分布式资源调度——YARN框架

YARN产生背景

YARN是Hadoop2.x才有的,所以在介绍YARN之前,我们先看一下MapReduce1.x时所存在的问题:

  • 单点故障
  • 节点压力大
  • 不易扩展

MapReduce1.x时的架构如下:

可以看到,1.x时也是Master/Slave这种主从结构,在集群上的表现就是一个JobTracker带多个TaskTracker。

JobTracker:负责资源管理和作业调度
TaskTracker:定期向JobTracker汇报本节点的健康状况、资源使用情况以及作业执行情况。还可以接收来自JobTracker的命令,例如启动任务或结束任务等。

那么这种架构存在哪些问题呢:

  1. 整个集群中只有一个JobTracker,就代表着会存在单点故障的情况
  2. JobTracker节点的压力很大,不仅要接收来自客户端的请求,还要接收大量TaskTracker节点的请求
  3. 由于JobTracker是单节点,所以容易成为集群中的瓶颈,而且也不易域扩展
  4. JobTracker承载的职责过多,基本整个集群中的事情都是JobTracker来管理
  5. 1.x版本的整个集群只支持MapReduce作业,其他例如Spark的作业就不支持了

由于1.x版本不支持其他框架的作业,所以导致我们需要根据不同的框架去搭建多个集群。这样就会导致资源利用率比较低以及运维成本过高,因为多个集群会导致服务环境比较复杂。如下图:

在上图中我们可以看到,不同的框架我不仅需要搭建不同的集群。而且这些集群很多时候并不是总是在工作,如上图可以看到,Hadoop集群在忙的时候Spark就比较闲,Spark集群比较忙的时候Hadoop集群就比较闲,而MPI集群则是整体并不是很忙。这样就无法高效的利用资源,因为这些不同的集群无法互相使用资源。除此之外,我们还得运维这些个不同的集群,而且文件系统是无法共享的。如果当需要将Hadoop集群上的HDFS里存储的数据传输到Spark集群上进行计算时,还会耗费相当大的网络IO流量。

所以我们就想着要把这些集群都合并在一起,让这些不同的框架能够运行在同一个集群上,这样就能解决这各种各样的问题了。如下图:

正是因为在1.x中,有各种各样的问题,才使得YARN得以诞生,而YARN就可以令这些不同的框架运行在同一个集群上,并为它们调度资源。我们来看看Hadoop2.x的架构图:

在上图中,我们可以看到,集群最底层的是HDFS,在其之上的就是YARN层,而在YARN层上则是各种不同的计算框架。所以不同计算框架可以共享同一个HDFS集群上的数据,享受整体的资源调度,进而提高集群资源的利用率,这也就是所谓的 xxx on YARN。


YARN架构

YARN概述:

  • YARN是资源调度框架
  • 通用的资源管理系统
  • 为上层应用提供统一的资源管理和调度

YARN架构图,也是Master/Slave结构的:

从上图中,我们可以看到YARN主要由以下几个核心组件构成:

1. ResourceManager, 简称RM,整个集群同一时间提供服务的RM只有一个,它负责集群资源的统一管理和调度。以及还需要处理客户端的请求,例如:提交作业或结束作业等。并且监控集群中的NM,一旦某个NM挂了,那么就需要将该NM上运行的任务告诉AM来如何进行处理。

2. NodeManager, 简称NM,整个集群中会有多个NM,它主要负责自己本身节点的资源管理和使用,以及定时向RM汇报本节点的资源使用情况。接收并处理来自RM的各种命令,例如:启动Container。NM还需要处理来自AM的命令,例如:AM会告诉NM需要启动多少个Container来跑task。

3. ApplicationMaster, 简称AM,每个应用程序都对应着一个AM。例如:MapReduce会对应一个、Spark会对应一个。它主要负责应用程序的管理,为应用程序向RM申请资源(Core、Memory),将资源分配给内部的task。AM需要与NM通信,以此来启动或停止task。task是运行在Container里面的,所以AM也是运行在Container里面。

