Apache Spark 在 reduceByKey 步骤上运行缓慢

Question

我在 /usr/local/share/data/ 中有一个 2MB 的纯文本文件。 然后我在 Apache Spark 中运行以下代码。

conf = SparkConf().setMaster("local[*]").setAppName("test").set("spark.executor.memory", "2g") 
sc = SparkContext(conf=conf)
doc_rdd = sc.textFile("/usr/local/share/data/")
unigrams = doc_rdd.flatMap(word_tokenize)
step1 = unigrams.flatMap(word_pos_tagging)
step2 = step1.filter(lambda x: filter_punctuation(x[0]))
step3 = step2.map(lambda x: (x, 1))
freq_unigrams = step3.reduceByKey(lambda x, y: x + y)

预期结果

[((u'showing', 'VBG'), 24), ((u'Ave', 'NNP'), 1), ((u'Scrilla364', 'NNP'), 1), ((u'internally', 'RB'), 4), ...]

但是返回预期的字数需要很长时间（6 分钟）。 它停留在 reduceByKey 步骤。 如何解决这个性能问题？

-- 参考--

硬件规格

型号名称：MacBook Air 型号 标识符：MacBookAir4,2 处理器名称：英特尔酷睿 i7 处理器 速度：1.8 GHz 处理器数量：1 核心总数：2 L2 缓存（每核）：256 KB 三级缓存：4 MB 内存：4 GB

日志

15/10/02 16:05:12 INFO HadoopRDD: Input split: file:/usr/local/share/data/enronsent01:0+873602
15/10/02 16:05:12 INFO HadoopRDD: Input split: file:/usr/local/share/data/enronsent01:873602+873602
15/10/02 16:09:11 INFO BlockManagerInfo: Removed broadcast_2_piece0 on localhost:53478 in memory (size: 4.1 KB, free: 530.0 MB)
15/10/02 16:09:11 INFO BlockManagerInfo: Removed broadcast_3_piece0 on localhost:53478 in memory (size: 4.6 KB, free: 530.0 MB)
15/10/02 16:09:11 INFO ContextCleaner: Cleaned accumulator 4
15/10/02 16:09:11 INFO ContextCleaner: Cleaned accumulator 3
15/10/02 16:09:11 INFO BlockManagerInfo: Removed broadcast_1_piece0 on localhost:53478 in memory (size: 3.9 KB, free: 530.0 MB)
15/10/02 16:09:11 INFO ContextCleaner: Cleaned accumulator 2
15/10/02 16:10:05 INFO PythonRDD: Times: total = 292892, boot = 8, init = 275, finish = 292609
15/10/02 16:10:05 INFO Executor: Finished task 1.0 in stage 3.0 (TID 4). 2373 bytes result sent to driver
15/10/02 16:10:05 INFO TaskSetManager: Finished task 1.0 in stage 3.0 (TID 4) in 292956 ms on localhost (1/2)
15/10/02 16:10:35 INFO PythonRDD: Times: total = 322562, boot = 5, init = 276, finish = 322281
15/10/02 16:10:35 INFO Executor: Finished task 0.0 in stage 3.0 (TID 3). 2373 bytes result sent to driver
15/10/02 16:10:35 INFO TaskSetManager: Finished task 0.0 in stage 3.0 (TID 3) in 322591 ms on localhost (2/2)

Answer 1

代码看起来不错。

您可以尝试几个选项来提高性能。

SparkConf().setMaster("local[*]").setAppName("test").set("spark.executor.memory", "2g")

local -> local[*] 如果任务中断 - 它可以占用机器上可用的核心数量
如果可能的话，增加程序可用的内存

附：为了欣赏 Spark - 您应该拥有大量数据，以便可以在集群上运行它