Haizea

# -------------------------------------------------------------------------- #

# Copyright 2006-2009, University of Chicago                                 #

# Copyright 2008-2009, Distributed Systems Architecture Group, Universidad   #

# Complutense de Madrid (dsa-research.org)                                   #

#                                                                            #

# Licensed under the Apache License, Version 2.0 (the "License"); you may    #

# not use this file except in compliance with the License. You may obtain    #

# a copy of the License at                                                   #

#                                                                            #

# http://www.apache.org/licenses/LICENSE-2.0                                 #

#                                                                            #

# Unless required by applicable law or agreed to in writing, software        #

# distributed under the License is distributed on an "AS IS" BASIS,          #

# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.   #

# See the License for the specific language governing permissions and        #

# limitations under the License.                                             #

# -------------------------------------------------------------------------- #

"""This module provides the main classes for Haizea's VM Scheduler. All the

scheduling code that decides when and where a lease is scheduled is contained

in the VMScheduler class (except for the code that specifically decides

what physical machines each virtual machine is mapped to, which is factored out

into the "mapper" module). This module also provides the classes for the

reservations that will be placed in the slot table and correspond to VMs.

"""

import haizea.common.constants as constants

from haizea.common.utils import round_datetime_delta, round_datetime, estimate_transfer_time, pretty_nodemap, get_config, get_clock, get_persistence, compute_suspend_resume_time

from haizea.core.leases import Lease, Capacity

from haizea.core.scheduler.slottable import ResourceReservation, ResourceTuple

from haizea.core.scheduler import ReservationEventHandler, RescheduleLeaseException, NormalEndLeaseException, EnactmentError, NotSchedulableException, InconsistentScheduleError, InconsistentLeaseStateError, MigrationResourceReservation

from operator import attrgetter, itemgetter

from mx.DateTime import TimeDelta

import logging

import operator

class VMScheduler(object):

    """The Haizea VM Scheduler

    This class is responsible for taking a lease and scheduling VMs to satisfy

    the requirements of that lease.

    """

    def __init__(self, slottable, resourcepool, mapper, max_in_future):

        """Constructor

        The constructor does little more than create the VM scheduler's

        attributes. However, it does expect (in the arguments) a fully-constructed

        SlotTable, ResourcePool, and Mapper (these are constructed in the

        Manager's constructor).

        Arguments:

        slottable -- Slot table

        resourcepool -- Resource pool where enactment commands will be sent to

        mapper -- Mapper

        """

        self.slottable = slottable

        self.resourcepool = resourcepool

        self.mapper = mapper

        self.logger = logging.getLogger("VMSCHED")

        # Register the handlers for the types of reservations used by

        # the VM scheduler

        self.handlers = {}

        self.handlers[VMResourceReservation] = ReservationEventHandler(

                                sched    = self,

                                on_start = VMScheduler._handle_start_vm,

                                on_end   = VMScheduler._handle_end_vm)

        self.handlers[ShutdownResourceReservation] = ReservationEventHandler(

                                sched    = self,

                                on_start = VMScheduler._handle_start_shutdown,

                                on_end   = VMScheduler._handle_end_shutdown)

        self.handlers[SuspensionResourceReservation] = ReservationEventHandler(

                                sched    = self,

                                on_start = VMScheduler._handle_start_suspend,

                                on_end   = VMScheduler._handle_end_suspend)

        self.handlers[ResumptionResourceReservation] = ReservationEventHandler(

                                sched    = self,

                                on_start = VMScheduler._handle_start_resume,

                                on_end   = VMScheduler._handle_end_resume)

        self.handlers[MemImageMigrationResourceReservation] = ReservationEventHandler(

                                sched    = self,

                                on_start = VMScheduler._handle_start_migrate,

                                on_end   = VMScheduler._handle_end_migrate)

        self.max_in_future = max_in_future

        self.future_leases = set()

    def schedule(self, lease, duration, nexttime, earliest, override_state = None):

        """ The scheduling function

        This particular function doesn't do much except call __schedule_asap

        and __schedule_exact (which do all the work).

