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
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
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
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
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
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
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
229
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
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
|
/*
* Licensed to the Apache Software Foundation (ASF) under one or more
* contributor license agreements. See the NOTICE file distributed with
* this work for additional information regarding copyright ownership.
* The ASF licenses this file to You 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.
*/
package org.apache.spark.scheduler.cluster
import java.lang.{Boolean => JBoolean}
import scala.collection.mutable.ArrayBuffer
import scala.collection.mutable.HashMap
import scala.collection.mutable.HashSet
import org.apache.spark._
import org.apache.spark.TaskState.TaskState
import org.apache.spark.scheduler._
import org.apache.spark.scheduler.cluster.SchedulingMode.SchedulingMode
import java.nio.ByteBuffer
import java.util.concurrent.atomic.AtomicLong
import java.util.{TimerTask, Timer}
/**
* The main TaskScheduler implementation, for running tasks on a cluster. Clients should first call
* initialize() and start(), then submit task sets through the runTasks method.
*
* This class can work with multiple types of clusters by acting through a SchedulerBackend.
* It handles common logic, like determining a scheduling order across jobs, waking up to launch
* speculative tasks, etc.
*
* THREADING: SchedulerBackends and task-submitting clients can call this class from multiple
* threads, so it needs locks in public API methods to maintain its state. In addition, some
* SchedulerBackends sycnchronize on themselves when they want to send events here, and then
* acquire a lock on us, so we need to make sure that we don't try to lock the backend while
* we are holding a lock on ourselves.
*/
private[spark] class ClusterScheduler(val sc: SparkContext)
extends TaskScheduler
with Logging
{
// How often to check for speculative tasks
val SPECULATION_INTERVAL = System.getProperty("spark.speculation.interval", "100").toLong
// Threshold above which we warn user initial TaskSet may be starved
val STARVATION_TIMEOUT = System.getProperty("spark.starvation.timeout", "15000").toLong
val activeTaskSets = new HashMap[String, TaskSetManager]
val taskIdToTaskSetId = new HashMap[Long, String]
val taskIdToExecutorId = new HashMap[Long, String]
val taskSetTaskIds = new HashMap[String, HashSet[Long]]
@volatile private var hasReceivedTask = false
@volatile private var hasLaunchedTask = false
private val starvationTimer = new Timer(true)
// Incrementing Mesos task IDs
val nextTaskId = new AtomicLong(0)
// Which executor IDs we have executors on
val activeExecutorIds = new HashSet[String]
// The set of executors we have on each host; this is used to compute hostsAlive, which
// in turn is used to decide when we can attain data locality on a given host
private val executorsByHost = new HashMap[String, HashSet[String]]
private val executorIdToHost = new HashMap[String, String]
// JAR server, if any JARs were added by the user to the SparkContext
var jarServer: HttpServer = null
// URIs of JARs to pass to executor
var jarUris: String = ""
// Listener object to pass upcalls into
var listener: TaskSchedulerListener = null
var backend: SchedulerBackend = null
val mapOutputTracker = SparkEnv.get.mapOutputTracker
var schedulableBuilder: SchedulableBuilder = null
var rootPool: Pool = null
// default scheduler is FIFO
val schedulingMode: SchedulingMode = SchedulingMode.withName(
System.getProperty("spark.scheduler.mode", "FIFO"))
override def setListener(listener: TaskSchedulerListener) {
this.listener = listener
}
def initialize(context: SchedulerBackend) {
backend = context
// temporarily set rootPool name to empty
rootPool = new Pool("", schedulingMode, 0, 0)
schedulableBuilder = {
schedulingMode match {
case SchedulingMode.FIFO =>
new FIFOSchedulableBuilder(rootPool)
case SchedulingMode.FAIR =>
new FairSchedulableBuilder(rootPool)
}
}
schedulableBuilder.buildPools()
}
def newTaskId(): Long = nextTaskId.getAndIncrement()
override def start() {
backend.start()
if (System.getProperty("spark.speculation", "false").toBoolean) {
new Thread("ClusterScheduler speculation check") {
setDaemon(true)
override def run() {
logInfo("Starting speculative execution thread")
while (true) {
try {
Thread.sleep(SPECULATION_INTERVAL)
} catch {
case e: InterruptedException => {}
}
checkSpeculatableTasks()
}
}
}.start()
}
}
override def submitTasks(taskSet: TaskSet) {
val tasks = taskSet.tasks
logInfo("Adding task set " + taskSet.id + " with " + tasks.length + " tasks")
this.synchronized {
val manager = new ClusterTaskSetManager(this, taskSet)
activeTaskSets(taskSet.id) = manager
schedulableBuilder.addTaskSetManager(manager, manager.taskSet.properties)
taskSetTaskIds(taskSet.id) = new HashSet[Long]()
if (!hasReceivedTask) {
starvationTimer.scheduleAtFixedRate(new TimerTask() {
override def run() {
if (!hasLaunchedTask) {
logWarning("Initial job has not accepted any resources; " +
"check your cluster UI to ensure that workers are registered " +
"and have sufficient memory")
} else {
this.cancel()
}
}
}, STARVATION_TIMEOUT, STARVATION_TIMEOUT)
}
hasReceivedTask = true
}
backend.reviveOffers()
}
def taskSetFinished(manager: TaskSetManager) {
this.synchronized {
activeTaskSets -= manager.taskSet.id
manager.parent.removeSchedulable(manager)
logInfo("Remove TaskSet %s from pool %s".format(manager.taskSet.id, manager.parent.name))
taskIdToTaskSetId --= taskSetTaskIds(manager.taskSet.id)
taskIdToExecutorId --= taskSetTaskIds(manager.taskSet.id)
taskSetTaskIds.remove(manager.taskSet.id)
}
}
/**
* Called by cluster manager to offer resources on slaves. We respond by asking our active task
* sets for tasks in order of priority. We fill each node with tasks in a round-robin manner so
* that tasks are balanced across the cluster.
