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
|
/*
* 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.shuffle.unsafe;
import java.io.*;
import java.nio.channels.FileChannel;
import java.util.Iterator;
import javax.annotation.Nullable;
import scala.Option;
import scala.Product2;
import scala.collection.JavaConversions;
import scala.reflect.ClassTag;
import scala.reflect.ClassTag$;
import com.google.common.annotations.VisibleForTesting;
import com.google.common.io.ByteStreams;
import com.google.common.io.Closeables;
import com.google.common.io.Files;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
import org.apache.spark.*;
import org.apache.spark.annotation.Private;
import org.apache.spark.io.CompressionCodec;
import org.apache.spark.io.CompressionCodec$;
import org.apache.spark.io.LZFCompressionCodec;
import org.apache.spark.executor.ShuffleWriteMetrics;
import org.apache.spark.network.util.LimitedInputStream;
import org.apache.spark.scheduler.MapStatus;
import org.apache.spark.scheduler.MapStatus$;
import org.apache.spark.serializer.SerializationStream;
import org.apache.spark.serializer.Serializer;
import org.apache.spark.serializer.SerializerInstance;
import org.apache.spark.shuffle.IndexShuffleBlockResolver;
import org.apache.spark.shuffle.ShuffleMemoryManager;
import org.apache.spark.shuffle.ShuffleWriter;
import org.apache.spark.storage.BlockManager;
import org.apache.spark.storage.TimeTrackingOutputStream;
import org.apache.spark.unsafe.PlatformDependent;
import org.apache.spark.unsafe.memory.TaskMemoryManager;
@Private
public class UnsafeShuffleWriter<K, V> extends ShuffleWriter<K, V> {
private final Logger logger = LoggerFactory.getLogger(UnsafeShuffleWriter.class);
private static final ClassTag<Object> OBJECT_CLASS_TAG = ClassTag$.MODULE$.Object();
@VisibleForTesting
static final int INITIAL_SORT_BUFFER_SIZE = 4096;
private final BlockManager blockManager;
private final IndexShuffleBlockResolver shuffleBlockResolver;
private final TaskMemoryManager memoryManager;
private final ShuffleMemoryManager shuffleMemoryManager;
private final SerializerInstance serializer;
private final Partitioner partitioner;
private final ShuffleWriteMetrics writeMetrics;
private final int shuffleId;
private final int mapId;
private final TaskContext taskContext;
private final SparkConf sparkConf;
private final boolean transferToEnabled;
private MapStatus mapStatus = null;
private UnsafeShuffleExternalSorter sorter = null;
/** Subclass of ByteArrayOutputStream that exposes `buf` directly. */
private static final class MyByteArrayOutputStream extends ByteArrayOutputStream {
public MyByteArrayOutputStream(int size) { super(size); }
public byte[] getBuf() { return buf; }
}
private MyByteArrayOutputStream serBuffer;
private SerializationStream serOutputStream;
/**
* Are we in the process of stopping? Because map tasks can call stop() with success = true
* and then call stop() with success = false if they get an exception, we want to make sure
* we don't try deleting files, etc twice.
*/
private boolean stopping = false;
public UnsafeShuffleWriter(
BlockManager blockManager,
IndexShuffleBlockResolver shuffleBlockResolver,
TaskMemoryManager memoryManager,
ShuffleMemoryManager shuffleMemoryManager,
UnsafeShuffleHandle<K, V> handle,
int mapId,
TaskContext taskContext,
SparkConf sparkConf) throws IOException {
final int numPartitions = handle.dependency().partitioner().numPartitions();
if (numPartitions > UnsafeShuffleManager.MAX_SHUFFLE_OUTPUT_PARTITIONS()) {
throw new IllegalArgumentException(
"UnsafeShuffleWriter can only be used for shuffles with at most " +
UnsafeShuffleManager.MAX_SHUFFLE_OUTPUT_PARTITIONS() + " reduce partitions");
}
this.blockManager = blockManager;
this.shuffleBlockResolver = shuffleBlockResolver;
this.memoryManager = memoryManager;
this.shuffleMemoryManager = shuffleMemoryManager;
this.mapId = mapId;
final ShuffleDependency<K, V, V> dep = handle.dependency();
this.shuffleId = dep.shuffleId();
this.serializer = Serializer.getSerializer(dep.serializer()).newInstance();
this.partitioner = dep.partitioner();
this.writeMetrics = new ShuffleWriteMetrics();
taskContext.taskMetrics().shuffleWriteMetrics_$eq(Option.apply(writeMetrics));
this.taskContext = taskContext;
this.sparkConf = sparkConf;
this.transferToEnabled = sparkConf.getBoolean("spark.file.transferTo", true);
open();
}
/**
* This convenience method should only be called in test code.