        Arguments:

        lease -- Lease to schedule

        nexttime -- The next time at which the scheduler can allocate resources.

        earliest -- The earliest possible starting times on each physical node

        """

        if lease.get_type() == Lease.BEST_EFFORT:

            return self.__schedule_asap(lease, duration, nexttime, earliest, allow_in_future = self.can_schedule_in_future())

        elif lease.get_type() == Lease.ADVANCE_RESERVATION:

            return self.__schedule_exact(lease, duration, nexttime, earliest)

        elif lease.get_type() == Lease.IMMEDIATE:

            return self.__schedule_asap(lease, duration, nexttime, earliest, allow_in_future = False)

        elif lease.get_type() == Lease.DEADLINE:

            return self.__schedule_deadline(lease, duration, nexttime, earliest, override_state)

    def estimate_migration_time(self, lease):

        """ Estimates the time required to migrate a lease's VMs

        This function conservatively estimates that all the VMs are going to

        be migrated to other nodes. Since all the transfers are intra-node,

        the bottleneck is the transfer from whatever node has the most

        memory to transfer.

        Note that this method only estimates the time to migrate the memory

        state files for the VMs. Migrating the software environment (which may

        or may not be a disk image) is the responsibility of the preparation

        scheduler, which has it's own set of migration scheduling methods.

        Arguments:

        lease -- Lease that might be migrated

        """

        migration = get_config().get("migration")

        if migration == constants.MIGRATE_YES:

            bandwidth = self.resourcepool.info.get_migration_bandwidth()

            vmrr = lease.get_last_vmrr()

            #mem_in_pnode = dict([(pnode,0) for pnode in set(vmrr.nodes.values())])

            transfer_time = 0

            for pnode in vmrr.nodes.values():

                mem = vmrr.resources_in_pnode[pnode].get_by_type(constants.RES_MEM)

                transfer_time += estimate_transfer_time(mem, bandwidth)

            return transfer_time

        elif migration == constants.MIGRATE_YES_NOTRANSFER:

            return TimeDelta(seconds=0)

    def schedule_migration(self, lease, vmrr, nexttime):

        """ Schedules migrations for a lease

        Arguments:

        lease -- Lease being migrated

        vmrr -- The VM reservation before which the migration will take place

        nexttime -- The next time at which the scheduler can allocate resources.

        """

        # Determine what migrations have to be done.

        last_vmrr = lease.get_last_vmrr()

        vnode_mappings = self.resourcepool.get_ram_image_mappings(lease)

        vnode_migrations = dict([(vnode, (vnode_mappings[vnode], vmrr.nodes[vnode])) for vnode in vmrr.nodes if vnode_mappings[vnode] != vmrr.nodes[vnode]])

        # Determine if we actually have to migrate

        mustmigrate = False

        for vnode in vnode_migrations:

            if vnode_migrations[vnode][0] != vnode_migrations[vnode][1]:

                mustmigrate = True

                break

        if not mustmigrate:

            return []

        # If Haizea is configured to migrate without doing any transfers,

        # then we just return a nil-duration migration RR

        if get_config().get("migration") == constants.MIGRATE_YES_NOTRANSFER:

            start = nexttime

            end = nexttime

            res = {}

            migr_rr = MemImageMigrationResourceReservation(lease, start, end, res, vmrr, vnode_migrations)

            migr_rr.state = ResourceReservation.STATE_SCHEDULED

            return [migr_rr]

        # Figure out what migrations can be done simultaneously

        migrations = []

        while len(vnode_migrations) > 0:

            pnodes = set()

            migration = {}

            for vnode in vnode_migrations:

                origin = vnode_migrations[vnode][0]

                dest = vnode_migrations[vnode][1]

                if not origin in pnodes and not dest in pnodes:

                    migration[vnode] = vnode_migrations[vnode]

                    pnodes.add(origin)

                    pnodes.add(dest)

            for vnode in migration:

                del vnode_migrations[vnode]

            migrations.append(migration)

        # Create migration RRs

        start = max(last_vmrr.post_rrs[-1].end, nexttime)

        bandwidth = self.resourcepool.info.get_migration_bandwidth()

        migr_rrs = []

        for m in migrations:

            vnodes_to_migrate = m.keys()

            max_mem_to_migrate = max([lease.requested_resources[vnode].get_quantity(constants.RES_MEM) for vnode in vnodes_to_migrate])

            migr_time = estimate_transfer_time(max_mem_to_migrate, bandwidth)

            end = start + migr_time

            res = {}

            for (origin,dest) in m.values():

                resorigin = Capacity([constants.RES_NETOUT])

