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
441
442
443
444
445
446
447
448
449
450
451
452
453
|
/*
* 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.api.r
import java.io._
import java.net.{InetAddress, ServerSocket}
import java.util.{Map => JMap}
import scala.collection.JavaConversions._
import scala.io.Source
import scala.reflect.ClassTag
import scala.util.Try
import org.apache.spark._
import org.apache.spark.api.java.{JavaPairRDD, JavaRDD, JavaSparkContext}
import org.apache.spark.broadcast.Broadcast
import org.apache.spark.rdd.RDD
import org.apache.spark.util.Utils
private abstract class BaseRRDD[T: ClassTag, U: ClassTag](
parent: RDD[T],
numPartitions: Int,
func: Array[Byte],
deserializer: String,
serializer: String,
packageNames: Array[Byte],
rLibDir: String,
broadcastVars: Array[Broadcast[Object]])
extends RDD[U](parent) with Logging {
protected var dataStream: DataInputStream = _
private var bootTime: Double = _
override def getPartitions: Array[Partition] = parent.partitions
override def compute(partition: Partition, context: TaskContext): Iterator[U] = {
// Timing start
bootTime = System.currentTimeMillis / 1000.0
// The parent may be also an RRDD, so we should launch it first.
val parentIterator = firstParent[T].iterator(partition, context)
// we expect two connections
val serverSocket = new ServerSocket(0, 2, InetAddress.getByName("localhost"))
val listenPort = serverSocket.getLocalPort()
// The stdout/stderr is shared by multiple tasks, because we use one daemon
// to launch child process as worker.
val errThread = RRDD.createRWorker(rLibDir, listenPort)
// We use two sockets to separate input and output, then it's easy to manage
// the lifecycle of them to avoid deadlock.
// TODO: optimize it to use one socket
// the socket used to send out the input of task
serverSocket.setSoTimeout(10000)
val inSocket = serverSocket.accept()
startStdinThread(inSocket.getOutputStream(), parentIterator, partition.index)
// the socket used to receive the output of task
val outSocket = serverSocket.accept()
val inputStream = new BufferedInputStream(outSocket.getInputStream)
dataStream = new DataInputStream(inputStream)
serverSocket.close()
try {
return new Iterator[U] {
def next(): U = {
val obj = _nextObj
if (hasNext) {
_nextObj = read()
}
obj
}
var _nextObj = read()
def hasNext(): Boolean = {
val hasMore = (_nextObj != null)
if (!hasMore) {
dataStream.close()
}
hasMore
}
}
} catch {
case e: Exception =>
throw new SparkException("R computation failed with\n " + errThread.getLines())
}
}
/**
* Start a thread to write RDD data to the R process.
*/
private def startStdinThread[T](
output: OutputStream,
iter: Iterator[T],
partition: Int): Unit = {
val env = SparkEnv.get
val bufferSize = System.getProperty("spark.buffer.size", "65536").toInt
val stream = new BufferedOutputStream(output, bufferSize)
new Thread("writer for R") {
override def run(): Unit = {
try {
SparkEnv.set(env)
val dataOut = new DataOutputStream(stream)
dataOut.writeInt(partition)
SerDe.writeString(dataOut, deserializer)
SerDe.writeString(dataOut, serializer)
dataOut.writeInt(packageNames.length)
dataOut.write(packageNames)
dataOut.writeInt(func.length)
dataOut.write(func)
dataOut.writeInt(broadcastVars.length)
broadcastVars.foreach { broadcast =>
// TODO(shivaram): Read a Long in R to avoid this cast
dataOut.writeInt(broadcast.id.toInt)
// TODO: Pass a byte array from R to avoid this cast ?
