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
|
package spark.scheduler.mesos
import com.google.protobuf.ByteString
import org.apache.mesos.{Scheduler => MScheduler}
import org.apache.mesos._
import org.apache.mesos.Protos.{TaskInfo => MesosTaskInfo, TaskState => MesosTaskState, _}
import spark.{SparkException, Utils, Logging, SparkContext}
import scala.collection.mutable.{ArrayBuffer, HashMap, HashSet}
import scala.collection.JavaConversions._
import java.io.File
import spark.scheduler.cluster._
import java.util.{ArrayList => JArrayList, List => JList}
import java.util.Collections
import spark.TaskState
/**
* A SchedulerBackend for running fine-grained tasks on Mesos. Each Spark task is mapped to a
* separate Mesos task, allowing multiple applications to share cluster nodes both in space (tasks
* from multiple apps can run on different cores) and in time (a core can switch ownership).
*/
private[spark] class MesosSchedulerBackend(
scheduler: ClusterScheduler,
sc: SparkContext,
master: String,
frameworkName: String)
extends SchedulerBackend
with MScheduler
with Logging {
// Memory used by each executor (in megabytes)
val EXECUTOR_MEMORY = {
if (System.getenv("SPARK_MEM") != null) {
Utils.memoryStringToMb(System.getenv("SPARK_MEM"))
// TODO: Might need to add some extra memory for the non-heap parts of the JVM
} else {
512
}
}
// Lock used to wait for scheduler to be registered
var isRegistered = false
val registeredLock = new Object()
// Driver for talking to Mesos
var driver: SchedulerDriver = null
// Which slave IDs we have executors on
val slaveIdsWithExecutors = new HashSet[String]
val taskIdToSlaveId = new HashMap[Long, String]
// An ExecutorInfo for our tasks
var executorInfo: ExecutorInfo = null
override def start() {
synchronized {
new Thread("MesosSchedulerBackend driver") {
setDaemon(true)
override def run() {
val scheduler = MesosSchedulerBackend.this
val fwInfo = FrameworkInfo.newBuilder().setUser("").setName(frameworkName).build()
driver = new MesosSchedulerDriver(scheduler, fwInfo, master)
try {
val ret = driver.run()
logInfo("driver.run() returned with code " + ret)
} catch {
case e: Exception => logError("driver.run() failed", e)
}
}
}.start()
executorInfo = createExecutorInfo()
waitForRegister()
}
}
def createExecutorInfo(): ExecutorInfo = {
val sparkHome = sc.getSparkHome() match {
case Some(path) =>
path
case None =>
throw new SparkException("Spark home is not set; set it through the spark.home system " +
"property, the SPARK_HOME environment variable or the SparkContext constructor")
}
val execScript = new File(sparkHome, "spark-executor").getCanonicalPath
val environment = Environment.newBuilder()
sc.executorEnvs.foreach { case (key, value) =>
environment.addVariables(Environment.Variable.newBuilder()
.setName(key)
.setValue(value)
.build())
}
val memory = Resource.newBuilder()
.setName("mem")
.setType(Value.Type.SCALAR)
.setScalar(Value.Scalar.newBuilder().setValue(EXECUTOR_MEMORY).build())
.build()
val command = CommandInfo.newBuilder()
.setValue(execScript)
.setEnvironment(environment)
.build()
ExecutorInfo.newBuilder()
.setExecutorId(ExecutorID.newBuilder().setValue("default").build())
.setCommand(command)
.setData(ByteString.copyFrom(createExecArg()))
.addResources(memory)
.build()
}
/**
* Create and serialize the executor argument to pass to Mesos. Our executor arg is an array
* containing all the spark.* system properties in the form of (String, String) pairs.
*/
private def createExecArg(): Array[Byte] = {
val props = new HashMap[String, String]
val iterator = System.getProperties.entrySet.iterator
while (iterator.hasNext) {
val entry = iterator.next
val (key, value) = (entry.getKey.toString, entry.getValue.toString)
if (key.startsWith("spark.")) {
props(key) = value
}
}
// Serialize the map as an array of (String, String) pairs
return Utils.serialize(props.toArray)
}
override def offerRescinded(d: SchedulerDriver, o: OfferID) {}
override def registered(d: SchedulerDriver, frameworkId: FrameworkID, masterInfo: MasterInfo) {
logInfo("Registered as framework ID " + frameworkId.getValue)
registeredLock.synchronized {
isRegistered = true
registeredLock.notifyAll()
}
}
def waitForRegister() {
registeredLock.synchronized {
while (!isRegistered) {
registeredLock.wait()
}
}
}
override def disconnected(d: SchedulerDriver) {}
override def reregistered(d: SchedulerDriver, masterInfo: MasterInfo) {}
/**
* Method called by Mesos to offer resources on slaves. We resond 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.