                resorigin.set_quantity(constants.RES_NETOUT, bandwidth)

                resdest = Capacity([constants.RES_NETIN])

                resdest.set_quantity(constants.RES_NETIN, bandwidth)

                res[origin] = self.slottable.create_resource_tuple_from_capacity(resorigin)

                res[dest] = self.slottable.create_resource_tuple_from_capacity(resdest)

            migr_rr = MemImageMigrationResourceReservation(lease, start, start + migr_time, res, vmrr, m)

            migr_rr.state = ResourceReservation.STATE_SCHEDULED

            migr_rrs.append(migr_rr)

            start = end

        return migr_rrs

    def cancel_vm(self, vmrr):

        """ Cancels a VM resource reservation

        Arguments:

        vmrr -- VM RR to be cancelled

        """

        # If this VM RR is part of a lease that was scheduled in the future,

        # remove that lease from the set of future leases.

        if vmrr.lease in self.future_leases:

            self.future_leases.remove(vmrr.lease)

            get_persistence().persist_future_leases(self.future_leases)

        # If there are any pre-RRs that are scheduled or active, remove them

        for rr in vmrr.pre_rrs:

            if rr.state != ResourceReservation.STATE_DONE:

                self.slottable.remove_reservation(rr)

        # If there are any post RRs, remove them

        for rr in vmrr.post_rrs:

            self.slottable.remove_reservation(rr)

        # Remove the reservation itself

        self.slottable.remove_reservation(vmrr)

    def can_suspend_at(self, lease, t, nexttime=None):

        """ Determines if it is possible to suspend a lease before a given time

        Arguments:

        vmrr -- VM RR to be preempted

        t -- Time by which the VM must be preempted

        """

        vmrr = lease.get_vmrr_at(t)

        if t < vmrr.start:

            return False # t could be during a resume

        by_t = min(vmrr.end, t) # t could be during a suspend

        if nexttime == None:

            time_until_suspend = t - vmrr.start

        else:

            time_until_suspend = min( (t - vmrr.start, t - nexttime))

        min_duration = self.__compute_scheduling_threshold(lease)

        can_suspend = time_until_suspend >= min_duration

        return can_suspend

    def preempt_vm(self, vmrr, t):

        """ Preempts a VM reservation at a given time

        This method assumes that the lease is, in fact, preemptable,

        that the VMs are running at the given time, and that there is

        enough time to suspend the VMs before the given time (all these

        checks are done in the lease scheduler).

        Arguments:

        vmrr -- VM RR to be preempted

        t -- Time by which the VM must be preempted

        """

        # Save original start and end time of the vmrr

        old_start = vmrr.start

        old_end = vmrr.end

        # Schedule the VM suspension

        self.__schedule_suspension(vmrr, t)

        # Update the VMRR in the slot table

        self.slottable.update_reservation(vmrr, old_start, old_end)

        # Add the suspension RRs to the VM's post-RRs

        for susprr in vmrr.post_rrs:

            self.slottable.add_reservation(susprr)

    def get_future_reschedulable_leases(self):

        """ Returns a list of future leases that are reschedulable.

        Currently, this list is just the best-effort leases scheduled

        in the future as determined by the backfilling algorithm.

        Advance reservation leases, by their nature, cannot be

        rescheduled to find a "better" starting time.

        """

        return list(self.future_leases)

    def can_schedule_in_future(self):

        """ Returns True if the backfilling algorithm would allow a lease

        to be scheduled in the future.

        """

        if self.max_in_future == -1: # Unlimited

            return True

        else:

            return len(self.future_leases) < self.max_in_future

    def get_utilization(self, time):

        """ Computes resource utilization (currently just CPU-based)

        This utilization information shows what

        portion of the physical resources is used by each type of reservation

        (e.g., 70% are running a VM, 5% are doing suspensions, etc.)

