Etl之HiveSql调优(设置map reduce 的数量)
前言：
最近发现hivesql的执行速度特别慢，前面我们已经说明了left和union的优化，下面咱们分析一下增加或者减少reduce的数量来提升hsql的速度。
参考：http://www.cnblogs.com/liqiu/p/4873238.html
分析：
select s.id,o.order_id from sight s left join order_sight o on o.sight_id=s.id where s.id=9718 and o.create_time = '2015-10-10';
上一篇博文已经说明了，需要8个map，1个reduce，执行的速度：52秒。详细记录参考：http://www.cnblogs.com/liqiu/p/4873238.html
增加Reduce的数量：
首先说明一下reduce默认的个数：（每个reduce任务处理的数据量，默认为1000^3=1G，参数是hive.exec.reducers.bytes.per.reducer）；（每个任务最大的reduce数，默认为999，参数是hive.exec.reducers.max）
即，如果reduce的输入（map的输出）总大小不超过1G，那么只会有一个reduce任务；
如果数据表b2c_money_trace的大小是2.4G，那么reduce的数量是3个，例如：
hive> select count() from b2c_money_trace where operate_time = '2015-10-10' group by operate_time;
Total MapReduce jobs =
Launching Job out of
Number of reduce tasks not specified. Estimated from input data size:
In order to change the average load for a reducer (in bytes):
set hive.exec.reducers.bytes.per.reducer=<number>
In order to limit the maximum number of reducers:
set hive.exec.reducers.max=<number>
In order to set a constant number of reducers:
set mapred.reduce.tasks=<number>
Starting Job = job_1434099279301_3623421, Tracking URL = http://l-hdpm4.data.cn5.qunar.com:9981/proxy/application_1434099279301_3623421/
Kill Command = /home/q/hadoop/hadoop-2.2./bin/hadoop job -kill job_1434099279301_3623421
Hadoop job information for Stage-: number of mappers: ; number of reducers:
那么继续说最开始的例子，例如：
set mapred.reduce.tasks = 8;
执行的结果：
hive> set mapred.reduce.tasks = 8;
hive> select s.id,o.order_id from sight s left join order_sight o on o.sight_id=s.id where s.id=9718 and o.create_time = '2015-10-10';
Total MapReduce jobs = 1
Launching Job 1 out of 1
Number of reduce tasks not specified. Defaulting to jobconf value of: 8
In order to change the average load for a reducer (in bytes):
set hive.exec.reducers.bytes.per.reducer=<number>
In order to limit the maximum number of reducers:
set hive.exec.reducers.max=<number>
In order to set a constant number of reducers:
set mapred.reduce.tasks=<number>
Cannot run job locally: Input Size (= 380265495) is larger than hive.exec.mode.local.auto.inputbytes.max (= 50000000)
Starting Job = job_1434099279301_3618454, Tracking URL = http://l-hdpm4.data.cn5.qunar.com:9981/proxy/application_1434099279301_3618454/
Kill Command = /home/q/hadoop/hadoop-2.2.0/bin/hadoop job -kill job_1434099279301_3618454
Hadoop job information for Stage-1: number of mappers: 8; number of reducers: 8
2015-10-14 15:31:55,570 Stage-1 map = 0%, reduce = 0%
2015-10-14 15:32:01,734 Stage-1 map = 25%, reduce = 0%, Cumulative CPU 4.63 sec
2015-10-14 15:32:02,760 Stage-1 map = 50%, reduce = 0%, Cumulative CPU 10.93 sec
2015-10-14 15:32:03,786 Stage-1 map = 50%, reduce = 0%, Cumulative CPU 10.93 sec
2015-10-14 15:32:04,812 Stage-1 map = 75%, reduce = 0%, Cumulative CPU 21.94 sec