*/
def resourceOffers(offers: Seq[WorkerOffer]): Seq[Seq[TaskDescription]] = synchronized {
SparkEnv.set(sc.env)
// Mark each slave as alive and remember its hostname
for (o <- offers) {
executorIdToHost(o.executorId) = o.host
if (!executorsByHost.contains(o.host)) {
executorsByHost(o.host) = new HashSet[String]()
executorGained(o.executorId, o.host)
}
}
// Build a list of tasks to assign to each worker
val tasks = offers.map(o => new ArrayBuffer[TaskDescription](o.cores))
val availableCpus = offers.map(o => o.cores).toArray
val sortedTaskSets = rootPool.getSortedTaskSetQueue()
for (taskSet <- sortedTaskSets) {
logDebug("parentName: %s, name: %s, runningTasks: %s".format(
taskSet.parent.name, taskSet.name, taskSet.runningTasks))
}
// Take each TaskSet in our scheduling order, and then offer it each node in increasing order
// of locality levels so that it gets a chance to launch local tasks on all of them.
var launchedTask = false
for (taskSet <- sortedTaskSets; maxLocality <- TaskLocality.values) {
do {
launchedTask = false
for (i <- 0 until offers.size) {
val execId = offers(i).executorId
val host = offers(i).host
for (task <- taskSet.resourceOffer(execId, host, availableCpus(i), maxLocality)) {
tasks(i) += task
val tid = task.taskId
taskIdToTaskSetId(tid) = taskSet.taskSet.id
taskSetTaskIds(taskSet.taskSet.id) += tid
taskIdToExecutorId(tid) = execId
activeExecutorIds += execId
executorsByHost(host) += execId
availableCpus(i) -= 1
launchedTask = true
}
}
} while (launchedTask)
}
if (tasks.size > 0) {
hasLaunchedTask = true
}
return tasks
}
def statusUpdate(tid: Long, state: TaskState, serializedData: ByteBuffer) {
var taskSetToUpdate: Option[TaskSetManager] = None
var failedExecutor: Option[String] = None
var taskFailed = false
synchronized {
try {
if (state == TaskState.LOST && taskIdToExecutorId.contains(tid)) {
// We lost this entire executor, so remember that it's gone
val execId = taskIdToExecutorId(tid)
if (activeExecutorIds.contains(execId)) {
removeExecutor(execId)
failedExecutor = Some(execId)
}
}
taskIdToTaskSetId.get(tid) match {
case Some(taskSetId) =>
if (activeTaskSets.contains(taskSetId)) {
taskSetToUpdate = Some(activeTaskSets(taskSetId))
}
if (TaskState.isFinished(state)) {
taskIdToTaskSetId.remove(tid)
if (taskSetTaskIds.contains(taskSetId)) {
taskSetTaskIds(taskSetId) -= tid
}
taskIdToExecutorId.remove(tid)
}
if (state == TaskState.FAILED) {
taskFailed = true
}
case None =>
logInfo("Ignoring update from TID " + tid + " because its task set is gone")
}
} catch {
case e: Exception => logError("Exception in statusUpdate", e)
}
}
// Update the task set and DAGScheduler without holding a lock on this, since that can deadlock
if (taskSetToUpdate != None) {
taskSetToUpdate.get.statusUpdate(tid, state, serializedData)
}
if (failedExecutor != None) {
listener.executorLost(failedExecutor.get)
backend.reviveOffers()
}
if (taskFailed) {
// Also revive offers if a task had failed for some reason other than host lost
backend.reviveOffers()
}
}
def error(message: String) {
synchronized {
if (activeTaskSets.size > 0) {
// Have each task set throw a SparkException with the error
for ((taskSetId, manager) <- activeTaskSets) {
try {
manager.error(message)
} catch {
case e: Exception => logError("Exception in error callback", e)
}
}
} else {
// No task sets are active but we still got an error. Just exit since this
// must mean the error is during registration.
