Using YARN with Cgroups testing in sparkml cluster

 

部署服务器:

sparkml 集群

########### sparkml ##########

sparkml-node1 # yarn resource manager
sparkml-node2 # nodemanager spark-2.0.0
sparkml-node3 # nodemanager spark-2.0.0
sparkml-node4 # nodemanager spark-2.0.0
sparkml-node5 # nodemanager spark-2.0.0

 

上线功能:

  1.  Cgroup 限制每个节点 yarn container 能占用的该节点 CPU 总量
  2.  每个 yarn container 能够按照被分配的 vcore 数目 share CPU 

 

测试方法:

功能一测试:

在不限制的情况下,我们跑一条 hive SQL 

test_hive_sql.sql

我们看看 container 分配情况:

 4 个 nodemanager 节点的 CPU 使用情况:

 

都接近 100 % 

我们现在尝试限制到 50%

设置 cpu.cfs_quota_us="1200000"; (计算方法:24 (逻辑CPU核心数)* 0.5(50% CPU 使用)* 100000(每个计算周期)  = 1200000)

重启 cgroup : /etc/init.d/cgconfig restart 

再跑一次同样的 SQL : 

基本同样的 container 分配

nodemanager 服务器上的 CPU 使用:

 

全部限制在 50% 以内

功能二,测试:

hive SQL 跑出来的 container 都只占用了 一个 vcore (mapred的特性?),因此我们用 spark 来进行测试:

我们跑这一段代码:

from __future__ import print_function

#
# Licensed to the Apache Software Foundation (ASF) under one or more
# contributor license agreements.  See the NOTICE file distributed with
# this work for additional information regarding copyright ownership.
# The ASF licenses this file to You under the Apache License, Version 2.0
# (the "License"); you may not use this file except in compliance with
# the License.  You may obtain a copy of the License at
#
#    http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#

import sys
from random import random
from operator import add

from pyspark import SparkContext

import time


if __name__ == "__main__":
    """
        Usage: pi [partitions]
    """
    sc = SparkContext(appName="PythonPi")
    partitions = int(sys.argv[1]) if len(sys.argv) > 1 else 2
    n = 100000 * partitions

    def f(_):
        for i in range(1,10000):
            x = random() * random() * random() - 1
            y = random() * random() * random() - 1
        #time.sleep(60)
        x = random() * random() * random() - 1
        y = random() * random() * random() - 1
        return 1 if x ** 2 + y ** 2 < 1 else 0

    count = sc.parallelize(range(1, n + 1), partitions).map(f).reduce(add)
    print("Pi is roughly %f" % (4.0 * count / n))

    sc.stop()

container 分配:

 

 跑了 1 个 container 4 个 vcore 的服务器上面:

跑测试的  hive SQL 

在 node4 这台服务器上:

spark_sc 的 CPU 占用只有 100,没有其他 vcore 为 1 的来自 hdfs 的 container 多

这是因为上述 python 代码没有并发,因此只能使用 一个 核

这台服务器上有 5 个 container :

 

只有 最后一个 container 的 cpu.shares 值是 4096 ,是别的 4 倍

 

 上述结果和我们观察到的 vcore 分配一致,在这里 python code 的 CPU 占用没有 hive SQL 生成的 container 多是因为 python 使用了 单进程,没有多核调度

 

测试结果:

对于功能一:生效

对于功能二:生效,通过控制 cpu.shares 来按照 vcore 分配 CPU ,缺乏直观的测试数据

 

配置参数:

yarn.nodemanager.container-executor.class : org.apache.hadoop.yarn.server.nodemanager.LinuxContainerExecutor

yarn.nodemanager.linux-container-executor.resources-handler.class : org.apache.hadoop.yarn.server.nodemanager.util.CgroupsLCEResourcesHandler

yarn.nodemanager.linux-container-executor.cgroups.hierarchy : /hadoop-yarn (对于 /cgroup/cpu/ 目录下的 cgroup hierarchy ,手动配置到 cgconfig.conf 文件里面)

yarn.nodemanager.linux-container-executor.cgroups.mount : true 

yarn.nodemanager.linux-container-executor.cgroups.mount-path : /cgroup (cgroup 文件系统根目录)

yarn.nodemanager.linux-container-executor.group : yarn 

yarn.nodemanager.linux-container-executor.nonsecure-mode.limit-users : false 

不生效参数:

yarn.nodemanager.resource.percentage-physical-cpu-limit : 100 (该参数控制 nodemanager 节点的总体CPU 使用,hadoop-2.5.0-cdh5.3.2 不支持,可以同在 在 cgconfig.conf 中配置 cpu.cfs_quota_us)

yarn.nodemanager.linux-container-executor.cgroups.strict-resource-usage : false (CPU use hard limit)

 

cgroup 配置:

 

#
#  Copyright IBM Corporation. 2007
#
#  Authors:    Balbir Singh <balbir@linux.vnet.ibm.com>
#  This program is free software; you can redistribute it and/or modify it
#  under the terms of version 2.1 of the GNU Lesser General Public License
#  as published by the Free Software Foundation.
#
#  This program is distributed in the hope that it would be useful, but
#  WITHOUT ANY WARRANTY; without even the implied warranty of
#  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
#
# See man cgconfig.conf for further details.
#
# By default, mount all controllers to /cgroup/<controller>

mount {
    cpuset    = /cgroup/cpuset;
    cpu    = /cgroup/cpu;
    cpuacct    = /cgroup/cpuacct;
    memory    = /cgroup/memory;
    devices    = /cgroup/devices;
    freezer    = /cgroup/freezer;
    net_cls    = /cgroup/net_cls;
    blkio    = /cgroup/blkio;
}

group hadoop-yarn {
     perm {
         task {
             uid = yarn;
             gid = hadoop;
         } admin {
             uid = yarn;
             gid = hadoop;
         }
     }
    cpu {
#             cpu.shares="1024";
#             cpu.cfs_period_us="100000";
#             cpu.cfs_quota_us="1200000";
    }
}

