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
454
455
456
457
458
459
460
461
462
463
464
465
|
/*
* 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.metrics
import java.io.{File, FileWriter, PrintWriter}
import scala.collection.mutable.ArrayBuffer
import org.apache.commons.lang3.RandomUtils
import org.apache.hadoop.conf.Configuration
import org.apache.hadoop.fs.{FileSystem, Path}
import org.apache.hadoop.io.{LongWritable, Text}
import org.apache.hadoop.mapred.lib.{CombineFileInputFormat => OldCombineFileInputFormat,
CombineFileRecordReader => OldCombineFileRecordReader, CombineFileSplit => OldCombineFileSplit}
import org.apache.hadoop.mapred.{JobConf, Reporter, FileSplit => OldFileSplit,
InputSplit => OldInputSplit, LineRecordReader => OldLineRecordReader,
RecordReader => OldRecordReader, TextInputFormat => OldTextInputFormat}
import org.apache.hadoop.mapreduce.lib.input.{CombineFileInputFormat => NewCombineFileInputFormat,
CombineFileRecordReader => NewCombineFileRecordReader, CombineFileSplit => NewCombineFileSplit,
FileSplit => NewFileSplit, TextInputFormat => NewTextInputFormat}
import org.apache.hadoop.mapreduce.lib.output.{TextOutputFormat => NewTextOutputFormat}
import org.apache.hadoop.mapreduce.{TaskAttemptContext, InputSplit => NewInputSplit,
RecordReader => NewRecordReader}
import org.scalatest.{BeforeAndAfter, FunSuite}
import org.apache.spark.SharedSparkContext
import org.apache.spark.deploy.SparkHadoopUtil
import org.apache.spark.scheduler.{SparkListener, SparkListenerTaskEnd}
import org.apache.spark.util.Utils
class InputOutputMetricsSuite extends FunSuite with SharedSparkContext
with BeforeAndAfter {
@transient var tmpDir: File = _
@transient var tmpFile: File = _
@transient var tmpFilePath: String = _
@transient val numRecords: Int = 100000
@transient val numBuckets: Int = 10
before {
tmpDir = Utils.createTempDir()
val testTempDir = new File(tmpDir, "test")
testTempDir.mkdir()
tmpFile = new File(testTempDir, getClass.getSimpleName + ".txt")
val pw = new PrintWriter(new FileWriter(tmpFile))
for (x <- 1 to numRecords) {
pw.println(RandomUtils.nextInt(0, numBuckets))
}
pw.close()
// Path to tmpFile
tmpFilePath = "file://" + tmpFile.getAbsolutePath
}
after {
Utils.deleteRecursively(tmpDir)
}
test("input metrics for old hadoop with coalesce") {
val bytesRead = runAndReturnBytesRead {
sc.textFile(tmpFilePath, 4).count()
}
val bytesRead2 = runAndReturnBytesRead {
sc.textFile(tmpFilePath, 4).coalesce(2).count()
}
assert(bytesRead != 0)
assert(bytesRead == bytesRead2)
assert(bytesRead2 >= tmpFile.length())
}
test("input metrics with cache and coalesce") {
// prime the cache manager
val rdd = sc.textFile(tmpFilePath, 4).cache()
rdd.collect()
val bytesRead = runAndReturnBytesRead {
rdd.count()
}
val bytesRead2 = runAndReturnBytesRead {
rdd.coalesce(4).count()
}
// for count and coelesce, the same bytes should be read.
assert(bytesRead != 0)
assert(bytesRead2 == bytesRead)
}
/**
* This checks the situation where we have interleaved reads from
* different sources. Currently, we only accumulate fron the first
* read method we find in the task. This test uses cartesian to create
* the interleaved reads.
*
* Once https://issues.apache.org/jira/browse/SPARK-5225 is fixed
* this test should break.