        Arguments:

        time -- Time at which to determine utilization

        """

        total = self.slottable.get_total_capacity(restype = constants.RES_CPU)

        util = {}

        reservations = self.slottable.get_reservations_at(time)

        for r in reservations:

            for node in r.resources_in_pnode:

                if isinstance(r, VMResourceReservation):

                    use = r.resources_in_pnode[node].get_by_type(constants.RES_CPU)

                    util[type(r)] = use + util.setdefault(type(r),0.0)

                elif isinstance(r, SuspensionResourceReservation) or isinstance(r, ResumptionResourceReservation) or isinstance(r, ShutdownResourceReservation):

                    use = r.vmrr.resources_in_pnode[node].get_by_type(constants.RES_CPU)

                    util[type(r)] = 0.0

                    #util[type(r)] = use + util.setdefault(type(r),0.0)

        util[None] = total - sum(util.values())

        if total != 0:

            for k in util:

                util[k] /= total

        return util

    def __schedule_exact(self, lease, duration, nexttime, earliest):

        """ Schedules VMs that must start at an exact time

        This type of lease is "easy" to schedule because we know the exact

        start time, which means that's the only starting time we have to

        check. So, this method does little more than call the mapper.

        Arguments:

        lease -- Lease to schedule

        nexttime -- The next time at which the scheduler can allocate resources.

        earliest -- The earliest possible starting times on each physical node

        """

        # Determine the start and end time

        shutdown_time = lease.estimate_shutdown_time()

        start = lease.start.requested

        end = start + lease.duration.requested + shutdown_time

        # Convert Capacity objects in lease object into ResourceTuples that

        # we can hand over to the mapper.

        requested_resources = dict([(k,self.slottable.create_resource_tuple_from_capacity(v)) for k,v in lease.requested_resources.items()])

        # Let the mapper do its magiv

        mapping, actualend, preemptions = self.mapper.map(lease,

                                                          requested_resources,

                                                          start,

                                                          end,

                                                          strictend = True,

                                                          allow_preemption = True)

        # If no mapping was found, tell the lease scheduler about it

        if mapping == None:

            raise NotSchedulableException, "Not enough resources in specified interval"

        # Create VM resource reservations

        res = {}

        for (vnode,pnode) in mapping.items():

            vnode_res = requested_resources[vnode]

            if res.has_key(pnode):

                res[pnode].incr(vnode_res)

            else:

                res[pnode] = ResourceTuple.copy(vnode_res)

        vmrr = VMResourceReservation(lease, start, end, mapping, res)

        vmrr.state = ResourceReservation.STATE_SCHEDULED

        # Schedule shutdown for the VM

        self.__schedule_shutdown(vmrr)

        return vmrr, preemptions

    def __schedule_asap(self, lease, duration, nexttime, earliest, allow_in_future = None, override_state = None):

        """ Schedules VMs as soon as possible

        This method is a bit more complex that __schedule_exact because

        we need to figure out what "as soon as possible" actually is.

        This involves attempting several mappings, at different points

        in time, before we can schedule the lease.

        This method will always check, at least, if the lease can be scheduled

        at the earliest possible moment at which the lease could be prepared

        (e.g., if the lease can't start until 1 hour in the future because that's

        the earliest possible time at which the disk images it requires can

        be transferred, then that's when the scheduler will check). Note, however,

        that this "earliest possible moment" is determined by the preparation

        scheduler.

        Additionally, if the lease can't be scheduled at the earliest

        possible moment, it can also check if the lease can be scheduled

        in the future. This partially implements a backfilling algorithm

        (the maximum number of future leases is stored in the max_in_future

        attribute of VMScheduler), the other part being implemented in the

        __process_queue method of LeaseScheduler.

        Note that, if the method is allowed to scheduled in the future,

        and assuming that the lease doesn't request more resources than

        the site itself, this method will always schedule the VMs succesfully

        (since there's always an empty spot somewhere in the future).

        Arguments:

        lease -- Lease to schedule

        nexttime -- The next time at which the scheduler can allocate resources.

        earliest -- The earliest possible starting times on each physical node

        allow_in_future -- Boolean indicating whether the scheduler is

        allowed to schedule the VMs in the future.

        """

        #

        # STEP 1: PROLEGOMENA

        #

        lease_id = lease.id

        remaining_duration = duration

        shutdown_time = lease.estimate_shutdown_time()

        if override_state != None:

            state = override_state

        else:

            state = lease.get_state()

        # We might be scheduling a suspended lease. If so, we will

        # also have to schedule its resumption. Right now, just

        # figure out if this is such a lease.

        mustresume = (state in (Lease.STATE_SUSPENDED_PENDING, Lease.STATE_SUSPENDED_QUEUED, Lease.STATE_SUSPENDED_SCHEDULED))

        # This is the minimum duration that we must be able to schedule.

        # See __compute_scheduling_threshold for more details.

        min_duration = self.__compute_scheduling_threshold(lease)

