Spark优化一则 - 减少Shuffle

看了Spark Summit 2014的A Deeper Understanding of Spark Internals，视频（要科学上网）详细讲解了Spark的工作原理，Slides的45页给原始算法和优化算法。

破砂锅用自己3节点的Spark集群试验了这个优化算法，并进一步找到更快的算法。测试数据是Sogou实验室的日志文件前10000000条数据。目标是对日志第2列数据，按照第一个字母合并，得到每个首字母有几条记录。

所有的方案都重新启动Spark shell，先用以下代码把日志第2列数据cache到内存里，Spark GUI显示cache有8个partition，约1GB内存。

val rdd = sc.textFile("hdfs://hadoop1:8000/input/SogouQ3.txt").map(_.split("\t")).map(_())

rdd.cache()

rdd.count()

// res1: Long = 10000000

Spark GUI

RDD Name	Storage Level	Cached Partitions	Fraction Cached	Size in Memory	Size in Tachyon	Size on Disk
3	Memory Deserialized 1x Replicated	8	100%	1089.4 MB	0.0 B	0.0 B

Slides原始方案

rdd.map(x => (x.charAt(), x)).groupByKey().mapValues({x => x.toSet.size}).collect()

// res2: Array[(Char, Int)] = Array((8,168189), (0,168338), (a,168228), (9,168018), (1,167647), (b,168404), (2,168731), (3,168206), (c,168991), (d,168095), (4,167523), (e,168179), (5,167967), (6,167907), (f,168174), (7,168718))

Spark stage GUI显示有关stage Id是1-2，累计耗时5s，产生140MB shuffle read和208MB shuffle write。

Stage Id	Description	Submitted	Duration	Tasks: Succeeded/Total	Shuffle Read	Shuffle Write
1	collect at <console>:15	2014/09/03 20:51:58	3 s	8/8	140.2 MB
2	map at <console>:15	2014/09/03 20:51:55	2 s	8/8		208.4 MB
0	count at <console>:15	2014/09/03 20:51:46	8 s	8/8

Slides优化方案

rdd.distinct(numPartitions = ).map(x => (x.charAt(), )).reduceByKey(_+_).collect()

// res2: Array[(Char, Int)] = Array((6,167907), (0,168338), (f,168174), (7,168718), (a,168228), (1,167647), (8,168189), (b,168404), (2,168731), (9,168018), (3,168206), (c,168991), (d,168095), (4,167523), (e,168179), (5,167967))

Spark stage GUI显示有关stage Id是1-3，累计耗时4.2s，生成50MB shuffle read和75MB shuffle write。虽然多了1个stage，shuffle read/write比原始方案减少超过60%，从而速度加快16%。

Stage Id	Description	Submitted	Duration	Tasks: Succeeded/Total	Shuffle Read	Shuffle Write
1	collect at <console>:15	2014/09/03 20:24:17	0.2 s	6/6	4.9 KB
2	reduceByKey at <console>:15	2014/09/03 20:24:15	2 s	6/6	50.4 MB	7.4 KB
3	distinct at <console>:15	2014/09/03 20:24:13	2 s	8/8		75.6 MB
0	count at <console>:15	2014/09/03 20:23:55	7 s	8/8

Zero Shuffle优化方案

既然减少shuffle可以加快速度，破砂锅想出以下的Zero Shuffle方案来。

rdd.map(x => (x.charAt(), x)).countByKey()

// res2: scala.collection.Map[Char,Long] = Map(e -> 623689, 2 -> 623914, 5 -> 619840, b -> 626111, 8 -> 620738, d -> 623515, 7 -> 620222, 1 -> 616184, 4 -> 616628, a -> 641623, c -> 630514, 6 -> 621346, f -> 624447, 0 -> 632735, 9 -> 637770, 3 -> 620724)

Spark stage GUI显示有关stage Id是1，累计耗时只有0.3s，没有shuffle read/write。这个方案有关的RDD只有narrow dependency，所以只有1个stage。

Stage Id	Description	Submitted	Duration	Tasks: Succeeded/Total	Shuffle Read	Shuffle Write
1	countByKey at <console>:15	2014/09/03 20:45:02	0.3 s	8/8
0	count at <console>:15	2014/09/03 20:44:32	8 s

小结

比较3种方案

方案	Shuffle Read	Shuffle Write	Time
Slides原始方案	140.2 MB	208.4 MB	5s
Slides优化方案	50.4 MB	75.6 MB	4.2s
Zero Shuffle优化方案	0	0	0.3s

Spark的优化之一是尽可能减少shuffle从而大幅减少缓慢的网络传输。熟悉RDD的函数对Spark优化有很大帮助。