*/
test("input metrics with mixed read method") {
// prime the cache manager
val numPartitions = 2
val rdd = sc.parallelize(1 to 100, numPartitions).cache()
rdd.collect()
val rdd2 = sc.textFile(tmpFilePath, numPartitions)
val bytesRead = runAndReturnBytesRead {
rdd.count()
}
val bytesRead2 = runAndReturnBytesRead {
rdd2.count()
}
val cartRead = runAndReturnBytesRead {
rdd.cartesian(rdd2).count()
}
assert(cartRead != 0)
assert(bytesRead != 0)
// We read from the first rdd of the cartesian once per partition.
assert(cartRead == bytesRead * numPartitions)
}
test("input metrics for new Hadoop API with coalesce") {
val bytesRead = runAndReturnBytesRead {
sc.newAPIHadoopFile(tmpFilePath, classOf[NewTextInputFormat], classOf[LongWritable],
classOf[Text]).count()
}
val bytesRead2 = runAndReturnBytesRead {
sc.newAPIHadoopFile(tmpFilePath, classOf[NewTextInputFormat], classOf[LongWritable],
classOf[Text]).coalesce(5).count()
}
assert(bytesRead != 0)
assert(bytesRead2 == bytesRead)
assert(bytesRead >= tmpFile.length())
}
test("input metrics when reading text file") {
val bytesRead = runAndReturnBytesRead {
sc.textFile(tmpFilePath, 2).count()
}
assert(bytesRead >= tmpFile.length())
}
test("input metrics on records read - simple") {
val records = runAndReturnRecordsRead {
sc.textFile(tmpFilePath, 4).count()
}
assert(records == numRecords)
}
test("input metrics on records read - more stages") {
val records = runAndReturnRecordsRead {
sc.textFile(tmpFilePath, 4)
.map(key => (key.length, 1))
.reduceByKey(_ + _)
.count()
}
assert(records == numRecords)
}
test("input metrics on records - New Hadoop API") {
val records = runAndReturnRecordsRead {
sc.newAPIHadoopFile(tmpFilePath, classOf[NewTextInputFormat], classOf[LongWritable],
classOf[Text]).count()
}
assert(records == numRecords)
}
test("input metrics on recordsd read with cache") {
// prime the cache manager
val rdd = sc.textFile(tmpFilePath, 4).cache()
rdd.collect()
val records = runAndReturnRecordsRead {
rdd.count()
}
assert(records == numRecords)
}
test("shuffle records read metrics") {
val recordsRead = runAndReturnShuffleRecordsRead {
sc.textFile(tmpFilePath, 4)
.map(key => (key, 1))
.groupByKey()
.collect()
}
assert(recordsRead == numRecords)
}
test("shuffle records written metrics") {
val recordsWritten = runAndReturnShuffleRecordsWritten {
sc.textFile(tmpFilePath, 4)
.map(key => (key, 1))
.groupByKey()
.collect()
}
assert(recordsWritten == numRecords)
}
/**
* Tests the metrics from end to end.
* 1) reading a hadoop file
* 2) shuffle and writing to a hadoop file.
* 3) writing to hadoop file.
*/
test("input read/write and shuffle read/write metrics all line up") {
var inputRead = 0L
var outputWritten = 0L
var shuffleRead = 0L
var shuffleWritten = 0L
sc.addSparkListener(new SparkListener() {
override def onTaskEnd(taskEnd: SparkListenerTaskEnd) {
val metrics = taskEnd.taskMetrics
metrics.inputMetrics.foreach(inputRead += _.recordsRead)
metrics.outputMetrics.foreach(outputWritten += _.recordsWritten)
metrics.shuffleReadMetrics.foreach(shuffleRead += _.recordsRead)
metrics.shuffleWriteMetrics.foreach(shuffleWritten += _.shuffleRecordsWritten)
}
})
val tmpFile = new File(tmpDir, getClass.getSimpleName)
sc.textFile(tmpFilePath, 4)
.map(key => (key, 1))
.reduceByKey(_ + _)
.saveAsTextFile("file://" + tmpFile.getAbsolutePath)
sc.listenerBus.waitUntilEmpty(500)
assert(inputRead == numRecords)
// Only supported on newer Hadoop
if (SparkHadoopUtil.get.getFSBytesWrittenOnThreadCallback().isDefined) {
assert(outputWritten == numBuckets)
}
assert(shuffleRead == shuffleWritten)
}
test("input metrics with interleaved reads") {
val numPartitions = 2
val cartVector = 0 to 9
val cartFile = new File(tmpDir, getClass.getSimpleName + "_cart.txt")
val cartFilePath = "file://" + cartFile.getAbsolutePath
// write files to disk so we can read them later.
