-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmapreduce.html
583 lines (373 loc) · 12.6 KB
/
mapreduce.html
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
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
<!DOCTYPE html>
<html>
<head>
<title>Title</title>
<meta charset="utf-8">
<style>
@import url(https://fonts.googleapis.com/css?family=Yanone+Kaffeesatz);
@import url(https://fonts.googleapis.com/css?family=Droid+Serif:400,700,400italic);
@import url(https://fonts.googleapis.com/css?family=Ubuntu+Mono:400,700,400italic);
body { font-family: 'Droid Serif'; }
h1, h2, h3 {
font-family: 'Yanone Kaffeesatz';
font-weight: normal;
}
th, td {
padding: 5px;
border: 1px solid black;
}
.remark-code, .remark-inline-code { font-family: 'Ubuntu Mono'; }
</style>
</head>
<body>
<textarea id="source">
class: center, middle
# MapReduce
---
# MapReduce: Simplified Data Processing on Large Clusters
### Authors: Jeffrey Dean, Sanjay Ghemawat
### Presenter: Neil Seward
---
# Contents
1. Defining Google's Problem
2. The MapReduce Model
3. Implementation of MapReduce
4. Fault Tolerance
5. Additions to the MapReduce Model
6. Evaluating Performance
7. Conclusions
---
# Defining Google's Problem
- Google needed a new method to work with all of their indexed data.
---
# Defining Google's Problem
- The new searching algorithm needed to work across a distributed set of thousands of computers.
<br>
<br>
- These networked computers were considered to be mid-low tier in processing power.
---
# Defining Google's Problem
The algorithm needed to be relatively simple to implement and understand, but still incorporate for:
- fault tolerance
- slow network bandwith
- hanging processes
---
# The MapReduce Model
- The MapReduce algorithm takes in a set of key/value pairs as input and produces a different set of key/value pairs as output.
### Input
Key | Value
------------- | -------------
words.txt | {Banana, Banana}
### Output
Key | Value
------------- | -------------
Banana | 2
---
# The MapReduce Model
The MapReduce algorithm has two main functions:
- Map
- Reduce
<br>
<br>
Both the Map function and the Reduce function can be written differently, depending on the user requirements of the MapReduce function.
---
# The MapReduce Model
### Map
- The Map function takes a document with a list of values, and emits a transitional key/value pair.
<br>
<br>
<img src="img/map-function.png" style="width: 600px; height: 250px"/>
---
# The MapReduce Model
## Map Example
### Input
Key | Value
------------- | -------------
words.txt | Banana, Banana
### Output
Key | Value
------------- | -------------
Banana | 1
Banana | 1
---
# The MapReduce Model
### Reduce
- The Reduce function accepts an intermediate key from the Map function along with the set of values for that key, and emits a subset of the given values.
<br>
<br>
<img src="img/reduce-function.png" style="width: 600px; height: 250px"/>
---
# The MapReduce Model
## Reduce Example
### Input
Key | Value
------------- | -------------
Banana | {1, 1}
### Output
Key | Value
------------- | -------------
Banana | 2
---
# MapReduce Implementation
### Components
- The MapReduce model uses two main components to carry out commands and to load files for input and to store output.
---
# MapReduce Implementation
### Components
<img src="img/master-worker.png" style="width: 700px; height: 450px"/>
---
# MapReduce Implementation
### Components - Master
Tasks include:
- assign Map and Reduce tasks to workers
- tracks worker status
- tracks local file locations from map workers
- handles faults in workers.
---
# MapReduce Implementation
### Components - Worker
Tasks include:
- performs Map and Reduce tasks
- sends updates to master
- performs read and write operations to files.
---
# MapReduce Implementation
### Specifying Size - Mapping
- Mapping tasks are divided up into M segments of input data.
- This is used to optimize the size of the input data.
---
# MapReduce Implementation
### Specifying Size - Reduce
- Reduce tasks read from R intermediary files produced by the partitioning function.
---
# MapReduce Implementation
### Splitting
- The MapReduce library will first split the input files into M smaller subsets.