*/
override def resourceOffers(d: SchedulerDriver, offers: JList[Offer]) {
synchronized {
// Build a big list of the offerable workers, and remember their indices so that we can
// figure out which Offer to reply to for each worker
val offerableIndices = new ArrayBuffer[Int]
val offerableWorkers = new ArrayBuffer[WorkerOffer]
def enoughMemory(o: Offer) = {
val mem = getResource(o.getResourcesList, "mem")
val slaveId = o.getSlaveId.getValue
mem >= EXECUTOR_MEMORY || slaveIdsWithExecutors.contains(slaveId)
}
for ((offer, index) <- offers.zipWithIndex if enoughMemory(offer)) {
offerableIndices += index
offerableWorkers += new WorkerOffer(
offer.getSlaveId.getValue,
offer.getHostname,
getResource(offer.getResourcesList, "cpus").toInt)
}
// Call into the ClusterScheduler
val taskLists = scheduler.resourceOffers(offerableWorkers)
// Build a list of Mesos tasks for each slave
val mesosTasks = offers.map(o => Collections.emptyList[MesosTaskInfo]())
for ((taskList, index) <- taskLists.zipWithIndex) {
if (!taskList.isEmpty) {
val offerNum = offerableIndices(index)
val slaveId = offers(offerNum).getSlaveId.getValue
slaveIdsWithExecutors += slaveId
mesosTasks(offerNum) = new JArrayList[MesosTaskInfo](taskList.size)
for (taskDesc <- taskList) {
taskIdToSlaveId(taskDesc.taskId) = slaveId
mesosTasks(offerNum).add(createMesosTask(taskDesc, slaveId))
}
}
}
// Reply to the offers
val filters = Filters.newBuilder().setRefuseSeconds(1).build() // TODO: lower timeout?
for (i <- 0 until offers.size) {
d.launchTasks(offers(i).getId, mesosTasks(i), filters)
}
}
}
/** Helper function to pull out a resource from a Mesos Resources protobuf */
def getResource(res: JList[Resource], name: String): Double = {
for (r <- res if r.getName == name) {
return r.getScalar.getValue
}
// If we reached here, no resource with the required name was present
throw new IllegalArgumentException("No resource called " + name + " in " + res)
}
/** Turn a Spark TaskDescription into a Mesos task */
def createMesosTask(task: TaskDescription, slaveId: String): MesosTaskInfo = {
val taskId = TaskID.newBuilder().setValue(task.taskId.toString).build()
val cpuResource = Resource.newBuilder()
.setName("cpus")
.setType(Value.Type.SCALAR)
.setScalar(Value.Scalar.newBuilder().setValue(1).build())
.build()
return MesosTaskInfo.newBuilder()
.setTaskId(taskId)
.setSlaveId(SlaveID.newBuilder().setValue(slaveId).build())
.setExecutor(executorInfo)
.setName(task.name)
.addResources(cpuResource)
.setData(ByteString.copyFrom(task.serializedTask))
.build()
}
/** Check whether a Mesos task state represents a finished task */
def isFinished(state: MesosTaskState) = {
state == MesosTaskState.TASK_FINISHED ||
state == MesosTaskState.TASK_FAILED ||
state == MesosTaskState.TASK_KILLED ||
state == MesosTaskState.TASK_LOST
}
override def statusUpdate(d: SchedulerDriver, status: TaskStatus) {
val tid = status.getTaskId.getValue.toLong
val state = TaskState.fromMesos(status.getState)
synchronized {
if (status.getState == MesosTaskState.TASK_LOST && taskIdToSlaveId.contains(tid)) {
// We lost the executor on this slave, so remember that it's gone
slaveIdsWithExecutors -= taskIdToSlaveId(tid)
}
if (isFinished(status.getState)) {
taskIdToSlaveId.remove(tid)
}
}
scheduler.statusUpdate(tid, state, status.getData.asReadOnlyByteBuffer)
}
override def error(d: SchedulerDriver, message: String) {
logError("Mesos error: " + message)
scheduler.error(message)
}
override def stop() {
if (driver != null) {
driver.stop()
}
}
override def reviveOffers() {
driver.reviveOffers()
}
override def frameworkMessage(d: SchedulerDriver, e: ExecutorID, s: SlaveID, b: Array[Byte]) {}
private def recordSlaveLost(d: SchedulerDriver, slaveId: SlaveID, reason: ExecutorLossReason) {
logInfo("Mesos slave lost: " + slaveId.getValue)
synchronized {
slaveIdsWithExecutors -= slaveId.getValue
}
scheduler.slaveLost(slaveId.getValue, reason)
}
override def slaveLost(d: SchedulerDriver, slaveId: SlaveID) {
recordSlaveLost(d, slaveId, SlaveLost())
}
override def executorLost(d: SchedulerDriver, executorId: ExecutorID,
slaveId: SlaveID, status: Int) {
logInfo("Executor lost: %s, marking slave %s as lost".format(executorId.getValue,
slaveId.getValue))
recordSlaveLost(d, slaveId, ExecutorExited(status))
}
// TODO: query Mesos for number of cores
override def defaultParallelism() = System.getProperty("spark.default.parallelism", "8").toInt
}
|