4. Container, 封装了CPU、Memory等资源的一个容器,相当于是一个任务运行环境的抽象。

5. Client, 客户端,它可以提交作业、查询作业的运行进度以及结束作业。

YARN官方文档地址如下:

https://hadoop.apache.org/docs/stable/hadoop-yarn/hadoop-yarn-site/YARN.html


YARN执行流程

假设客户端向ResourceManager提交一个作业,ResourceManager则会为这个作业分配一个Container。所以ResourceManager会与NodeManager进行通信,要求这个NodeManager启动一个Container。而这个Container是用来启动ApplicationMaster的,ApplicationMaster启动完之后会与ResourceManager进行一个注册。这时候客户端就可以通过ResourceManager查询作业的运行情况了。然后ApplicationMaster还会到ResourceManager上申请作业所需要的资源,申请到以后就会到对应的NodeManager之上运行客户端所提交的作业,然后NodeManager就会把task运行在启动的Container里。

如下图:

另外找到两篇关于YARN执行流程不错的文章:

  • 【图文】YARN 工作流程
  • Yarn应用程序运行流程剖析

YARN环境搭建

介绍完基本的理论部分之后,我们来搭建一个伪分布式的单节点YARN环境,使用的hadoop版本如下:

  • hadoop-2.6.0-cdh5.7.0

官方的安装文档地址如下:

https://hadoop.apache.org/docs/stable/hadoop-project-dist/hadoop-common/SingleCluster.html

1.下载并解压好hadoop-2.6.0-cdh5.7.0,这一步可以参考我之前写的一篇关于HDFS伪分布式环境搭建的文章,我这里就不再赘述了。

确保HDFS是正常启动状态:

[root@localhost ~]# jps
3827 Jps
3383 NameNode
3500 DataNode
3709 SecondaryNameNode
[root@localhost ~]# 

2.编辑mapred-site.xml配置文件,在文件中增加如下内容:

[root@localhost ~]# cd /usr/local/hadoop-2.6.0-cdh5.7.0/etc/hadoop
[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/etc/hadoop]# cp mapred-site.xml.template mapred-site.xml  # 拷贝模板文件
[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/etc/hadoop]# vim mapred-site.xml  # 增加如下内容

    mapreduce.framework.name
    yarn

3.编辑yarn-site.xml配置文件,在文件中增加如下内容:

[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/etc/hadoop]# vim yarn-site.xml  # 增加如下内容

    yarn.nodemanager.aux-services
    mapreduce_shuffle

4.启动ResourceManager进程以及NodeManager进程:

[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/etc/hadoop]# cd ../../sbin/
[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/sbin]# ./start-yarn.sh
starting yarn daemons
starting resourcemanager, logging to /usr/local/hadoop-2.6.0-cdh5.7.0/logs/yarn-root-resourcemanager-localhost.out
localhost: starting nodemanager, logging to /usr/local/hadoop-2.6.0-cdh5.7.0/logs/yarn-root-nodemanager-localhost.out
[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/sbin]# jps
3984 NodeManager  # 启动成功后可以看到多出了NodeManager
4947 DataNode
5252 Jps
5126 SecondaryNameNode
3884 ResourceManager  # 和ResourceManager进程,这样才是正常的。
4813 NameNode
[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/sbin]# netstat -lntp |grep java
tcp        0      0 0.0.0.0:50090           0.0.0.0:*               LISTEN      5126/java
tcp        0      0 127.0.0.1:42602         0.0.0.0:*               LISTEN      4947/java
tcp        0      0 192.168.77.130:8020     0.0.0.0:*               LISTEN      4813/java
tcp        0      0 0.0.0.0:50070           0.0.0.0:*               LISTEN      4813/java
tcp        0      0 0.0.0.0:50010           0.0.0.0:*               LISTEN      4947/java 
tcp        0      0 0.0.0.0:50075           0.0.0.0:*               LISTEN      4947/java
tcp        0      0 0.0.0.0:50020           0.0.0.0:*               LISTEN      4947/java 
tcp6       0      0 :::8040                 :::*                    LISTEN      5566/java 
tcp6       0      0 :::8042                 :::*                    LISTEN      5566/java
tcp6       0      0 :::8088                 :::*                    LISTEN      5457/java 
tcp6       0      0 :::13562                :::*                    LISTEN      5566/java
tcp6       0      0 :::8030                 :::*                    LISTEN      5457/java 
tcp6       0      0 :::8031                 :::*                    LISTEN      5457/java
tcp6       0      0 :::8032                 :::*                    LISTEN      5457/java 
tcp6       0      0 :::48929                :::*                    LISTEN      5566/java
tcp6       0      0 :::8033                 :::*                    LISTEN      5457/java
[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/sbin]# 

5.通过浏览器来访问ResourceManager,默认端口是8088,例如192.168.77.130:8088,就会访问到这样的一个页面上:

错误解决:

从上图中,可以看到有一个不健康的节点,也就是说我们的单节点环境有问题,点击红色框框中标记的数字可以进入到详细的信息页面,在该页面中看到了如下信息:

于是查看yarn的日志文件:yarn-root-nodemanager-localhost.log,发现如下警告与异常:

很明显是因为磁盘的使用空间达到了90%,所以我们需要删除一些没有的数据,或者扩容磁盘空间才行。于是删除了一堆安装包,让磁盘空间降低到90%以下了:

[root@localhost /usr/local]# df -h
Filesystem               Size  Used Avail Use% Mounted on
/dev/mapper/centos-root   19G   14G  4.5G  76% /
devtmpfs                 3.9G     0  3.9G   0% /dev
tmpfs                    3.9G     0  3.9G   0% /dev/shm
tmpfs                    3.9G  8.7M  3.9G   1% /run
tmpfs                    3.9G     0  3.9G   0% /sys/fs/cgroup
/dev/sdb                  50G   14G   34G  29% /kvm_data
/dev/sda1                497M  127M  371M  26% /boot
tmpfs                    781M     0  781M   0% /run/user/0
[root@localhost /usr/local]# 

这时再次刷新页面,可以发现这个节点就正常了:

到此为止,我们的yarn环境就搭建完成了。

如果需要关闭进程则使用以下命令:

[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/sbin]# stop-yarn.sh

初识提交PI的MapReduce作业到YARN上执行

虽然我们没有搭建MapReduce的环境,但是我们可以使用Hadoop自带的一些测试例子来演示一下如何提交作业到YARN上执行。Hadoop把example的包放在了如下路径,可以看到有好几个jar包:

[root@localhost ~]# cd /usr/local/hadoop-2.6.0-cdh5.7.0/share/hadoop/mapreduce/
[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/share/hadoop/mapreduce]# ls
hadoop-mapreduce-client-app-2.6.0-cdh5.7.0.jar
hadoop-mapreduce-client-common-2.6.0-cdh5.7.0.jar
hadoop-mapreduce-client-core-2.6.0-cdh5.7.0.jar
hadoop-mapreduce-client-hs-2.6.0-cdh5.7.0.jar
hadoop-mapreduce-client-hs-plugins-2.6.0-cdh5.7.0.jar
hadoop-mapreduce-client-jobclient-2.6.0-cdh5.7.0.jar
hadoop-mapreduce-client-jobclient-2.6.0-cdh5.7.0-tests.jar
hadoop-mapreduce-client-nativetask-2.6.0-cdh5.7.0.jar
hadoop-mapreduce-client-shuffle-2.6.0-cdh5.7.0.jar
hadoop-mapreduce-examples-2.6.0-cdh5.7.0.jar
lib
lib-examples
sources
[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/share/hadoop/mapreduce]# 

在这里我们使用hadoop-mapreduce-examples-2.6.0-cdh5.7.0.jar这个jar包来进行演示:

[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/share/hadoop/mapreduce]# hadoop jar hadoop-mapreduce-examples-2.6.0-cdh5.7.0.jar pi 2 3

命令说明:

  • hadoop jar 执行一个jar包作业的命令
  • hadoop-mapreduce-examples-2.6.0-cdh5.7.0.jar 需要被执行的jar包路径
  • pi 表示计算圆周率,可以写其他的
  • 末尾的两个数据分别表示指定运行2次map, 以及指定每个map任务取样3次,两数相乘即为总的取样数。

运行以上命令后,到浏览器页面上进行查看,会有以下三个阶段:

1.接收资源,这个阶段就是ApplicationMaster到ResourceManager上申请作业所需要的资源:

2.运行作业,这时候NodeManager就会把task运行在启动的Container里:

3.作业完成:

终端输出信息如下:

