当前位置:网站首页>[yarn] CDP cluster yarn configuration capacity scheduler batch allocation
[yarn] CDP cluster yarn configuration capacity scheduler batch allocation
2022-07-06 11:32:00 【kiraraLou】
One 、 Preface
It's going to be upgraded recently CDH Cluster into CDP colony ,CDH In the cluster Yarn By default, the service uses fair Scheduler ,CDP The cluster uses capacity Scheduler , We've been there before The scheduler is set unreasonably due to batch allocation , As a result, tasks are allocated to certain nodes , Make the cluster resource load extremely unbalanced .
To avoid the same problem in CDP On the cluster , We conduct research in advance . Look at using Will the scheduler also have centralized allocation . But in the process of research , There are some unexpected phenomena , Waiting for follow-up .
Two 、CDH Centralized distribution
As mentioned earlier ,CDH 5.8-Hadoop2.6.0 After version , in the light of Fair dispatch , There are several configurations as follows , For task types ( Small tasks ) Accelerate the optimization of allocation .
| Configuration name | explain |
|---|---|
yarn.scheduler.fair.max.assign | Maximum allocation : If assignmultiple by true And dynamic.max.assign by false, Then the maximum number of containers that can be allocated in a heartbeat . |
yarn.scheduler.fair.assignmultiple | Assign multiple : Whether multiple containers are allowed to be allocated in a heartbeat . |
yarn.scheduler.fair.dynamic.max.assign | If assignmultiple It's true , Whether to dynamically determine the amount of resources that a heartbeat can allocate . After opening , About half of the unallocated resources on the node will be allocated to the container in a heartbeat . Default to true . |
Through reasonable configuration , We can use centralized allocation , It will not expand the cluster load difference .
CDH How to configure a cluster is not covered here .
3、 ... and 、CDP Centralized distribution
CDP Already used in the cluster Capacity scheduling As the default scheduler , By consulting the official and Cloudera file , Find out Scheduling is also possible through heartbeat NodeManager Allocate multiple containers . The configuration is as follows :
| Configuration name | explain |
|---|---|
yarn.scheduler.capacity.per-node-heartbeat.multiple-assignments-enabled | Whether to allow in a NodeManager Allocate multiple containers in the heartbeat . Default to true . |
yarn.scheduler.capacity.per-node-heartbeat.maximum-container-assignments | If multiple-assignments-enabled by true, In a NodeManager The maximum number of containers that can be allocated in the heartbeat . The default is -1, No restrictions . |
yarn.scheduler.capacity.per-node-heartbeat.maximum-offswitch-assignments | If multiple-assignments-enabled by true, In a NodeManager The maximum that can be allocated in the heartbeat off-switch Number of containers . The default is 1, Indicates that only one off switch is allowed to be assigned in a heartbeat . |
How to configure
stay
Cloudera Managerin , Select cluster >YARN Queue manager UI service.
stay
YARNIn the queue manager window , Click the scheduler configuration tab .
stay “ Scheduler configuration ” Window

Choose
Enable Multiple Assignments Per HeartbeatCheck box to allow in aNodeManagerAllocate multiple containers in the heartbeatConfigure the following
NodeManagerHeartbeat properties :
Maximum Container Assignments Per Heartbeat: In aNodeManagerThe maximum number of containers that can be allocated in the heartbeat . Set this value to -1 This restriction will be disabled .Maximum Off-Switch Assignments Per Heartbeat: Can be in aNodeManagerThe maximum number of closed switch containers allocated in the heartbeat .
3、 ... and 、 summary
capacitySchedulers have similarfairThe heartbeat batch allocation configuration of the scheduler .- CDP colony
capacityThe scheduler enables batch allocation by default , And the assigned quantity is 100, This value needs to be reduced . - Now it's tested , It is found that the configuration does not seem to be effective , This needs to be followed up by experts .
Reference resources
https://docs.cloudera.com/cdp-private-cloud-base/7.1.7/yarn-allocate-resources/topics/yarn-set-user-limits.htmlhttps://docs.cloudera.com/cdp-private-cloud-base/7.1.7/yarn-allocate-resources/topics/yarn-configure-nm-heartbeat.htmlhttps://hadoop.apache.org/docs/r3.1.1/hadoop-yarn/hadoop-yarn-site/CapacityScheduler.html#Reviewing_the_configuration_of_the_CapacityScheduler
边栏推荐
- Error reporting solution - io UnsupportedOperation: can‘t do nonzero end-relative seeks
- Case analysis of data inconsistency caused by Pt OSC table change
- AcWing 179.阶乘分解 题解
- neo4j安装教程
- Pytorch基础
- Password free login of distributed nodes
- 保姆级出题教程
- How to build a new project for keil5mdk (with super detailed drawings)
- vs2019 第一个MFC应用程序
- 快来走进JVM吧
猜你喜欢

MySQL与c语言连接(vs2019版)

Nanny level problem setting tutorial

{一周总结}带你走进js知识的海洋

Why can't I use the @test annotation after introducing JUnit

How to build a new project for keil5mdk (with super detailed drawings)

软件测试与质量学习笔记3--白盒测试

Basic use of redis
![[download app for free]ineukernel OCR image data recognition and acquisition principle and product application](/img/1b/ed39a8b9181660809a081798eb8a24.jpg)
[download app for free]ineukernel OCR image data recognition and acquisition principle and product application

Install mongdb tutorial and redis tutorial under Windows

Introduction and use of automatic machine learning framework (flaml, H2O)
随机推荐
{一周总结}带你走进js知识的海洋
In the era of DFI dividends, can TGP become a new benchmark for future DFI?
What does usart1 mean
double转int精度丢失问题
Punctual atom stm32f103zet6 download serial port pin
Learn winpwn (3) -- sEH from scratch
QT creator specifies dependencies
L2-006 tree traversal (25 points)
Why can't STM32 download the program
Antlr4 uses keywords as identifiers
Record a problem of raspberry pie DNS resolution failure
第4阶段 Mysql数据库
AcWing 1298.曹冲养猪 题解
Pytorch基础
jS数组+数组方法重构
PyCharm中无法调用numpy,报错ModuleNotFoundError: No module named ‘numpy‘
Summary of numpy installation problems
Introduction and use of automatic machine learning framework (flaml, H2O)
搞笑漫画:程序员的逻辑
Test objects involved in safety test