elastic_scheduler.py 11.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
# Copyright (c) 2017, Daniele Venzano
#
# 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.

"""
The Elastic scheduler is the implementation of the scheduling algorithm presented in this paper:
https://arxiv.org/abs/1611.09528
"""

from collections import namedtuple
import logging
import threading
import time

from zoe_lib.state import Execution, SQLManager
27
from zoe_master.exceptions import ZoeException
28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45

from zoe_master.backends.interface import terminate_execution, get_platform_state, start_elastic, start_essential
from zoe_master.scheduler.simulated_platform import SimulatedPlatform
from zoe_master.exceptions import UnsupportedSchedulerPolicyError

log = logging.getLogger(__name__)

ExecutionProgress = namedtuple('ExecutionProgress', ['last_time_scheduled', 'progress_sequence'])


class ZoeElasticScheduler:
    """The Scheduler class for size-based scheduling. Policy can be "FIFO" or "SIZE"."""
    def __init__(self, state: SQLManager, policy):
        if policy != 'FIFO' and policy != 'SIZE':
            raise UnsupportedSchedulerPolicyError
        self.trigger_semaphore = threading.Semaphore(0)
        self.policy = policy
        self.queue = []
46
        self.queue_running = []
47 48 49
        self.additional_exec_state = {}
        self.async_threads = []
        self.loop_quit = False
Daniele Venzano's avatar
Daniele Venzano committed
50
        self.loop_th = threading.Thread(target=self._thread_wrapper, name='scheduler')
51 52
        self.loop_th.start()
        self.state = state
53 54 55 56 57
        for execution in self.state.execution_list(status='running'):
            if execution.all_services_running:
                self.queue_running.append(execution)
            else:
                self.queue.append(execution)
58
                self.additional_exec_state[execution.id] = ExecutionProgress(0, [])
59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83

    def trigger(self):
        """Trigger a scheduler run."""
        self.trigger_semaphore.release()

    def incoming(self, execution: Execution):
        """
        This method adds the execution to the end of the FIFO queue and triggers the scheduler.
        :param execution: The execution
        :return:
        """
        self.queue.append(execution)
        exec_data = ExecutionProgress(0, [])
        self.additional_exec_state[execution.id] = exec_data
        self.trigger()

    def terminate(self, execution: Execution) -> None:
        """
        Inform the master that an execution has been terminated. This can be done asynchronously.
        :param execution: the terminated execution
        :return: None
        """
        def async_termination(e):
            """Actual termination runs in a thread."""
            with e.termination_lock:
84 85 86 87 88
                try:
                    terminate_execution(e)
                except ZoeException as ex:
                    log.error('Error in termination thread: {}'.format(ex))
                    return
89 90 91 92 93 94
                self.trigger()
            log.debug('Execution {} terminated successfully'.format(e.id))

        try:
            self.queue.remove(execution)
        except ValueError:
95 96 97 98
            try:
                self.queue_running.remove(execution)
            except ValueError:
                log.error('Terminating execution {} that is not in any queue'.format(execution.id))
99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126

        try:
            del self.additional_exec_state[execution.id]
        except KeyError:
            pass

        th = threading.Thread(target=async_termination, name='termination_{}'.format(execution.id), args=(execution,))
        th.start()
        self.async_threads.append(th)

    def _cleanup_async_threads(self):
        counter = len(self.async_threads)
        while counter > 0:
            if len(self.async_threads) == 0:
                break
            th = self.async_threads.pop(0)
            th.join(0.1)
            if th.isAlive():  # join failed
                log.debug('Thread {} join failed'.format(th.name))
                self.async_threads.append(th)
            counter -= 1

    def _refresh_execution_sizes(self):
        for execution in self.queue:  # type: Execution
            exec_data = self.additional_exec_state[execution.id]
            if exec_data.last_time_scheduled == 0:
                progress = 0
            else:
127
                last_progress = (time.time() - exec_data.last_time_scheduled) / ((execution.services_count / execution.running_services_count) * execution.size)
128 129
                exec_data.progress_sequence.append(last_progress)
                progress = sum(exec_data.progress_sequence)
130
            remaining_execution_time = (1 - progress) * execution.size
131 132 133 134 135 136 137 138 139 140 141 142 143 144
            execution.size = remaining_execution_time * execution.services_count

    def _pop_all_with_same_size(self):
        out_list = []
        while len(self.queue) > 0:
            job = self.queue.pop(0)  # type: Execution
            ret = job.termination_lock.acquire(blocking=False)
            if ret and job.status != Execution.TERMINATED_STATUS:
                out_list.append(job)
            else:
                log.debug('While popping, throwing away execution {} that has the termination lock held'.format(job.id))

        return out_list

Daniele Venzano's avatar
Daniele Venzano committed
145 146 147 148 149 150 151 152 153 154
    def _thread_wrapper(self):
        while True:
            try:
                self.loop_start_th()
            except BaseException:  # pylint: disable=broad-except
                log.exception('Unmanaged exception in scheduler loop')
            else:
                log.debug('Scheduler thread terminated')
                break