2015-10-14 15:32:05,837 Stage-1 map = 75%, reduce = 0%, Cumulative CPU 21.94 sec
2015-10-14 15:32:06,892 Stage-1 map = 75%, reduce = 0%, Cumulative CPU 21.94 sec
2015-10-14 15:32:07,947 Stage-1 map = 75%, reduce = 0%, Cumulative CPU 21.94 sec
2015-10-14 15:32:08,983 Stage-1 map = 75%, reduce = 0%, Cumulative CPU 21.94 sec
2015-10-14 15:32:10,039 Stage-1 map = 75%, reduce = 0%, Cumulative CPU 21.94 sec
2015-10-14 15:32:11,088 Stage-1 map = 75%, reduce = 0%, Cumulative CPU 21.94 sec
2015-10-14 15:32:12,114 Stage-1 map = 75%, reduce = 0%, Cumulative CPU 21.94 sec
2015-10-14 15:32:13,143 Stage-1 map = 75%, reduce = 19%, Cumulative CPU 24.28 sec
2015-10-14 15:32:14,170 Stage-1 map = 75%, reduce = 25%, Cumulative CPU 27.94 sec
2015-10-14 15:32:15,197 Stage-1 map = 75%, reduce = 25%, Cumulative CPU 27.94 sec
2015-10-14 15:32:16,224 Stage-1 map = 75%, reduce = 25%, Cumulative CPU 28.58 sec
2015-10-14 15:32:17,250 Stage-1 map = 75%, reduce = 25%, Cumulative CPU 28.95 sec
2015-10-14 15:32:18,277 Stage-1 map = 75%, reduce = 25%, Cumulative CPU 37.02 sec
2015-10-14 15:32:19,305 Stage-1 map = 75%, reduce = 25%, Cumulative CPU 48.93 sec
2015-10-14 15:32:20,332 Stage-1 map = 75%, reduce = 25%, Cumulative CPU 49.31 sec
2015-10-14 15:32:21,359 Stage-1 map = 100%, reduce = 25%, Cumulative CPU 57.99 sec
2015-10-14 15:32:22,385 Stage-1 map = 100%, reduce = 67%, Cumulative CPU 61.88 sec
2015-10-14 15:32:23,411 Stage-1 map = 100%, reduce = 100%, Cumulative CPU 71.56 sec
2015-10-14 15:32:24,435 Stage-1 map = 100%, reduce = 100%, Cumulative CPU 71.56 sec
MapReduce Total cumulative CPU time: 1 minutes 11 seconds 560 msec
Ended Job = job_1434099279301_3618454
MapReduce Jobs Launched:
Job 0: Map: 8 Reduce: 8 Cumulative CPU: 71.56 sec HDFS Read: 380267639 HDFS Write: 330 SUCCESS
Total MapReduce CPU Time Spent: 1 minutes 11 seconds 560 msec
OK
9718 210296076
9718 210299105
9718 210295344
9718 210295277
9718 210295586
9718 210295050
9718 210301363
9718 210297733
9718 210298066
9718 210295566
9718 210298219
9718 210296438
9718 210298328
9718 210298008
9718 210299712
9718 210295239
9718 210297567
9718 210295525
9718 210294949
9718 210296318
9718 210294421
9718 210295840
Time taken: 36.978 seconds, Fetched: 22 row(s)
可见8个reduce使得reduce的时间明显提升了。
增加Map的数量：
数据表大小：
map的数量就不能用上面的事例，那么看这个数据表：
hive> dfs -ls -h /user/ticketdev/hive/warehouse/business_mirror.db/b2c_money_trace;
Found 4 items
-rw-r--r-- 3 ticketdev ticketdev 600.0 M 2015-10-14 02:13 /user/ticketdev/hive/warehouse/business_mirror.db/b2c_money_trace/24f19a74-ca91-4fb2-9b79-1b1235f1c6f8
-rw-r--r-- 3 ticketdev ticketdev 597.2 M 2015-10-14 02:13 /user/ticketdev/hive/warehouse/business_mirror.db/b2c_money_trace/34ca13a3-de44-402e-9548-e6b9f92fde67
-rw-r--r-- 3 ticketdev ticketdev 590.6 M 2015-10-14 02:13 /user/ticketdev/hive/warehouse/business_mirror.db/b2c_money_trace/ac249f44-60eb-4bf7-9c1a-6f643873b823