<br>
<br>
<img allign="left" src="img/splitting-step.png" style="width: 500px; height: 350px"/>
[1]
---
# MapReduce Implementation
### Mapping
- Each worker assigned to mapping will read a subset input file and Map intermediate key/value pairs.
<br>
<br>
<img src="img/mapping-step.png" style="width: 400px; height: 300px"/>
[1]
---
# MapReduce Implementation
### Mapping
- Once mapped, these pairs are written as buffers into memory to be sorted.
<br>
<br>
<img src="img/map-to-file.png" style="width: 350px; height: 300px"/>
---
# MapReduce Implementation
### Shuffling
- The partitioning function read the pairs from memory, sorts the pairs and stores the sorted partitions into local storage.
- The locations of these buffered pairs are stored in the master and are passed to the Reduce workers.
<br>
<br>
<img src="img/shuffle-step.png" style="width: 300px; height: 270px"/>
[1]
---
# MapReduce Implementation
### Reducing
- The Reduce workers read the sorted pairs from disk and creates subsets from the pairs.
<br>
<br>
<img src="img/reduce-step.png" style="width: 330px; height: 300px"/>
[1]
---
# MapReduce Implementation
### Reducing
- Once sorted, the Reduce workers perform the Reduce function and write the output pairs to output files.
<br>
<br>
<img src="img/reduce-to-file.png" style="width: 330px; height: 300px"/>
---
# MapReduce Implementation
### Output
- After appending to output files, the Reduce workers signal the master component for completion.
<br>
<br>
- The master component then notifies the use program of completion, returning the results.
<br>
<br>
<img src="img/final-step.png" style="width: 300px; height: 270px"/>
[1]
---
# Fault Tolerance
The MapReduce library has many mechanisms that handle failure in operations and also handle hanging processes.
Some failures and bottlenecks that are handled by the library include:
- Worker failure
- Master failure
- Slow network bandwidth
- Straggler tasks
---
# Fault Tolerance
### Execution States
The master component tracks the status of each Map and Reduce worker throughout the execution of the functions.
<br>
<br>
These states include:
- idle
- in-progress
- completed
---
# Fault Tolerance
### Worker Failure
- If stuck in the same state for too long, the master resets the worker back to idle.
---
# Fault Tolerance
### Worker Failure - Map
- When a Map worker is determined to be in a failed state, the master resets the worker back to idle.
- The output from the Map worker is stored locally and cannot be accessed by the master, so the Map tasks are re-scheduled on another worker.
---
# Fault Tolerance
### Worker Failure - Map
- Due to the large numbers of workers, tasks can be re-scheduled with ease.
<br>
<br>
<img src="img/reschedule-fault.png" style="width: 500px; height: 350px"/>
---
# Fault Tolerance
### Worker Failure - Reduce
- When Reduce workers fail, they are not re-scheduled due to their output being globally accessible.
---
# Fault Tolerance
### Master Failure
- The master component is central to the execution of the MapReduce program.
- There are no copies of the master component.
- When in failed state, the master is aborted and the client program is notified.
---
# Fault Tolerance
### Network Bandwidth
- The MapReduce model is robust in that it can still perform the function operation on a network with low bandwidth.
- The input files for the workers are initially stored as copies across the cluster of computers.
---
# Fault Tolerance
### Network Bandwidth - Locality
- When a worker fails, the master re-schedules jobs based on the closest machine that has a copy of the original worker data.
<img src="img/locality.png" style="width: 500px; height: 350px"/>
[3]
---
# MapReduce Additions
### Backup Tasks
- When a worker is taking a long time to complete a Map or a Reduce function while in-progress, the master component considers it a straggler.
- After straggler detection, the master creates backup executions of the in-progress task.
---
# MapReduce Additions
### Backup Tasks
- The task is determined to be complete when either the backup task or the original task completes the operation.
<br>
<br>
<img src="img/backup-task.png" style="width: 500px; height: 300px"/>
[2]
---
# MapReduce Additions
### Partitioning
- When creating inputs for Reduce tasks, a Partition function can be used to evenly distribute partitions.
- This is the shuffling step of MapReduce.