        #

        # STEP 2: FIND THE CHANGEPOINTS

        #

        # Find the changepoints, and the available nodes at each changepoint

        # We need to do this because the preparation scheduler may have

        # determined that some nodes might require more time to prepare

        # than others (e.g., if using disk image caching, some nodes

        # might have the required disk image predeployed, while others

        # may require transferring the image to that node).

        #

        # The end result of this step is a list (cps) where each entry

        # is a (t,nodes) pair, where "t" is the time of the changepoint

        # and "nodes" is the set of nodes that are available at that time.

        if not mustresume:

            # If this is not a suspended lease, then the changepoints

            # are determined based on the "earliest" parameter.

            cps = [(node, e.time) for node, e in earliest.items()]

            cps.sort(key=itemgetter(1))

            curcp = None

            changepoints = []

            nodes = []

            for node, time in cps:

                nodes.append(node)

                if time != curcp:

                    changepoints.append([time, set(nodes)])

                    curcp = time

                else:

                    changepoints[-1][1] = set(nodes)

        else:

            # If the lease is suspended, we take into account that, if

            # migration is disabled, we can only schedule the lease

            # on the nodes it is currently scheduled on.

            if get_config().get("migration") == constants.MIGRATE_NO:

                vmrr = lease.get_last_vmrr()

                onlynodes = set(vmrr.nodes.values())

            else:

                onlynodes = None

            changepoints = list(set([x.time for x in earliest.values()]))

            changepoints.sort()

            changepoints = [(x, onlynodes) for x in changepoints]

        # If we can schedule VMs in the future,

        # we also consider future changepoints

        if allow_in_future:

            res = self.slottable.get_reservations_ending_after(changepoints[-1][0])

            # We really only care about changepoints where VMs end (which is

            # when resources become available)

            futurecp = [r.get_final_end() for r in res if isinstance(r, VMResourceReservation)]

            # Corner case: Sometimes we're right in the middle of a ShutdownReservation, so it won't be

            # included in futurecp.

            futurecp += [r.end for r in res if isinstance(r, ShutdownResourceReservation) and not r.vmrr in res]

            if not mustresume:

                futurecp = [(p,None) for p in futurecp]

            else:

                futurecp = [(p,onlynodes) for p in futurecp]

        else:

            futurecp = []

        futurecp.sort()

        if lease.deadline != None:

            changepoints = [cp for cp in changepoints if cp[0] <= lease.deadline - duration]

            futurecp = [cp for cp in futurecp if cp[0] <= lease.deadline - duration]

        #

        # STEP 3: FIND A MAPPING

        #

        # In this step we find a starting time and a mapping for the VMs,

        # which involves going through the changepoints in order and seeing

        # if we can find a mapping.

        # Most of the work is done in the __find_fit_at_points

        # If resuming, we also have to allocate enough time for the resumption

        if mustresume:

            duration = remaining_duration + lease.estimate_resume_time()

        else:

            duration = remaining_duration

        duration += shutdown_time

        in_future = False

        # Convert Capacity objects in lease object into ResourceTuples that

        # we can hand over to the mapper.

        requested_resources = dict([(k,self.slottable.create_resource_tuple_from_capacity(v)) for k,v in lease.requested_resources.items()])

        # First, try to find a mapping assuming we can't schedule in the future

        start, end, mapping, preemptions = self.__find_fit_at_points(lease,

                                                                     requested_resources,

                                                                     changepoints,

                                                                     duration,

                                                                     min_duration,

                                                                     shutdown_time)

        if start == None and not allow_in_future:

            # We did not find a suitable starting time. This can happen

            # if we're unable to schedule in the future

            raise NotSchedulableException, "Could not find enough resources for this request"

        # If we haven't been able to fit the lease, check if we can

        # reserve it in the future

        if start == None and allow_in_future:

            start, end, mapping, preemptions = self.__find_fit_at_points(lease,

                                                                         requested_resources,

                                                                         futurecp,

                                                                         duration,

                                                                         min_duration,

                                                                         shutdown_time

                                                                         )