// It might be good to do something smarter here in the future.
logError("Exiting due to error from cluster scheduler: " + message)
System.exit(1)
}
}
}
override def stop() {
if (backend != null) {
backend.stop()
}
if (jarServer != null) {
jarServer.stop()
}
// sleeping for an arbitrary 5 seconds : to ensure that messages are sent out.
// TODO: Do something better !
Thread.sleep(5000L)
}
override def defaultParallelism() = backend.defaultParallelism()
// Check for speculatable tasks in all our active jobs.
def checkSpeculatableTasks() {
var shouldRevive = false
synchronized {
shouldRevive = rootPool.checkSpeculatableTasks()
}
if (shouldRevive) {
backend.reviveOffers()
}
}
// Check for pending tasks in all our active jobs.
def hasPendingTasks: Boolean = {
synchronized {
rootPool.hasPendingTasks()
}
}
def executorLost(executorId: String, reason: ExecutorLossReason) {
var failedExecutor: Option[String] = None
synchronized {
if (activeExecutorIds.contains(executorId)) {
val hostPort = executorIdToHost(executorId)
logError("Lost executor %s on %s: %s".format(executorId, hostPort, reason))
removeExecutor(executorId)
failedExecutor = Some(executorId)
} else {
// We may get multiple executorLost() calls with different loss reasons. For example, one
// may be triggered by a dropped connection from the slave while another may be a report
// of executor termination from Mesos. We produce log messages for both so we eventually
// report the termination reason.
logError("Lost an executor " + executorId + " (already removed): " + reason)
}
}
// Call listener.executorLost without holding the lock on this to prevent deadlock
if (failedExecutor != None) {
listener.executorLost(failedExecutor.get)
backend.reviveOffers()
}
}
/** Remove an executor from all our data structures and mark it as lost */
private def removeExecutor(executorId: String) {
activeExecutorIds -= executorId
val host = executorIdToHost(executorId)
val execs = executorsByHost.getOrElse(host, new HashSet)
execs -= executorId
if (execs.isEmpty) {
executorsByHost -= host
}
executorIdToHost -= executorId
rootPool.executorLost(executorId, host)
}
def executorGained(execId: String, host: String) {
listener.executorGained(execId, host)
}
def getExecutorsAliveOnHost(host: String): Option[Set[String]] = synchronized {
executorsByHost.get(host).map(_.toSet)
}
def hasExecutorsAliveOnHost(host: String): Boolean = synchronized {
executorsByHost.contains(host)
}
def isExecutorAlive(execId: String): Boolean = synchronized {
activeExecutorIds.contains(execId)
}
// By default, rack is unknown
def getRackForHost(value: String): Option[String] = None
}
object ClusterScheduler {
/**
* Used to balance containers across hosts.
*
* Accepts a map of hosts to resource offers for that host, and returns a prioritized list of
* resource offers representing the order in which the offers should be used. The resource
* offers are ordered such that we'll allocate one container on each host before allocating a
* second container on any host, and so on, in order to reduce the damage if a host fails.
*
* For example, given <h1, [o1, o2, o3]>, <h2, [o4]>, <h1, [o5, o6]>, returns
* [o1, o5, o4, 02, o6, o3]
*/
def prioritizeContainers[K, T] (map: HashMap[K, ArrayBuffer[T]]): List[T] = {
val _keyList = new ArrayBuffer[K](map.size)
_keyList ++= map.keys
// order keyList based on population of value in map
val keyList = _keyList.sortWith(
(left, right) => map(left).size > map(right).size
)
val retval = new ArrayBuffer[T](keyList.size * 2)
var index = 0
var found = true
while (found) {
found = false
for (key <- keyList) {
val containerList: ArrayBuffer[T] = map.get(key).getOrElse(null)
assert(containerList != null)
// Get the index'th entry for this host - if present
if (index < containerList.size){
retval += containerList.apply(index)
found = true
}
}
index += 1
}
retval.toList
}
}
|