155 156 157 158 159
    def loop_start_th(self):
        """The Scheduler thread loop."""
        auto_trigger_base = 60  # seconds
        auto_trigger = auto_trigger_base
        while True:
Daniele Venzano's avatar
Daniele Venzano committed
160 161 162 163 164 165 166 167 168 169
            ret = self.trigger_semaphore.acquire(timeout=1)
            if not ret:  # Semaphore timeout, do some thread cleanup
                self._cleanup_async_threads()
                auto_trigger -= 1
                if auto_trigger == 0:
                    auto_trigger = auto_trigger_base
                    self.trigger()
                continue
            if self.loop_quit:
                break
170

Daniele Venzano's avatar
Daniele Venzano committed
171 172 173 174
            if len(self.queue) == 0:
                log.debug("Scheduler loop has been triggered, but the queue is empty")
                continue
            log.debug("Scheduler loop has been triggered")
175

Daniele Venzano's avatar
Daniele Venzano committed
176 177
            while True:  # Inner loop will run until no new executions can be started or the queue is empty
                self._refresh_execution_sizes()
178

Daniele Venzano's avatar
Daniele Venzano committed
179 180
                if self.policy == "SIZE":
                    self.queue.sort(key=lambda execution: execution.size)
181

Daniele Venzano's avatar
Daniele Venzano committed
182 183 184 185 186 187 188 189 190 191
                log.debug('--> Queue dump after sorting')
                for j in self.queue:
                    log.debug(str(j))
                log.debug('--> End dump')

                jobs_to_attempt_scheduling = self._pop_all_with_same_size()
                log.debug('Scheduler inner loop, jobs to attempt scheduling:')
                for job in jobs_to_attempt_scheduling:
                    log.debug("-> {}".format(job))

192 193 194 195 196 197 198 199 200
                try:
                    platform_state = get_platform_state()
                except ZoeException:
                    log.error('Cannot retrieve platform state, cannot schedule')
                    for job in jobs_to_attempt_scheduling:
                        job.termination_lock.release()
                    self.queue = jobs_to_attempt_scheduling + self.queue
                    break

Daniele Venzano's avatar
Daniele Venzano committed
201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228
                cluster_status_snapshot = SimulatedPlatform(platform_state)
                log.debug(str(cluster_status_snapshot))

                jobs_to_launch = []
                free_resources = cluster_status_snapshot.aggregated_free_memory()

                # Try to find a placement solution using a snapshot of the platform status
                for job in jobs_to_attempt_scheduling:  # type: Execution
                    jobs_to_launch_copy = jobs_to_launch.copy()

                    # remove all elastic services from the previous simulation loop
                    for job_aux in jobs_to_launch:  # type: Execution
                        cluster_status_snapshot.deallocate_elastic(job_aux)

                    job_can_start = False
                    if not job.is_running:
                        job_can_start = cluster_status_snapshot.allocate_essential(job)

                    if job_can_start or job.is_running:
                        jobs_to_launch.append(job)

                    # Try to put back the elastic services
                    for job_aux in jobs_to_launch:
                        cluster_status_snapshot.allocate_elastic(job_aux)

                    current_free_resources = cluster_status_snapshot.aggregated_free_memory()
                    if current_free_resources >= free_resources:
                        jobs_to_launch = jobs_to_launch_copy
229
                        break
Daniele Venzano's avatar
Daniele Venzano committed
230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267
                    free_resources = current_free_resources

                log.debug('Allocation after simulation: {}'.format(cluster_status_snapshot.get_service_allocation()))

                # We port the results of the simulation into the real cluster
                for job in jobs_to_launch:  # type: Execution
                    if not job.essential_services_running:
                        ret = start_essential(job)
                        if ret == "fatal":
                            jobs_to_attempt_scheduling.remove(job)
                            continue  # trow away the execution
                        elif ret == "requeue":
                            self.queue.insert(0, job)
                            continue
                        elif ret == "ok":
                            job.set_running()
                        assert ret == "ok"

                    start_elastic(job)

                    if job.all_services_active:
                        log.debug('execution {}: all services are active'.format(job.id))
                        job.termination_lock.release()
                        jobs_to_attempt_scheduling.remove(job)
                        self.queue_running.append(job)

                for job in jobs_to_attempt_scheduling:
                    job.termination_lock.release()
                    # self.queue.insert(0, job)

                self.queue = jobs_to_attempt_scheduling + self.queue

                if len(self.queue) == 0:
                    log.debug('empty queue, exiting inner loop')
                    break
                if len(jobs_to_launch) == 0:
                    log.debug('No executions could be started, exiting inner loop')
                    break
268 269 270 271 272 273 274 275 276

    def quit(self):
        """Stop the scheduler thread."""
        self.loop_quit = True
        self.trigger()
        self.loop_th.join()

    def stats(self):
        """Scheduler statistics."""
277 278 279 280 281
        if self.policy == "SIZE":
            queue = sorted(self.queue, key=lambda execution: execution.size)
        else:
            queue = self.queue

282 283
        return {
            'queue_length': len(self.queue),
284
            'running_length': len(self.queue_running),
285 286 287 288
            'termination_threads_count': len(self.async_threads),
            'queue': [s.id for s in queue],
            'running_queue': [s.id for s in self.queue_running],
            'platform_stats': get_platform_state().serialize()
289
        }