            # TODO: The following will also raise an exception if a lease

            # makes a request that could *never* be satisfied with the

            # current resources.

            if start == None:

                if lease.deadline != None:

                    raise NotSchedulableException, "Could not find enough resources for this request before deadline"

                else:

                    raise InconsistentScheduleError, "Could not find a mapping in the future (this should not happen)"

            in_future = True

        #

        # STEP 4: CREATE RESERVATIONS

        #

        # At this point, the lease is feasible. We just need to create

        # the reservations for the VMs and, possibly, for the VM resumption,

        # suspension, and shutdown.

        # VM resource reservation

        res = {}

        for (vnode,pnode) in mapping.items():

            vnode_res = requested_resources[vnode]

            if res.has_key(pnode):

                res[pnode].incr(vnode_res)

            else:

                res[pnode] = ResourceTuple.copy(vnode_res)

        vmrr = VMResourceReservation(lease, start, end, mapping, res)

        vmrr.state = ResourceReservation.STATE_SCHEDULED

        # VM resumption resource reservation

        if mustresume:

            self.__schedule_resumption(vmrr, start)

        # If the mapper couldn't find a mapping for the full duration

        # of the lease, then we need to schedule a suspension.

        mustsuspend = (vmrr.end - vmrr.start) - shutdown_time < remaining_duration

        if mustsuspend:

            self.__schedule_suspension(vmrr, end)

        else:

            # Compensate for any overestimation

            if (vmrr.end - vmrr.start) > remaining_duration + shutdown_time:

                vmrr.end = vmrr.start + remaining_duration + shutdown_time

            self.__schedule_shutdown(vmrr)

        if in_future:

            self.future_leases.add(lease)

            get_persistence().persist_future_leases(self.future_leases)

        susp_str = res_str = ""

        if mustresume:

            res_str = " (resuming)"

        if mustsuspend:

            susp_str = " (suspending)"

        self.logger.info("Lease #%i can be scheduled on nodes %s from %s%s to %s%s" % (lease.id, mapping.values(), start, res_str, end, susp_str))

        return vmrr, preemptions

    def __schedule_deadline(self, lease, duration, nexttime, earliest, override_state):

        earliest_time = nexttime

        for n in earliest:

            earliest[n].time = max(lease.start.requested, earliest[n].time)

            earliest_time = max(earliest_time, earliest[n].time)

        slack = (lease.deadline - lease.start.requested) / lease.duration.requested

        if slack <= 2.0:

            try:

                self.logger.debug("Trying to schedule lease #%i as an advance reservation..." % lease.id)

                vmrr, preemptions = self.__schedule_exact(lease, duration, nexttime, earliest)

                # Don't return preemptions. They have already been preempted by the deadline mapper

                return vmrr, []

            except NotSchedulableException:

                self.logger.debug("Lease #%i cannot be scheduled as an advance reservation, trying as best-effort..." % lease.id)

                try:

                    vmrr, preemptions = self.__schedule_asap(lease, duration, nexttime, earliest, allow_in_future = True, override_state=override_state)

                except NotSchedulableException:

                    vmrr = None

                    preemptions = []

                if vmrr == None or vmrr.end - vmrr.start != duration or vmrr.end > lease.deadline or len(preemptions)>0:

                    self.logger.debug("Lease #%i cannot be scheduled before deadline using best-effort." % lease.id)

                    #raise NotSchedulableException, "Could not schedule before deadline without making other leases miss deadline"

                else:

                    return vmrr, preemptions

        else:

            self.logger.debug("Trying to schedule lease #%i as best-effort..." % lease.id)

            try:

                vmrr, preemptions = self.__schedule_asap(lease, duration, nexttime, earliest, allow_in_future = True, override_state=override_state)

            except NotSchedulableException:

                vmrr = None

                preemptions = []

        if vmrr == None or vmrr.end - vmrr.start != duration or vmrr.end > lease.deadline or len(preemptions)>0:

            self.logger.debug("Trying to schedule lease #%i by rescheduling other leases..." % lease.id)

            dirtynodes = set()

            dirtytime = earliest_time

            future_vmrrs = self.slottable.get_reservations_on_or_after(earliest_time)

            future_vmrrs.sort(key=operator.attrgetter("start"))

            future_vmrrs = [rr for rr in future_vmrrs

                            if isinstance(rr, VMResourceReservation)

                            and rr.state == ResourceReservation.STATE_SCHEDULED

                            and reduce(operator.and_, [(prerr.state == ResourceReservation.STATE_SCHEDULED) for prerr in rr.pre_rrs], True)]

            leases = list(set([future_vmrr.lease for future_vmrr in future_vmrrs]))

            self.slottable.push_state(leases)

            for future_vmrr in future_vmrrs:

                #print "REMOVE", future_vmrr.lease.id, future_vmrr.start, future_vmrr.end

                future_vmrr.lease.remove_vmrr(future_vmrr)

                self.cancel_vm(future_vmrr)

            orig_vmrrs = dict([(l,[rr for rr in future_vmrrs if rr.lease == l]) for l in leases])

            leases.append(lease)

            leases.sort(key= lambda l: (l.deadline - earliest_time) / l.get_remaining_duration_at(nexttime))

            new_vmrrs = {}

            self.logger.debug("Attempting to reschedule leases %s" % [l.id for l in leases])

            # First pass

            scheduled = set()

            for lease2 in leases:

                last_vmrr = lease2.get_last_vmrr()

                if last_vmrr != None and last_vmrr.is_suspending():

                    override_state = Lease.STATE_SUSPENDED_PENDING

                    l_earliest_time = max(last_vmrr.post_rrs[-1].end, earliest_time)

                else:

                    override_state = None

                    l_earliest_time = earliest_time

                for n in earliest:

                    earliest[n].time = max(lease2.start.requested, l_earliest_time)

                self.logger.debug("Rescheduling lease %s" % lease2.id)

                dur = lease2.get_remaining_duration_at(l_earliest_time)

                try:

                    vmrr, preemptions = self.__schedule_asap(lease2, dur, nexttime, earliest, allow_in_future = True, override_state=override_state)

                except NotSchedulableException:

                    vmrr = None

                    preemptions = []

                if vmrr == None or vmrr.end - vmrr.start != dur or vmrr.end > lease2.deadline or len(preemptions) != 0:

                    self.logger.debug("Lease %s could not be rescheduled, undoing changes." % lease2.id)

                    self.slottable.pop_state()

                    raise NotSchedulableException, "Could not schedule before deadline without making other leases miss deadline"

                dirtytime = max(vmrr.end, dirtytime)

                dirtynodes.update(vmrr.resources_in_pnode.keys())

                for rr in vmrr.pre_rrs:

                    self.slottable.add_reservation(rr)

                self.slottable.add_reservation(vmrr)

                for rr in vmrr.post_rrs:

                    self.slottable.add_reservation(rr)

                scheduled.add(lease2)

                if lease2 == lease:

                    return_vmrr = vmrr

                    break

                else:

                    new_vmrrs[lease2] = vmrr

            # We've scheduled the lease. Now we try to schedule the rest of the leases but,

            # since now we know the nodes the new lease is in, we can do a few optimizations

            # Restore the leases in nodes we haven't used, and that would not be

            # affected by the new lease. We need to find what this set of nodes is.

            to_schedule = [l for l in leases if l not in scheduled]

            dirtynodes, cleanleases = self.find_dirty_nodes(to_schedule, dirtynodes, orig_vmrrs)

            self.logger.debug("Rescheduling only leases on nodes %s" % dirtynodes)

            self.logger.debug("Leases %s can be skipped" % [l.id for l in cleanleases])

            # Restore the leases

            restored_leases = set()

            for l in leases:

                if l in cleanleases:

                    # Restore

                    for l_vmrr in orig_vmrrs[l]:

                        for rr in l_vmrr.pre_rrs:

                            self.slottable.add_reservation(rr)

                        self.slottable.add_reservation(l_vmrr)

                        for rr in l_vmrr.post_rrs:

                            self.slottable.add_reservation(rr)

                        l.append_vmrr(l_vmrr)

                        scheduled.add(l)

                        restored_leases.add(l)

            to_schedule = [l for l in leases if l not in scheduled]

            try:

                (more_scheduled, add_vmrrs, dirtytime) = self.reschedule_deadline_leases(to_schedule, orig_vmrrs, earliest_time, earliest, nexttime, dirtytime)

                scheduled.update(more_scheduled)

                new_vmrrs.update(add_vmrrs)

            except NotSchedulableException:

                self.logger.debug("Lease %s could not be rescheduled, undoing changes." % l.id)

                self.slottable.pop_state()

                raise

            self.slottable.pop_state(discard=True)

            for l in leases:

                if l not in scheduled:

                    for l_vmrr in orig_vmrrs[l]:

