Create deployment group
766 # New guids can be added to a same deployment group later on
770 group = self._group_id_generator.next()
772 if group not in self._groups:
773 self._groups[group] = []
775 self._groups[group].extend(guids)
777 def wait_all_and_start(group):
778 # Function that checks if all resources are READY
779 # before scheduling a start_with_conditions for each RM
782 # Get all guids in group
783 guids = self._groups[group]
786 if self.state(guid) < ResourceState.READY:
791 callback = functools.partial(wait_all_and_start, group)
792 self.schedule("1s", callback)
794 # If all resources are ready, we schedule the start
796 rm = self.get_resource(guid)
797 self.schedule("0s", rm.start_with_conditions)
799 if rm.conditions.get(ResourceAction.STOP):
800 # Only if the RM has STOP conditions we
801 # schedule a stop. Otherwise the RM will stop immediately
802 self.schedule("0s", rm.stop_with_conditions)
804 if wait_all_ready and new_group:
805 # Schedule a function to check that all resources are
806 # READY, and only then schedule the start.
807 # This aims at reducing the number of tasks looping in the
809 # Instead of having many start tasks, we will have only one for
811 callback = functools.partial(wait_all_and_start, group)
812 self.schedule("0s", callback)
815 rm = self.get_resource(guid)
816 rm.deployment_group = group
817 self.schedule("0s", rm.deploy_with_conditions)
819 if not wait_all_ready:
820 self.schedule("0s", rm.start_with_conditions)
822 if rm.conditions.get(ResourceAction.STOP):
823 # Only if the RM has STOP conditions we
824 # schedule a stop. Otherwise the RM will stop immediately
825 self.schedule("0s", rm.stop_with_conditions)
827 def release(self, guids = None):
828 """ Releases all ResourceManagers in the guids list.
830 If the argument 'guids' is not given, all RMs registered
831 in the experiment are released.
833 :param guids: List of RM guids
838 guids = self.resources
840 # Remove all pending tasks from the scheduler queue
841 for tid in list(self._scheduler.pending):
842 self._scheduler.remove(tid)
847 rm = self.get_resource(guid)
848 self.schedule("0s", rm.release)
850 self.wait_released(guids)
853 """ Releases all resources and stops the ExperimentController
856 # If there was a major failure we can't exit gracefully
857 if self._state == ECState.FAILED:
858 raise RuntimeError("EC failure. Can not exit gracefully")
862 # Mark the EC state as TERMINATED
863 self._state = ECState.TERMINATED
865 # Stop processing thread
868 # Notify condition to wake up the processing thread
871 if self._thread.is_alive():
874 def schedule(self, date, callback, track = False):
875 """ Schedules a callback to be executed at time 'date'.
877 :param date: string containing execution time for the task.
878 It can be expressed as an absolute time, using
879 timestamp format, or as a relative time matching
880 ^\d+.\d+(h|m|s|ms|us)$
882 :param callback: code to be executed for the task. Must be a
883 Python function, and receives args and kwargs
886 :param track: if set to True, the task will be retrievable with
887 the get_task() method
889 :return : The Id of the task
893 timestamp = stabsformat(date)
894 task = Task(timestamp, callback)
895 task = self._scheduler.schedule(task)
898 self._tasks[task.id] = task
900 # Notify condition to wake up the processing thread
906 """ Process scheduled tasks.
910 Tasks are scheduled by invoking the schedule method with a target
911 callback and an execution time.
912 The schedule method creates a new Task object with that callback
913 and execution time, and pushes it into the '_scheduler' queue.
914 The execution time and the order of arrival of tasks are used
915 to order the tasks in the queue.
917 The _process method is executed in an independent thread held by
918 the ExperimentController for as long as the experiment is running.
919 This method takes tasks from the '_scheduler' queue in a loop
920 and processes them in parallel using multithreading.
921 The environmental variable NEPI_NTHREADS can be used to control
922 the number of threads used to process tasks. The default value is
925 To execute tasks in parallel, a ParallelRunner (PR) object is used.
926 This object keeps a pool of threads (workers), and a queue of tasks
927 scheduled for 'immediate' execution.
929 On each iteration, the '_process' loop will take the next task that
930 is scheduled for 'future' execution from the '_scheduler' queue,
931 and if the execution time of that task is >= to the current time,
932 it will push that task into the PR for 'immediate execution'.
933 As soon as a worker is free, the PR will assign the next task to
936 Upon receiving a task to execute, each PR worker (thread) will
937 invoke the _execute method of the EC, passing the task as
939 The _execute method will then invoke task.callback inside a
940 try/except block. If an exception is raised by the tasks.callback,
941 it will be trapped by the try block, logged to standard error
942 (usually the console), and the task will be marked as failed.
948 while not self._stop:
952 task = self._scheduler.next()
955 # No task to execute. Wait for a new task to be scheduled.
958 # The task timestamp is in the future. Wait for timeout
959 # or until another task is scheduled.
961 if now < task.timestamp:
962 # Calculate timeout in seconds
963 timeout = tdiffsec(task.timestamp, now)
965 # Re-schedule task with the same timestamp
966 self._scheduler.schedule(task)
970 # Wait timeout or until a new task awakes the condition
971 self._cond.wait(timeout)
976 # Process tasks in parallel
977 self._runner.put(self._execute, task)
980 err = traceback.format_exc()
981 self.logger.error("Error while processing tasks in the EC: %s" % err)
983 # Set the EC to FAILED state
984 self._state = ECState.FAILED
986 # Set the FailureManager failure level to EC failure
987 self._fm.set_ec_failure()
989 self.logger.debug("Exiting the task processing loop ... ")
992 self._runner.destroy()
994 def _execute(self, task):
995 """ Executes a single task.
997 :param task: Object containing the callback to execute
1003 task.result = task.callback()
1004 task.status = TaskStatus.DONE
1007 err = traceback.format_exc()
1009 task.status = TaskStatus.ERROR
1011 self.logger.error("Error occurred while executing task: %s" % err)
1014 """ Awakes the processing thread if it is blocked waiting
1015 for new tasks to arrive
1018 self._cond.acquire()
1020 self._cond.release()