[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/share/hadoop/mapreduce]# hadoop jar hadoop-mapreduce-examples-2.6.0-cdh5.7.0.jar pi 2 3
Number of Maps  = 2
Samples per Map = 3
Wrote input for Map #0
Wrote input for Map #1
Starting Job
18/03/27 23:00:01 INFO client.RMProxy: Connecting to ResourceManager at /0.0.0.0:8032
18/03/27 23:00:01 INFO input.FileInputFormat: Total input paths to process : 2
18/03/27 23:00:01 INFO mapreduce.JobSubmitter: number of splits:2
18/03/27 23:00:02 INFO mapreduce.JobSubmitter: Submitting tokens for job: job_1522162696272_0001
18/03/27 23:00:02 INFO impl.YarnClientImpl: Submitted application application_1522162696272_0001
18/03/27 23:00:02 INFO mapreduce.Job: The url to track the job: http://localhost:8088/proxy/application_1522162696272_0001/
18/03/27 23:00:02 INFO mapreduce.Job: Running job: job_1522162696272_0001
18/03/27 23:00:10 INFO mapreduce.Job: Job job_1522162696272_0001 running in uber mode : false
18/03/27 23:00:10 INFO mapreduce.Job:  map 0% reduce 0%
18/03/27 23:00:15 INFO mapreduce.Job:  map 50% reduce 0%
18/03/27 23:00:16 INFO mapreduce.Job:  map 100% reduce 0%
18/03/27 23:00:19 INFO mapreduce.Job:  map 100% reduce 100%
18/03/27 23:00:20 INFO mapreduce.Job: Job job_1522162696272_0001 completed successfully
18/03/27 23:00:20 INFO mapreduce.Job: Counters: 49
    File System Counters
        FILE: Number of bytes read=50
        FILE: Number of bytes written=335298
        FILE: Number of read operations=0
        FILE: Number of large read operations=0
        FILE: Number of write operations=0
        HDFS: Number of bytes read=536
        HDFS: Number of bytes written=215
        HDFS: Number of read operations=11
        HDFS: Number of large read operations=0
        HDFS: Number of write operations=3
    Job Counters 
        Launched map tasks=2
        Launched reduce tasks=1
        Data-local map tasks=2
        Total time spent by all maps in occupied slots (ms)=7108
        Total time spent by all reduces in occupied slots (ms)=2066
        Total time spent by all map tasks (ms)=7108
        Total time spent by all reduce tasks (ms)=2066
        Total vcore-seconds taken by all map tasks=7108
        Total vcore-seconds taken by all reduce tasks=2066
        Total megabyte-seconds taken by all map tasks=7278592
        Total megabyte-seconds taken by all reduce tasks=2115584
    Map-Reduce Framework
        Map input records=2
        Map output records=4
        Map output bytes=36
        Map output materialized bytes=56
        Input split bytes=300
        Combine input records=0
        Combine output records=0
        Reduce input groups=2
        Reduce shuffle bytes=56
        Reduce input records=4
        Reduce output records=0
        Spilled Records=8
        Shuffled Maps =2
        Failed Shuffles=0
        Merged Map outputs=2
        GC time elapsed (ms)=172
        CPU time spent (ms)=2990
        Physical memory (bytes) snapshot=803618816
        Virtual memory (bytes) snapshot=8354324480
        Total committed heap usage (bytes)=760217600
    Shuffle Errors
        BAD_ID=0
        CONNECTION=0
        IO_ERROR=0
        WRONG_LENGTH=0
        WRONG_MAP=0
        WRONG_REDUCE=0
    File Input Format Counters 
        Bytes Read=236
    File Output Format Counters 
        Bytes Written=97
Job Finished in 19.96 seconds
Estimated value of Pi is 4.00000000000000000000
[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/share/hadoop/mapreduce]#

以上这个例子计算了一个PI值,下面我们再来演示一个hadoop中比较经典的例子:wordcount ,这是一个经典的词频统计的例子。首先创建好用于测试的文件:

[root@localhost ~]# mkdir /tmp/input
[root@localhost ~]# cd /tmp/input/
[root@localhost /tmp/input]# echo "hello word" > file1.txt
[root@localhost /tmp/input]# echo "hello hadoop" > file2.txt
[root@localhost /tmp/input]# echo "hello mapreduce" >> file2.txt
[root@localhost /tmp/input]# hdfs dfs -mkdir /wc_input
[root@localhost /tmp/input]# hdfs dfs -put ./file* /wc_input
[root@localhost /tmp/input]# hdfs dfs -ls /wc_input
Found 2 items
-rw-r--r--   1 root supergroup         11 2018-03-27 23:11 /wc_input/file1.txt
-rw-r--r--   1 root supergroup         29 2018-03-27 23:11 /wc_input/file2.txt
[root@localhost /tmp/input]# 

然后执行以下命令:

[root@localhost /tmp/input]# cd /usr/local/hadoop-2.6.0-cdh5.7.0/share/hadoop/mapreduce
[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/share/hadoop/mapreduce]# hadoop jar ./hadoop-mapreduce-examples-2.6.0-cdh5.7.0.jar wordcount /wc_input /wc_output