*/
@VisibleForTesting
public void write(Iterator<Product2<K, V>> records) throws IOException {
write(JavaConversions.asScalaIterator(records));
}
@Override
public void write(scala.collection.Iterator<Product2<K, V>> records) throws IOException {
boolean success = false;
try {
while (records.hasNext()) {
insertRecordIntoSorter(records.next());
}
closeAndWriteOutput();
success = true;
} finally {
if (!success) {
sorter.cleanupAfterError();
}
}
}
private void open() throws IOException {
assert (sorter == null);
sorter = new UnsafeShuffleExternalSorter(
memoryManager,
shuffleMemoryManager,
blockManager,
taskContext,
INITIAL_SORT_BUFFER_SIZE,
partitioner.numPartitions(),
sparkConf,
writeMetrics);
serBuffer = new MyByteArrayOutputStream(1024 * 1024);
serOutputStream = serializer.serializeStream(serBuffer);
}
@VisibleForTesting
void closeAndWriteOutput() throws IOException {
serBuffer = null;
serOutputStream = null;
final SpillInfo[] spills = sorter.closeAndGetSpills();
sorter = null;
final long[] partitionLengths;
try {
partitionLengths = mergeSpills(spills);
} finally {
for (SpillInfo spill : spills) {
if (spill.file.exists() && ! spill.file.delete()) {
logger.error("Error while deleting spill file {}", spill.file.getPath());
}
}
}
shuffleBlockResolver.writeIndexFile(shuffleId, mapId, partitionLengths);
mapStatus = MapStatus$.MODULE$.apply(blockManager.shuffleServerId(), partitionLengths);
}
@VisibleForTesting
void insertRecordIntoSorter(Product2<K, V> record) throws IOException {
final K key = record._1();
final int partitionId = partitioner.getPartition(key);
serBuffer.reset();
serOutputStream.writeKey(key, OBJECT_CLASS_TAG);
serOutputStream.writeValue(record._2(), OBJECT_CLASS_TAG);
serOutputStream.flush();
final int serializedRecordSize = serBuffer.size();
assert (serializedRecordSize > 0);
sorter.insertRecord(
serBuffer.getBuf(), PlatformDependent.BYTE_ARRAY_OFFSET, serializedRecordSize, partitionId);
}
@VisibleForTesting
void forceSorterToSpill() throws IOException {
assert (sorter != null);
sorter.spill();
}
/**
* Merge zero or more spill files together, choosing the fastest merging strategy based on the
* number of spills and the IO compression codec.
*
* @return the partition lengths in the merged file.
*/
private long[] mergeSpills(SpillInfo[] spills) throws IOException {
final File outputFile = shuffleBlockResolver.getDataFile(shuffleId, mapId);
final boolean compressionEnabled = sparkConf.getBoolean("spark.shuffle.compress", true);
final CompressionCodec compressionCodec = CompressionCodec$.MODULE$.createCodec(sparkConf);
final boolean fastMergeEnabled =
sparkConf.getBoolean("spark.shuffle.unsafe.fastMergeEnabled", true);
final boolean fastMergeIsSupported =
!compressionEnabled || compressionCodec instanceof LZFCompressionCodec;
try {
if (spills.length == 0) {
new FileOutputStream(outputFile).close(); // Create an empty file
return new long[partitioner.numPartitions()];
} else if (spills.length == 1) {
// Here, we don't need to perform any metrics updates because the bytes written to this
// output file would have already been counted as shuffle bytes written.