sc.parallelize(cartVector).saveAsTextFile(cartFilePath)
val aRdd = sc.textFile(cartFilePath, numPartitions)
val tmpRdd = sc.textFile(tmpFilePath, numPartitions)
val firstSize = runAndReturnBytesRead {
aRdd.count()
}
val secondSize = runAndReturnBytesRead {
tmpRdd.count()
}
val cartesianBytes = runAndReturnBytesRead {
aRdd.cartesian(tmpRdd).count()
}
// Computing the amount of bytes read for a cartesian operation is a little involved.
// Cartesian interleaves reads between two partitions eg. p1 and p2.
// Here are the steps:
// 1) First it creates an iterator for p1
// 2) Creates an iterator for p2
// 3) Reads the first element of p1 and then all the elements of p2
// 4) proceeds to the next element of p1
// 5) Creates a new iterator for p2
// 6) rinse and repeat.
// As a result we read from the second partition n times where n is the number of keys in
// p1. Thus the math below for the test.
assert(cartesianBytes != 0)
assert(cartesianBytes == firstSize * numPartitions + (cartVector.length * secondSize))
}
private def runAndReturnBytesRead(job: => Unit): Long = {
runAndReturnMetrics(job, _.taskMetrics.inputMetrics.map(_.bytesRead))
}
private def runAndReturnRecordsRead(job: => Unit): Long = {
runAndReturnMetrics(job, _.taskMetrics.inputMetrics.map(_.recordsRead))
}
private def runAndReturnRecordsWritten(job: => Unit): Long = {
runAndReturnMetrics(job, _.taskMetrics.outputMetrics.map(_.recordsWritten))
}
private def runAndReturnShuffleRecordsRead(job: => Unit): Long = {
runAndReturnMetrics(job, _.taskMetrics.shuffleReadMetrics.map(_.recordsRead))
}
private def runAndReturnShuffleRecordsWritten(job: => Unit): Long = {
runAndReturnMetrics(job, _.taskMetrics.shuffleWriteMetrics.map(_.shuffleRecordsWritten))
}
private def runAndReturnMetrics(job: => Unit,
collector: (SparkListenerTaskEnd) => Option[Long]): Long = {
val taskMetrics = new ArrayBuffer[Long]()
sc.addSparkListener(new SparkListener() {
override def onTaskEnd(taskEnd: SparkListenerTaskEnd) {
collector(taskEnd).foreach(taskMetrics += _)
}
})
job
sc.listenerBus.waitUntilEmpty(500)
taskMetrics.sum
}
test("output metrics on records written") {
// Only supported on newer Hadoop
if (SparkHadoopUtil.get.getFSBytesWrittenOnThreadCallback().isDefined) {
val file = new File(tmpDir, getClass.getSimpleName)
val filePath = "file://" + file.getAbsolutePath
val records = runAndReturnRecordsWritten {
sc.parallelize(1 to numRecords).saveAsTextFile(filePath)
}
assert(records == numRecords)
}
}
test("output metrics on records written - new Hadoop API") {
// Only supported on newer Hadoop
if (SparkHadoopUtil.get.getFSBytesWrittenOnThreadCallback().isDefined) {
val file = new File(tmpDir, getClass.getSimpleName)
val filePath = "file://" + file.getAbsolutePath
val records = runAndReturnRecordsWritten {
sc.parallelize(1 to numRecords).map(key => (key.toString, key.toString))
.saveAsNewAPIHadoopFile[NewTextOutputFormat[String, String]](filePath)
}
assert(records == numRecords)
}
}
test("output metrics when writing text file") {
val fs = FileSystem.getLocal(new Configuration())
val outPath = new Path(fs.getWorkingDirectory, "outdir")
if (SparkHadoopUtil.get.getFSBytesWrittenOnThreadCallback().isDefined) {
val taskBytesWritten = new ArrayBuffer[Long]()
sc.addSparkListener(new SparkListener() {
override def onTaskEnd(taskEnd: SparkListenerTaskEnd) {
taskBytesWritten += taskEnd.taskMetrics.outputMetrics.get.bytesWritten
}
})
val rdd = sc.parallelize(Array("a", "b", "c", "d"), 2)