                        for rr in l_vmrr.pre_rrs:

                            self.slottable.add_reservation(rr)

                        self.slottable.add_reservation(l_vmrr)

                        for rr in l_vmrr.post_rrs:

                            self.slottable.add_reservation(rr)

                        l.append_vmrr(l_vmrr)

                        restored_leases.add(l)

            for lease2, vmrr in new_vmrrs.items():

                lease2.append_vmrr(vmrr)

            # Remove from slottable, because lease_scheduler is the one that actually

            # adds the RRs

            for rr in return_vmrr.pre_rrs:

                self.slottable.remove_reservation(rr)

            self.slottable.remove_reservation(return_vmrr)

            for rr in return_vmrr.post_rrs:

                self.slottable.remove_reservation(rr)

            for l in leases:

                if l in scheduled:

                    self.logger.vdebug("Lease %i after rescheduling:" % l.id)

                    l.print_contents()

            return return_vmrr, []

        else:

            return vmrr, preemptions

    def find_dirty_nodes(self, to_schedule, dirtynodes, orig_vmrrs):

        dirtynodes = set(dirtynodes)

        done = False

        while not done:

            stable = True

            cleanleases = set()

            for l in to_schedule:

                pnodes = set()

                for l_vmrr in orig_vmrrs[l]:

                    pnodes.update(l_vmrr.resources_in_pnode.keys())

                in_dirty = dirtynodes & pnodes

                in_clean = pnodes - dirtynodes

                if len(in_dirty) > 0 and len(in_clean) > 0:

                    stable = False

                    dirtynodes.update(in_clean)

                if len(in_clean) > 0 and len(in_dirty) == 0:

                    cleanleases.add(l)

            if stable == True:

                done = True

        return dirtynodes, cleanleases

    def reschedule_deadline_leases(self, leases, orig_vmrrs, earliest_time, earliest, nexttime, dirtytime):

        scheduled = set()

        new_vmrrs = {}

        for l in leases:

            if len(scheduled) < len(leases) and dirtytime != None:

                min_future_start = min([min([rr.start for rr in lrr]) for l2, lrr in orig_vmrrs.items() if l2 in leases and l2 not in scheduled])

                if min_future_start > dirtytime:

                    break

            last_vmrr = l.get_last_vmrr()

            if last_vmrr != None and last_vmrr.is_suspending():

                override_state = Lease.STATE_SUSPENDED_PENDING

                l_earliest_time = max(last_vmrr.post_rrs[-1].end, earliest_time)

            else:

                override_state = None

                l_earliest_time = earliest_time

            for n in earliest:

                earliest[n].time = max(l.start.requested, l_earliest_time)

            self.logger.debug("Rescheduling lease %s" % l.id)

            dur = l.get_remaining_duration_at(l_earliest_time)

            try:

                vmrr, preemptions = self.__schedule_asap(l, dur, nexttime, earliest, allow_in_future = True, override_state=override_state)

            except NotSchedulableException:

                vmrr = None

                preemptions = []

            if vmrr == None or vmrr.end - vmrr.start != dur or vmrr.end > l.deadline or len(preemptions) != 0:

                raise NotSchedulableException, "Could not schedule before deadline without making other leases miss deadline"

            if dirtytime != None:

                dirtytime = max(vmrr.end, dirtytime)

            for rr in vmrr.pre_rrs:

                self.slottable.add_reservation(rr)

            self.slottable.add_reservation(vmrr)

            for rr in vmrr.post_rrs:

                self.slottable.add_reservation(rr)

            new_vmrrs[l] = vmrr

            scheduled.add(l)

        return scheduled, new_vmrrs, dirtytime

    def reschedule_deadline(self, lease, duration, nexttime, earliest, override_state = None):

        for n in earliest:

            earliest[n].time = max(lease.start.requested, earliest[n].time)

        try:

            vmrr, preemptions = self.__schedule_asap(lease, duration, nexttime, earliest, allow_in_future = True, override_state=override_state)

        except NotSchedulableException:

            vmrr = None

            preemptions = []

        if vmrr == None or vmrr.end - vmrr.start != duration or vmrr.end > lease.deadline or len(preemptions)>0:

            self.logger.debug("Lease #%i cannot be scheduled before deadline using best-effort." % lease.id)

            raise NotSchedulableException, "Could not schedule before deadline without making other leases miss deadline"