Files.move(spills[0].file, outputFile);
return spills[0].partitionLengths;
} else {
final long[] partitionLengths;
// There are multiple spills to merge, so none of these spill files' lengths were counted
// towards our shuffle write count or shuffle write time. If we use the slow merge path,
// then the final output file's size won't necessarily be equal to the sum of the spill
// files' sizes. To guard against this case, we look at the output file's actual size when
// computing shuffle bytes written.
//
// We allow the individual merge methods to report their own IO times since different merge
// strategies use different IO techniques. We count IO during merge towards the shuffle
// shuffle write time, which appears to be consistent with the "not bypassing merge-sort"
// branch in ExternalSorter.
if (fastMergeEnabled && fastMergeIsSupported) {
// Compression is disabled or we are using an IO compression codec that supports
// decompression of concatenated compressed streams, so we can perform a fast spill merge
// that doesn't need to interpret the spilled bytes.
if (transferToEnabled) {
logger.debug("Using transferTo-based fast merge");
partitionLengths = mergeSpillsWithTransferTo(spills, outputFile);
} else {
logger.debug("Using fileStream-based fast merge");
partitionLengths = mergeSpillsWithFileStream(spills, outputFile, null);
}
} else {
logger.debug("Using slow merge");
partitionLengths = mergeSpillsWithFileStream(spills, outputFile, compressionCodec);
}
// When closing an UnsafeShuffleExternalSorter that has already spilled once but also has
// in-memory records, we write out the in-memory records to a file but do not count that
// final write as bytes spilled (instead, it's accounted as shuffle write). The merge needs
// to be counted as shuffle write, but this will lead to double-counting of the final
// SpillInfo's bytes.
writeMetrics.decShuffleBytesWritten(spills[spills.length - 1].file.length());
writeMetrics.incShuffleBytesWritten(outputFile.length());
return partitionLengths;
}
} catch (IOException e) {
if (outputFile.exists() && !outputFile.delete()) {
logger.error("Unable to delete output file {}", outputFile.getPath());
}
throw e;
}
}
/**
* Merges spill files using Java FileStreams. This code path is slower than the NIO-based merge,
* {@link UnsafeShuffleWriter#mergeSpillsWithTransferTo(SpillInfo[], File)}, so it's only used in
* cases where the IO compression codec does not support concatenation of compressed data, or in
* cases where users have explicitly disabled use of {@code transferTo} in order to work around
* kernel bugs.
*
* @param spills the spills to merge.
* @param outputFile the file to write the merged data to.
* @param compressionCodec the IO compression codec, or null if shuffle compression is disabled.
* @return the partition lengths in the merged file.
*/
private long[] mergeSpillsWithFileStream(
SpillInfo[] spills,
File outputFile,
@Nullable CompressionCodec compressionCodec) throws IOException {
assert (spills.length >= 2);
final int numPartitions = partitioner.numPartitions();
final long[] partitionLengths = new long[numPartitions];
final InputStream[] spillInputStreams = new FileInputStream[spills.length];
OutputStream mergedFileOutputStream = null;
boolean threwException = true;
try {
for (int i = 0; i < spills.length; i++) {
spillInputStreams[i] = new FileInputStream(spills[i].file);
}
for (int partition = 0; partition < numPartitions; partition++) {
final long initialFileLength = outputFile.length();
mergedFileOutputStream =
new TimeTrackingOutputStream(writeMetrics, new FileOutputStream(outputFile, true));
if (compressionCodec != null) {
mergedFileOutputStream = compressionCodec.compressedOutputStream(mergedFileOutputStream);
}
for (int i = 0; i < spills.length; i++) {
final long partitionLengthInSpill = spills[i].partitionLengths[partition];
if (partitionLengthInSpill > 0) {
InputStream partitionInputStream =
new LimitedInputStream(spillInputStreams[i], partitionLengthInSpill);
if (compressionCodec != null) {
partitionInputStream = compressionCodec.compressedInputStream(partitionInputStream);
}
ByteStreams.copy(partitionInputStream, mergedFileOutputStream);
}
}
mergedFileOutputStream.flush();
mergedFileOutputStream.close();
partitionLengths[partition] = (outputFile.length() - initialFileLength);
}
threwException = false;
} finally {
// To avoid masking exceptions that caused us to prematurely enter the finally block, only
// throw exceptions during cleanup if threwException == false.
for (InputStream stream : spillInputStreams) {
Closeables.close(stream, threwException);
}
Closeables.close(mergedFileOutputStream, threwException);
}
return partitionLengths;
}
/**
* Merges spill files by using NIO's transferTo to concatenate spill partitions' bytes.
