Spark优化一则 - 减少Shuffle
看了Spark Summit 2014的A Deeper Understanding of Spark Internals,视频(要***)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(_(1)) 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(0), 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 | 2014/09/03 20:51:58 | 3 s | 8/8 | 140.2 MB | ||
2 | 2014/09/03 20:51:55 | 2 s | 8/8 | 208.4 MB | ||
0 | 2014/09/03 20:51:46 | 8 s | 8/8 |
Slides优化方案
rdd.distinct(numPartitions = 6).map(x => (x.charAt(0), 1)).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 | 2014/09/03 20:24:17 | 0.2 s | 6/6 | 4.9 KB | ||
2 | 2014/09/03 20:24:15 | 2 s | 6/6 | 50.4 MB | 7.4 KB | |
3 | 2014/09/03 20:24:13 | 2 s | 8/8 | 75.6 MB | ||
0 | 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(0), 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 | 2014/09/03 20:45:02 | 0.3 s | 8/8 | |||
0 | 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。