val broadcastByteArr = broadcast.value.asInstanceOf[Array[Byte]]
dataOut.writeInt(broadcastByteArr.length)
dataOut.write(broadcastByteArr)
}
dataOut.writeInt(numPartitions)
if (!iter.hasNext) {
dataOut.writeInt(0)
} else {
dataOut.writeInt(1)
}
val printOut = new PrintStream(stream)
def writeElem(elem: Any): Unit = {
if (deserializer == SerializationFormats.BYTE) {
val elemArr = elem.asInstanceOf[Array[Byte]]
dataOut.writeInt(elemArr.length)
dataOut.write(elemArr)
} else if (deserializer == SerializationFormats.ROW) {
dataOut.write(elem.asInstanceOf[Array[Byte]])
} else if (deserializer == SerializationFormats.STRING) {
// write string(for StringRRDD)
printOut.println(elem)
}
}
for (elem <- iter) {
elem match {
case (key, value) =>
writeElem(key)
writeElem(value)
case _ =>
writeElem(elem)
}
}
stream.flush()
} catch {
// TODO: We should propogate this error to the task thread
case e: Exception =>
logError("R Writer thread got an exception", e)
} finally {
Try(output.close())
}
}
}.start()
}
protected def readData(length: Int): U
protected def read(): U = {
try {
val length = dataStream.readInt()
length match {
case SpecialLengths.TIMING_DATA =>
// Timing data from R worker
val boot = dataStream.readDouble - bootTime
val init = dataStream.readDouble
val broadcast = dataStream.readDouble
val input = dataStream.readDouble
val compute = dataStream.readDouble
val output = dataStream.readDouble
logInfo(
("Times: boot = %.3f s, init = %.3f s, broadcast = %.3f s, " +
"read-input = %.3f s, compute = %.3f s, write-output = %.3f s, " +
"total = %.3f s").format(
boot,
init,
broadcast,
input,
compute,
output,
boot + init + broadcast + input + compute + output))
read()
case length if length >= 0 =>
readData(length)
}
} catch {
case eof: EOFException =>
throw new SparkException("R worker exited unexpectedly (cranshed)", eof)
}
}
}
/**
* Form an RDD[(Int, Array[Byte])] from key-value pairs returned from R.
* This is used by SparkR's shuffle operations.
*/
private class PairwiseRRDD[T: ClassTag](
parent: RDD[T],
numPartitions: Int,
hashFunc: Array[Byte],
deserializer: String,
packageNames: Array[Byte],
rLibDir: String,
broadcastVars: Array[Object])
extends BaseRRDD[T, (Int, Array[Byte])](
parent, numPartitions, hashFunc, deserializer,
SerializationFormats.BYTE, packageNames, rLibDir,
broadcastVars.map(x => x.asInstanceOf[Broadcast[Object]])) {
override protected def readData(length: Int): (Int, Array[Byte]) = {
length match {
case length if length == 2 =>
val hashedKey = dataStream.readInt()
val contentPairsLength = dataStream.readInt()
val contentPairs = new Array[Byte](contentPairsLength)
dataStream.readFully(contentPairs)
(hashedKey, contentPairs)
case _ => null
}
}
lazy val asJavaPairRDD : JavaPairRDD[Int, Array[Byte]] = JavaPairRDD.fromRDD(this)
}
/**
* An RDD that stores serialized R objects as Array[Byte].
*/
private class RRDD[T: ClassTag](
parent: RDD[T],
func: Array[Byte],
deserializer: String,
serializer: String,
packageNames: Array[Byte],
rLibDir: String,
broadcastVars: Array[Object])
extends BaseRRDD[T, Array[Byte]](
parent, -1, func, deserializer, serializer, packageNames, rLibDir,
broadcastVars.map(x => x.asInstanceOf[Broadcast[Object]])) {
override protected def readData(length: Int): Array[Byte] = {
length match {
case length if length > 0 =>
val obj = new Array[Byte](length)
dataStream.readFully(obj)
obj
case _ => null
}
}
lazy val asJavaRDD : JavaRDD[Array[Byte]] = JavaRDD.fromRDD(this)
}
/**
* An RDD that stores R objects as Array[String].
*/
private class StringRRDD[T: ClassTag](
parent: RDD[T],
func: Array[Byte],
deserializer: String,
packageNames: Array[Byte],
rLibDir: String,
broadcastVars: Array[Object])
extends BaseRRDD[T, String](
parent, -1, func, deserializer, SerializationFormats.STRING, packageNames, rLibDir,
broadcastVars.map(x => x.asInstanceOf[Broadcast[Object]])) {
override protected def readData(length: Int): String = {
length match {
case length if length > 0 =>
SerDe.readStringBytes(dataStream, length)
case _ => null
}
}
lazy val asJavaRDD : JavaRDD[String] = JavaRDD.fromRDD(this)
}
private object SpecialLengths {
val TIMING_DATA = -1
}
private[r] class BufferedStreamThread(
in: InputStream,
name: String,
errBufferSize: Int) extends Thread(name) with Logging {
val lines = new Array[String](errBufferSize)
var lineIdx = 0
override def run() {
for (line <- Source.fromInputStream(in).getLines) {
synchronized {
lines(lineIdx) = line
lineIdx = (lineIdx + 1) % errBufferSize
}
logInfo(line)
}
}
def getLines(): String = synchronized {
(0 until errBufferSize).filter { x =>
lines((x + lineIdx) % errBufferSize) != null
}.map { x =>
lines((x + lineIdx) % errBufferSize)
}.mkString("\n")
}
}
private[r] object RRDD {
// Because forking processes from Java is expensive, we prefer to launch
// a single R daemon (daemon.R) and tell it to fork new workers for our tasks.