在yarn页面上显示的阶段信息:


终端输出信息如下:

[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/share/hadoop/mapreduce]# hadoop jar ./hadoop-mapreduce-examples-2.6.0-cdh5.7.0.jar wordcount /wc_input /wc_output
18/03/27 23:12:54 INFO client.RMProxy: Connecting to ResourceManager at /0.0.0.0:8032
18/03/27 23:12:55 INFO input.FileInputFormat: Total input paths to process : 2
18/03/27 23:12:55 INFO mapreduce.JobSubmitter: number of splits:2
18/03/27 23:12:55 INFO mapreduce.JobSubmitter: Submitting tokens for job: job_1522162696272_0002
18/03/27 23:12:56 INFO impl.YarnClientImpl: Submitted application application_1522162696272_0002
18/03/27 23:12:56 INFO mapreduce.Job: The url to track the job: http://localhost:8088/proxy/application_1522162696272_0002/
18/03/27 23:12:56 INFO mapreduce.Job: Running job: job_1522162696272_0002
18/03/27 23:13:02 INFO mapreduce.Job: Job job_1522162696272_0002 running in uber mode : false
18/03/27 23:13:02 INFO mapreduce.Job:  map 0% reduce 0%
18/03/27 23:13:06 INFO mapreduce.Job:  map 50% reduce 0%
18/03/27 23:13:07 INFO mapreduce.Job:  map 100% reduce 0%
18/03/27 23:13:11 INFO mapreduce.Job:  map 100% reduce 100%
18/03/27 23:13:12 INFO mapreduce.Job: Job job_1522162696272_0002 completed successfully
18/03/27 23:13:12 INFO mapreduce.Job: Counters: 49
    File System Counters
        FILE: Number of bytes read=70
        FILE: Number of bytes written=334375
        FILE: Number of read operations=0
        FILE: Number of large read operations=0
        FILE: Number of write operations=0
        HDFS: Number of bytes read=260
        HDFS: Number of bytes written=36
        HDFS: Number of read operations=9
        HDFS: Number of large read operations=0
        HDFS: Number of write operations=2
    Job Counters 
        Launched map tasks=2
        Launched reduce tasks=1
        Data-local map tasks=2
        Total time spent by all maps in occupied slots (ms)=5822
        Total time spent by all reduces in occupied slots (ms)=1992
        Total time spent by all map tasks (ms)=5822
        Total time spent by all reduce tasks (ms)=1992
        Total vcore-seconds taken by all map tasks=5822
        Total vcore-seconds taken by all reduce tasks=1992
        Total megabyte-seconds taken by all map tasks=5961728
        Total megabyte-seconds taken by all reduce tasks=2039808
    Map-Reduce Framework
        Map input records=3
        Map output records=6
        Map output bytes=64
        Map output materialized bytes=76
        Input split bytes=220
        Combine input records=6
        Combine output records=5
        Reduce input groups=4
        Reduce shuffle bytes=76
        Reduce input records=5
        Reduce output records=4
        Spilled Records=10
        Shuffled Maps =2
        Failed Shuffles=0
        Merged Map outputs=2
        GC time elapsed (ms)=157
        CPU time spent (ms)=2290
        Physical memory (bytes) snapshot=800239616
        Virtual memory (bytes) snapshot=8352272384
        Total committed heap usage (bytes)=762314752
    Shuffle Errors
        BAD_ID=0
        CONNECTION=0
        IO_ERROR=0
        WRONG_LENGTH=0
        WRONG_MAP=0
        WRONG_REDUCE=0
    File Input Format Counters 
        Bytes Read=40
    File Output Format Counters 
        Bytes Written=36
[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/share/hadoop/mapreduce]# 

查看输出的结果文件:

[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/share/hadoop/mapreduce]# hdfs dfs -ls /wc_output
Found 2 items
-rw-r--r--   1 root supergroup          0 2018-03-27 23:13 /wc_output/_SUCCESS
-rw-r--r--   1 root supergroup         36 2018-03-27 23:13 /wc_output/part-r-00000
[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/share/hadoop/mapreduce]# hdfs dfs -cat /wc_output/part-r-00000   # 实际输出结果在part-r-00000中
hadoop  1
hello   3
mapreduce   1
word    1
[root@localhost /usr/local/hadoop-2.6.0-cdh5.7.0/share/hadoop/mapreduce]#

转载于:https://blog.51cto.com/zero01/2091635

你可能感兴趣的:(分布式资源调度——YARN框架)