try {
rdd.saveAsTextFile(outPath.toString)
sc.listenerBus.waitUntilEmpty(500)
assert(taskBytesWritten.length == 2)
val outFiles = fs.listStatus(outPath).filter(_.getPath.getName != "_SUCCESS")
taskBytesWritten.zip(outFiles).foreach { case (bytes, fileStatus) =>
assert(bytes >= fileStatus.getLen)
}
} finally {
fs.delete(outPath, true)
}
}
}
test("input metrics with old CombineFileInputFormat") {
val bytesRead = runAndReturnBytesRead {
sc.hadoopFile(tmpFilePath, classOf[OldCombineTextInputFormat], classOf[LongWritable],
classOf[Text], 2).count()
}
assert(bytesRead >= tmpFile.length())
}
test("input metrics with new CombineFileInputFormat") {
val bytesRead = runAndReturnBytesRead {
sc.newAPIHadoopFile(tmpFilePath, classOf[NewCombineTextInputFormat], classOf[LongWritable],
classOf[Text], new Configuration()).count()
}
assert(bytesRead >= tmpFile.length())
}
}
/**
* Hadoop 2 has a version of this, but we can't use it for backwards compatibility
*/
class OldCombineTextInputFormat extends OldCombineFileInputFormat[LongWritable, Text] {
override def getRecordReader(split: OldInputSplit, conf: JobConf, reporter: Reporter)
: OldRecordReader[LongWritable, Text] = {
new OldCombineFileRecordReader[LongWritable, Text](conf,
split.asInstanceOf[OldCombineFileSplit], reporter, classOf[OldCombineTextRecordReaderWrapper]
.asInstanceOf[Class[OldRecordReader[LongWritable, Text]]])
}
}
class OldCombineTextRecordReaderWrapper(
split: OldCombineFileSplit,
conf: Configuration,
reporter: Reporter,
idx: Integer) extends OldRecordReader[LongWritable, Text] {
val fileSplit = new OldFileSplit(split.getPath(idx),
split.getOffset(idx),
split.getLength(idx),
split.getLocations())
val delegate: OldLineRecordReader = new OldTextInputFormat().getRecordReader(fileSplit,
conf.asInstanceOf[JobConf], reporter).asInstanceOf[OldLineRecordReader]
override def next(key: LongWritable, value: Text): Boolean = delegate.next(key, value)
override def createKey(): LongWritable = delegate.createKey()
override def createValue(): Text = delegate.createValue()
override def getPos(): Long = delegate.getPos
override def close(): Unit = delegate.close()
override def getProgress(): Float = delegate.getProgress
}
/**
* Hadoop 2 has a version of this, but we can't use it for backwards compatibility
*/
class NewCombineTextInputFormat extends NewCombineFileInputFormat[LongWritable, Text] {
def createRecordReader(split: NewInputSplit, context: TaskAttemptContext)
: NewRecordReader[LongWritable, Text] = {
new NewCombineFileRecordReader[LongWritable, Text](split.asInstanceOf[NewCombineFileSplit],
context, classOf[NewCombineTextRecordReaderWrapper])
}
}
class NewCombineTextRecordReaderWrapper(
split: NewCombineFileSplit,
context: TaskAttemptContext,
idx: Integer) extends NewRecordReader[LongWritable, Text] {
val fileSplit = new NewFileSplit(split.getPath(idx),
split.getOffset(idx),
split.getLength(idx),
split.getLocations())
val delegate = new NewTextInputFormat().createRecordReader(fileSplit, context)
override def initialize(split: NewInputSplit, context: TaskAttemptContext): Unit = {
delegate.initialize(fileSplit, context)
}
override def nextKeyValue(): Boolean = delegate.nextKeyValue()
override def getCurrentKey(): LongWritable = delegate.getCurrentKey
override def getCurrentValue(): Text = delegate.getCurrentValue
override def getProgress(): Float = delegate.getProgress
override def close(): Unit = delegate.close()
}
|