- *partition = hash(key)**mod**R*
---
# MapReduce Additions
### Combining
- After Map tasks are finished, a simple combination can be performed after the Map task on the same worker to simplify the Reduce task load.
- The combination function works similar to the Reduce function.
<br>
<br>
<img src="img/combiner-function.png" style="width: 450px; height: 200px"/>
---
# MapReduce Performance
### Configuration
- 1800 machines on network.
- Two 2GHz Xeon CPUs.
- 4GB RAM.
- Two 160GB IDE storage.
- 100-200 Gbps network bandwidth.
- 1 TB of data.
---
# MapReduce Performance
### Grep - Summary
- Searches through data for rare three letter combinations.
- Input data is split into 64MB peices (M = 15,000).
- Output data is single file (R = 1).
---
# MapReduce Performance
### Grep - Performance
- The entire computation takes 150 seconds (including 60 seconds of startup).
<img src="img/data-rate.png" style="width: 500px; height: 300px"/>
---
# MapReduce Performance
### Sort - Summary
- Sorts data based on extracted keys from data.
- Input data is split into 64MB peices (M = 15,000).
- Output data is written to 4000 files (R = 4000).
---
# MapReduce Performance
### Sort - Performance (Normal, Input Rate)
- This graph shows the rate at which the data is read from input.
- All of the reading is completed at around 200 seconds in.
<img src="img/data-rate-sort-normal.png" style="width: 500px; height: 300px"/>
---
# MapReduce Performance
### Sort - Performance (Normal, Network Transfer Rate)
- This graph shows the rate at which the data is transfered from the Map phase to the Reduce phase.
- This happens specifically during the shuffling/partitioning phase.
- All of the data is transfered at around 600 seconds in.
<img src="img/network-rate-sort-normal.png" style="width: 500px; height: 270px"/>
---
# MapReduce Performance
### Sort - Performance (Normal, Output Rate)
- This graph shows the rate at which the data is written to the output files.
- All of the data is written at around 850 seconds in.
- With startup, this version of sort takes 891 seconds.
<img src="img/output-rate-sort-normal.png" style="width: 500px; height: 270px"/>
---
# MapReduce Performance
### Sort - Performance (No Backup, Input Rate)
<img src="img/data-rate-sort-backup.png" style="width: 500px; height: 300px"/>
---
# MapReduce Performance
### Sort - Performance (No Backup, Network Transfer Rate)
<img src="img/network-rate-sort-backup.png" style="width: 500px; height: 270px"/>
---
# MapReduce Performance
### Sort - Performance (No Backup, Output Rate)
- Most of the data is written before 960 seconds in.
- Without backup tasks, there are five stragglers that cause the end point to be at 1283 seconds.
<img src="img/output-rate-sort-backup.png" style="width: 500px; height: 270px"/>
---
# MapReduce Performance
### Sort - Performance (200 Task Killed, Input Rate)
- The negative rate is due to information loss during re-scheduling failed tasks.
<img src="img/data-rate-sort-fault.png" style="width: 500px; height: 300px"/>
---
# MapReduce Performance
### Sort - Performance (200 Task Killed, Network Transfer Rate)
<img src="img/network-rate-sort-fault.png" style="width: 500px; height: 270px"/>
---
# MapReduce Performance
### Sort - Performance (200 Task Killed, Output Rate)
- With startup, this configuration of sort completes at 933 seconds in.
<img src="img/output-rate-sort-fault.png" style="width: 500px; height: 270px"/>
---
# Conclusions
- With the MapReduce library, making simple queries on large sets of data in short periods is very easy to implement.
- The additions to the model, such as combining, backup tasks, and network locality make the distributed library even more powerful in terms of processing.
---
# References
1. http://xiaochongzhang.me/blog/wp-content/uploads/2013/05/MapReduce_Work_Structure.png
2. http://images.slideplayer.com/14/4413895/slides/slide_13.jpg
3. http://www.sis.pitt.edu/bpalan/research.html
</textarea>
<script src="https://gnab.github.io/remark/downloads/remark-latest.min.js">
</script>
<script>
var slideshow = remark.create();
</script>
</body>
</html>