-rw-r--r-- 3 ticketdev ticketdev 606.5 M 2015-10-14 02:13 /user/ticketdev/hive/warehouse/business_mirror.db/b2c_money_trace/f587fec9-60da-4f18-8b47-406999d95fd1
共2.4G
数据块大小：
hive> set dfs.block.size;
dfs.block.size=134217728
注意：134217728L是128M的意思！
map数量
文件大小是600M*4个，每个数据块是128M，即：取整(600/128)*4=20个Mapper
hive> select count(1) from b2c_money_trace;
Total MapReduce jobs = 1
Launching Job 1 out of 1
Number of reduce tasks determined at compile time: 1
In order to change the average load for a reducer (in bytes):
set hive.exec.reducers.bytes.per.reducer=<number>
In order to limit the maximum number of reducers:
set hive.exec.reducers.max=<number>
In order to set a constant number of reducers:
set mapred.reduce.tasks=<number>
Starting Job = job_1434099279301_3620170, Tracking URL = http://l-hdpm4.data.cn5.qunar.com:9981/proxy/application_1434099279301_3620170/
Kill Command = /home/q/hadoop/hadoop-2.2.0/bin/hadoop job -kill job_1434099279301_3620170
Hadoop job information for Stage-1: number of mappers: 20; number of reducers: 1
注意上面的红色部分，说明mappers的数量是20。
那么设置划分map的文件大小
set mapred.max.split.size=50000000;
set mapred.min.split.size.per.node=50000000;
set mapred.min.split.size.per.rack=50000000;
set hive.input.format=org.apache.hadoop.hive.ql.io.CombineHiveInputFormat;
大概解释一下：
50000000表示50M；
set hive.input.format=org.apache.hadoop.hive.ql.io.CombineHiveInputFormat;这个参数表示执行前进行小文件合并，当然这里没有使用到。
其他三个参数说明按照50M来划分数据块。
执行结果：
hive> select count(1) from b2c_money_trace;
Total MapReduce jobs = 1
Launching Job 1 out of 1
Number of reduce tasks determined at compile time: 1
In order to change the average load for a reducer (in bytes):
set hive.exec.reducers.bytes.per.reducer=<number>
In order to limit the maximum number of reducers:
set hive.exec.reducers.max=<number>
In order to set a constant number of reducers:
set mapred.reduce.tasks=<number>
Starting Job = job_1434099279301_3620223, Tracking URL = http://l-hdpm4.data.cn5.qunar.com:9981/proxy/application_1434099279301_3620223/
Kill Command = /home/q/hadoop/hadoop-2.2.0/bin/hadoop job -kill job_1434099279301_3620223
Hadoop job information for Stage-1: number of mappers: 36; number of reducers: 1
每个文件600M，正好12个Mapper，所以36个Mappers，注意上面的红色部分。
结论：
并非map和reduce数量越多越好，因为越多占用的资源越多，同时处理的时间未必一定增加，最好根据实际情况调整到一个合理的数量。
参考文章
http://lxw1234.com/archives/2015/04/15.htm
Etl之HiveSql调优(设置map reduce 的数量)的更多相关文章
- Etl之HiveSql调优(left join where的位置)
一.前言 公司实用Hadoop构建数据仓库,期间不可避免的实用HiveSql,在Etl过程中,速度成了避无可避的问题.本人有过几个数据表关联跑1个小时的经历,你可能觉得无所谓,可是多次Etl就要多个小 ...
- Etl之HiveSql调优(union all)
相信在Etl的过程中不可避免的实用union all来拼装数据,那么这就涉及到是否并行处理的问题了. 在hive中是否适用并行map,可以通过参数来设定: set hive.exec.parallel ...
- Tomcat6 一些调优设置内存和连接数
Tomcat6 一些调优设置内存和连接数 博客分类: java TomcatJVMLinux应用服务器网络应用 公司的一个服务器使用Tomcat6默认配置,在后台一阵全点击服务器就报废了,查了一下就 ...