 

原理简述:

cgroup 通过 cgroup hierarchy  来将 subsystem 和 task 联系起来,每次 yarn 在启动 container 的时候都会将在指定的 hadoop-yarn cgroup hierarchy  下面新建属于每个 container 的 hierarchy   

开始跑 container 以后 

由于总体的 节点 CPU 限制在线上版本不支持(YarnConfiguration.java 里面没有读入 yarn.nodemanager.resource.percentage-physical-cpu-limit 参数,也没有在 CgroupsLCEResourcesHandler 有相关实现,具体实现参考 : YARN-2440

我们在 hadoop-yarn 里面配置 设置 cpu.cfs_quota_us ,在 hadoop-yarn  下属的所有 container cgroup hierarchy   都不能超过父 hierarchy   的限制

 

对于功能二:

通过 YARN-600 加入到  CgroupsLCEResourcesHandler 类 

if (isCpuWeightEnabled()) {
  createCgroup(CONTROLLER_CPU, containerName);
  int cpuShares = CPU_DEFAULT_WEIGHT * containerResource.getVirtualCores();

  // absolute minimum of 10 shares for zero CPU containers
 cpuShares = Math.max(cpuShares, 10);

  updateCgroup(CONTROLLER_CPU, containerName, "shares",
      String.valueOf(cpuShares));
}

 

cpuShares 最少值 为 10 ,按照 VirtualCores 给予每个 container 相应的 cpu.shares 值 

Linux cfs 调度器会根据 cpu.shares 值作用到 CPU 调度,具体参考:cpu.shares 作用原理

 

部署流程:

yarn-site.xml 

<property>
<name>yarn.nodemanager.container-executor.class</name>
<value>org.apache.hadoop.yarn.server.nodemanager.LinuxContainerExecutor</value>
</property>
<property>
<name>yarn.nodemanager.linux-container-executor.resources-handler.class</name>
<value>org.apache.hadoop.yarn.server.nodemanager.util.CgroupsLCEResourcesHandler</value>
</property>
<property>
<name>yarn.nodemanager.linux-container-executor.cgroups.hierarchy</name>
<value>/hadoop-yarn</value>
</property>
<property>
<name>yarn.nodemanager.linux-container-executor.cgroups.mount</name>
<value>true</value>
</property>
<property>
<name>yarn.nodemanager.linux-container-executor.cgroups.mount-path</name>
<value>/cgroup</value>
</property>
<property>
<name>yarn.nodemanager.linux-container-executor.group</name>
<value>yarn</value>
</property>
<property>
<name>yarn.nodemanager.resource.percentage-physical-cpu-limit</name>
<value>100</value>
</property>
<property>
<name>yarn.nodemanager.linux-container-executor.cgroups.strict-resource-usage</name>
<value>false</value>
</property>
<property>
<name>yarn.nodemanager.linux-container-executor.nonsecure-mode.limit-users</name>
<value>false</value>
</property>

 

部署 cgroup 

 

重新编译 container-executor : 

cd ${HADOOP_HOME}/hadoop-2.6.0-src/hadoop-yarn-project/hadoop-yarn/hadoop-yarn-server/hadoop-yarn-server-nodemanager/

cmake src -DHADOOP_CONF_DIR=/etc/hadoop
make

cd targe/usr/local/bin/即可获得需要的container-executor文件

配置 container-executor.cfg 

yarn.nodemanager.linux-container-executor.group=yarn
banned.users=bin
min.user.id=0
allowed.system.users=hdfs,yarn

启动 cgroup 

重启 yarn 

 

参考文献:

yarn 新特性 - cgroup

Using YARN with Cgroups

Using YARN with Cgroups 参数配置 Apache 官网

cgroup 使用文档

YARN配置Kerberos认证

container executor 简介

按照 vcore 计算 container CPU 使用

AMBARI-9376

美团概述:Cgroup 简介

Cgroup – 从CPU资源隔离说起

cgroup 部署测试指南

 

后续跟进:

调查 yarn 是否支持灰度上 cgroup 

我们使用在外围不停 cgclassify 来上 cgroup 

#!/bin/bash 

echo ""
echo ""

containerPid=` su - yarn -c ' jps | grep -v NodeManager | grep -v -i jps ' | awk '{print $1}' `

containerList=` su - yarn -c ' jps | grep -v NodeManager | grep -v -i jps ' ` 

echo " We will begin to move ${containerList} of yarn to cgroup "

for pid in ${containerPid}
do
  cgclassify -g cpu:hadoop-yarn $pid
done 

echo " Move to cgroup per minute done "

taskID=` cat /cgroup/cpu/hadoop-yarn/tasks `

echo " Content in hadoop-yarn hierarchy is : ${taskID} "

date

echo ""
echo ""

 

部署 crontab job 一分钟一次,看效果 

待续

 

posted @ 2017-01-13 12:08  David_Lin  阅读(483)  评论(0)    收藏  举报