// This daemon currently only works on UNIX-based systems now, so we should
// also fall back to launching workers (worker.R) directly.
private[this] var errThread: BufferedStreamThread = _
private[this] var daemonChannel: DataOutputStream = _
def createSparkContext(
master: String,
appName: String,
sparkHome: String,
jars: Array[String],
sparkEnvirMap: JMap[Object, Object],
sparkExecutorEnvMap: JMap[Object, Object]): JavaSparkContext = {
val sparkConf = new SparkConf().setAppName(appName)
.setSparkHome(sparkHome)
.setJars(jars)
// Override `master` if we have a user-specified value
if (master != "") {
sparkConf.setMaster(master)
} else {
// If conf has no master set it to "local" to maintain
// backwards compatibility
sparkConf.setIfMissing("spark.master", "local")
}
for ((name, value) <- sparkEnvirMap) {
sparkConf.set(name.asInstanceOf[String], value.asInstanceOf[String])
}
for ((name, value) <- sparkExecutorEnvMap) {
sparkConf.setExecutorEnv(name.asInstanceOf[String], value.asInstanceOf[String])
}
new JavaSparkContext(sparkConf)
}
/**
* Start a thread to print the process's stderr to ours
*/
private def startStdoutThread(proc: Process): BufferedStreamThread = {
val BUFFER_SIZE = 100
val thread = new BufferedStreamThread(proc.getInputStream, "stdout reader for R", BUFFER_SIZE)
thread.setDaemon(true)
thread.start()
thread
}
private def createRProcess(rLibDir: String, port: Int, script: String): BufferedStreamThread = {
val rCommand = "Rscript"
val rOptions = "--vanilla"
val rExecScript = rLibDir + "/SparkR/worker/" + script
val pb = new ProcessBuilder(List(rCommand, rOptions, rExecScript))
// Unset the R_TESTS environment variable for workers.
// This is set by R CMD check as startup.Rs
// (http://svn.r-project.org/R/trunk/src/library/tools/R/testing.R)
// and confuses worker script which tries to load a non-existent file
pb.environment().put("R_TESTS", "")
pb.environment().put("SPARKR_RLIBDIR", rLibDir)
pb.environment().put("SPARKR_WORKER_PORT", port.toString)
pb.redirectErrorStream(true) // redirect stderr into stdout
val proc = pb.start()
val errThread = startStdoutThread(proc)
errThread
}
/**
* ProcessBuilder used to launch worker R processes.
*/
def createRWorker(rLibDir: String, port: Int): BufferedStreamThread = {
val useDaemon = SparkEnv.get.conf.getBoolean("spark.sparkr.use.daemon", true)
if (!Utils.isWindows && useDaemon) {
synchronized {
if (daemonChannel == null) {
// we expect one connections
val serverSocket = new ServerSocket(0, 1, InetAddress.getByName("localhost"))
val daemonPort = serverSocket.getLocalPort
errThread = createRProcess(rLibDir, daemonPort, "daemon.R")
// the socket used to send out the input of task
serverSocket.setSoTimeout(10000)
val sock = serverSocket.accept()
daemonChannel = new DataOutputStream(new BufferedOutputStream(sock.getOutputStream))
serverSocket.close()
}
try {
daemonChannel.writeInt(port)
daemonChannel.flush()
} catch {
case e: IOException =>
// daemon process died
daemonChannel.close()
daemonChannel = null
errThread = null
// fail the current task, retry by scheduler
throw e
}
errThread
}
} else {
createRProcess(rLibDir, port, "worker.R")
}
}
/**
* Create an RRDD given a sequence of byte arrays. Used to create RRDD when `parallelize` is
* called from R.
*/
def createRDDFromArray(jsc: JavaSparkContext, arr: Array[Array[Byte]]): JavaRDD[Array[Byte]] = {
JavaRDD.fromRDD(jsc.sc.parallelize(arr, arr.length))
}
}
|