- jmeter --JVM调优设置
JMeter用户可根据运行的计算机配置,来适当调整JMeter.bat中的JVM调优设置,如下所示: set HEAP=-Xms512m -Xmx512m set NEW=-XX:NewSize=12 ...
- HiveSql调优系列之Hive严格模式，如何合理使用Hive严格模式
目录 综述 1.严格模式 1.1 参数设置 1.2 查看参数 1.3 严格模式限制内容及对应参数设置 2.实际操作 2.1 分区表查询时必须指定分区 2.2 order by必须指定limit 2.3 ...
- [大牛翻译系列]Hadoop（10）MapReduce 性能调优：诊断reduce性能瓶颈
6.2.3 Reduce的性能问题 Reduce的性能问题有和map类似的方面,也有和map不同的方面.图6.13是reduce任务的具体的执行各阶段,标识了可能影响性能的区域. 这一章将介绍影响re ...
- Oracle 内存参数调优设置
Oracle 数据库系统中起到调节作用的参数叫初始化参数,数据库管理员根据实际情况需要适当调整这些 初始化参数以优化Oracle系统. 1 主要系统参数调优介绍 2 系统内存参数的分配 2.1 Ora ...
- Hadoop map reduce 任务数量优化
mapred.tasktracker.map.tasks.maximum 官方解释:The maximum number of map tasks that will be run simultan ...
- HiveSql调优经验
背景 在刚使用hive的过程中,碰到过很多问题,任务经常需要运行7,8个小时甚至更久,在此记录一下这个过程中,我的一些收获 join长尾 背景 SQL在Join执行阶段会将Join Key相同的数据分 ...
随机推荐
- CSS设计资料
CSS实现垂直居中的5种方法 网页阶级配色:http://tools.jb51.net/tools/peise.htm
- JUnit 测试
Junit 使用 1.忽略测试方法.在使用@Test的方法上使用@Ignore,将不会对此方法进行测试 2.测试套件 解决的问题: 1.对测试类进行统一测试,而不必在单独测试类上一个一个进行测试. 使 ...
- jenkins插件 查看job修改历史
文章来自:http://www.ciandcd.com文中的代码来自可以从github下载: https://github.com/ciandcd 插件jobConfigHistory(https:/ ...
- java5 ReadWriteLock用法--读写锁实现
读写锁:分为读锁和写锁,多个读锁不互斥,读锁与写锁互斥,这是由jvm自己控制的,你只要上好相应的锁即可.如果你的代码只读数据,可以很多人同时读,但不能同时写,那就上读锁:如果你的代码修改数据,只能有一 ...
- iOS开发——运行时OC篇&使用运行时获取系统的属性：使用自己的手势修改系统自带的手势
使用运行时获取系统的属性:使用自己的手势修改系统自带的手势 有的时候我需要实现一个功能,但是没有想到很好的方法或者想到了方法只是那个方法实现起来太麻烦,一或者确实为了装逼,我们就会想到iOS开发中最牛 ...
- cached过高导致内存溢出 java head space
最近公司线上遇到老是内存溢出检查后发现cached过高 命令:free -m 命令:sync //将缓存写入硬盘 cat /etc/redhat-release 这个是查看系统版本的命令c ...
- SQLite 批量insert - 如何加速SQLite的插入操作
本人翻译, 原文见: http://tech.vg.no/2011/04/04/speeding-up-sqlite-insert-operations/ 我正在开发一个Android程序, 它使用S ...
- UIScrollView offset in UINavigationController
转:UIScrollView offset in UINavigationController 通过设置viewCtronller的 self.automaticallyAdjustsScrollVi ...
- 用GO语言开发editplus编辑器插件(附源码)
我要开发的插件功能极为简单,就是对用户选中的内容进行base64编码或解密工作. 其中所涉及的技术部分主要是GO语言程序开发和editplus插件配置的部分,首先我们来看一下GO语言代码的写法,如下: ...
- 那些在学习iOS开发前就应该知道的事（part 2）
英文原文:Things I wish I had known before starting iOS development—Part 2 http://www.cocoachina.com/ios/ ...