* This is only safe when the IO compression codec and serializer support concatenation of
* serialized streams.
*
* @return the partition lengths in the merged file.
*/
private long[] mergeSpillsWithTransferTo(SpillInfo[] spills, File outputFile) throws IOException {
assert (spills.length >= 2);
final int numPartitions = partitioner.numPartitions();
final long[] partitionLengths = new long[numPartitions];
final FileChannel[] spillInputChannels = new FileChannel[spills.length];
final long[] spillInputChannelPositions = new long[spills.length];
FileChannel mergedFileOutputChannel = null;
boolean threwException = true;
try {
for (int i = 0; i < spills.length; i++) {
spillInputChannels[i] = new FileInputStream(spills[i].file).getChannel();
}
// This file needs to opened in append mode in order to work around a Linux kernel bug that
// affects transferTo; see SPARK-3948 for more details.
mergedFileOutputChannel = new FileOutputStream(outputFile, true).getChannel();
long bytesWrittenToMergedFile = 0;
for (int partition = 0; partition < numPartitions; partition++) {
for (int i = 0; i < spills.length; i++) {
final long partitionLengthInSpill = spills[i].partitionLengths[partition];
long bytesToTransfer = partitionLengthInSpill;
final FileChannel spillInputChannel = spillInputChannels[i];
final long writeStartTime = System.nanoTime();
while (bytesToTransfer > 0) {
final long actualBytesTransferred = spillInputChannel.transferTo(
spillInputChannelPositions[i],
bytesToTransfer,
mergedFileOutputChannel);
spillInputChannelPositions[i] += actualBytesTransferred;
bytesToTransfer -= actualBytesTransferred;
}
writeMetrics.incShuffleWriteTime(System.nanoTime() - writeStartTime);
bytesWrittenToMergedFile += partitionLengthInSpill;
partitionLengths[partition] += partitionLengthInSpill;
}
}
// Check the position after transferTo loop to see if it is in the right position and raise an
// exception if it is incorrect. The position will not be increased to the expected length
// after calling transferTo in kernel version 2.6.32. This issue is described at
// https://bugs.openjdk.java.net/browse/JDK-7052359 and SPARK-3948.
if (mergedFileOutputChannel.position() != bytesWrittenToMergedFile) {
throw new IOException(
"Current position " + mergedFileOutputChannel.position() + " does not equal expected " +
"position " + bytesWrittenToMergedFile + " after transferTo. Please check your kernel" +
" version to see if it is 2.6.32, as there is a kernel bug which will lead to " +
"unexpected behavior when using transferTo. You can set spark.file.transferTo=false " +
"to disable this NIO feature."
);
}
threwException = false;
} finally {
// To avoid masking exceptions that caused us to prematurely enter the finally block, only
// throw exceptions during cleanup if threwException == false.
for (int i = 0; i < spills.length; i++) {
assert(spillInputChannelPositions[i] == spills[i].file.length());
Closeables.close(spillInputChannels[i], threwException);
}
Closeables.close(mergedFileOutputChannel, threwException);
}
return partitionLengths;
}
@Override
public Option<MapStatus> stop(boolean success) {
try {
if (stopping) {
return Option.apply(null);
} else {
stopping = true;
if (success) {
if (mapStatus == null) {
throw new IllegalStateException("Cannot call stop(true) without having called write()");
}
return Option.apply(mapStatus);
} else {
// The map task failed, so delete our output data.
shuffleBlockResolver.removeDataByMap(shuffleId, mapId);
return Option.apply(null);
}
}
} finally {
if (sorter != null) {
// If sorter is non-null, then this implies that we called stop() in response to an error,
// so we need to clean up memory and spill files created by the sorter
sorter.cleanupAfterError();
}
}
}
}
|