Slurm集羣搭建

一、系統環境

a) Centos 7.8
b) 兩臺機器規劃IP 分別爲10.10.0.20 與10.10.0.21
c) 機器主機名分別爲node1、node2
d) 均關閉了防火牆以及selinux
e) 已經配置好SSH 免密、NIS、NFS

二、配置munge 及Slurm

a) 安裝依賴
	yum install -y epel-release
	yum install -y gtk2 gtk2-devel munge munge-devel python python3
b) 配置munge #兩臺機器均要配置
	chown slurm:slurm /etc/munge
	chown slurm:slurm /var/run/munge
	chown slurm:slurm /var/lib/munge
	chown slurm:slurm /var/log/munge
	create-munge-key #此步驟只在node1 節點做即可
	scp /etc/munge/munge.key node2:/etc/munge/
	chown slurm:slurm /etc/munge/munge.key
	su - slurm #所有節點使用slurm 用戶啓動munged
	munged
c) 配置Slurm #所有節點配置相同
	tar xvf slurm-20.11.5.tar.bz2
	cd slurm-20.11.5/
	./configure
	make -j3
	make install -j3
	cp etc/{slurmctld.service,slurmdbd.service,slurmd.service}
	/usr/lib/systemd/system
	vi /usr/local/etc/slurm.conf #配置文件附在PS 中
	scp /usr/local/etc/slurm.conf node2:/usr/local/etc/
	chown slurm:slurm /var/spool/
	systemctl start slurmctld #主節點啓動slurmctld、slurmd
	systemctl start slurmd #從節點只需啓動slurmd 即可

三、測試

a) 系統命令測試
	[root[@node1](https://my.oschina.net/u/4273199) ~]# sinfo
	PARTITION AVAIL TIMELIMIT NODES STATE NODELIST
	control up infinite 1 idle node1
	compute* up infinite 2 idle node[1-2]
	[root[@node1](https://my.oschina.net/u/4273199) ~]# srun -N 2 -l hostname # -N 2 執行的節點數
	0: node1
	1: node2
	至此Slurm 集羣搭建完成
b) mpi 測試
	vi test.c #測試程序在PS 中
	mpicc -o test test.c #編譯mpi 程序
	vi tj.sh #作業腳本在PS 中
	sbatch tj.sh #提交作業
	squeue #查看作業狀態
	[root[@node1](https://my.oschina.net/u/4273199) ~]# cat test.out #查看作業結果
	node2: Hello world from process 2
	number of processes: 4
	...node1: Hello world from process 0
	node2: Hello world from process 3
	node1: Hello world from process 1

四、PS

a) tj.sh 作業腳本
	#!/bin/sh
	#SBATCH -o /root/test.out #結果輸出到test.out
	#SBATCH --nodes=2 #節點數：2
	#SBATCH --ntasks-per-node=2
	mpirun /root/test
b) slurm.conf 配置文件
	SlurmctldHost=node1 #主節點
	MpiDefault=none
	ProctrackType=proctrack/pgid #通過網站配置注意修改
	ReturnToService=1
	SlurmctldPidFile=/var/run/slurmctld.pid
	SlurmctldPort=6817
	SlurmdPidFile=/var/run/slurmd.pid
	SlurmdPort=6818
	SlurmdSpoolDir=/var/spool/slurmd
	SlurmUser=slurm #slurm 管理用戶
	StateSaveLocation=/var/spool
	SwitchType=switch/none
	TaskPlugin=task/affinity
	InactiveLimit=0
	KillWait=30
	MinJobAge=300
	SlurmctldTimeout=120
	SlurmdTimeout=300
	Waittime=0
	SchedulerType=sched/backfill
	SelectType=select/cons_tres
	SelectTypeParameters=CR_Core
	AccountingStorageType=accounting_storage/none
	AccountingStoreJobComment=YES
	ClusterName=siton #集羣名
	JobCompType=jobcomp/none
	JobAcctGatherFrequency=30
	JobAcctGatherType=jobacct_gather/none
	SlurmctldDebug=info
	SlurmdDebug=info
	NodeName=node1,node2 CPUs=4 RealMemory=2 Sockets=4 CoresPerSocket=1 ThreadsPerCore=1 State=UNKNOWN
		/*節點名稱，CPUs 核數，corepersocket,threadspersocket,使用lscpu 查看，realmemory 實際分配給slurm 內存，procs 是實際CPU 個數，/proc/cpuinfo 裏查看state=unknown是啓動集羣的時候爲unknown，之後會變成idle*/
	PartitionName=control Nodes=node1 Default=YES MaxTime=INFINITE State=UP	
	PartitionName=compute Nodes=node1,node2 Default=YES MaxTime=INFINITE State=UP
		/*partitionname 是分成control 和compute，default=yes 是說這個用來計算*/
		可以通過https://slurm.schedmd.com/configurator.html 生成配置文件
c) mpi 測試程序
	#include <stdio.h>
	#include <mpi.h>
	#include <stdlib.h>
	#include <time.h>
	int main(int argc, char* argv[])
	{
	int myid, numprocs, namelen;
	char processor_name[MPI_MAX_PROCESSOR_NAME];
	MPI_Init(&argc, &argv);
	MPI_Comm_rank(MPI_COMM_WORLD, &myid);
	MPI_Comm_size(MPI_COMM_WORLD, &numprocs);
	MPI_Get_processor_name(processor_name, &namelen);
	if (myid == 0) printf("number of
	processes: %d\n...", numprocs);
	printf("%s: Hello world from process %d \n",
	processor_name, myid);
	MPI_Finalize();
	return 0;
	}

五、進階（GPU）

修改slurm.conf 文件
GresTypes=gpu
NodeName=slave3 Sockets=2 Procs=32 CoresPerSocket=8 ThreadsPerCore=2
RealMemory=3000 Gres=gpu:tesla:2 State=UNKNOWN NodeAddr=10.135.12.29
此外，slave3 這個機器需配置GPU 信息，編輯/usr/local/etc/gres.conf 文件
Name=gpu Type=tesla File=/dev/nvidia0
Name=gpu Type=tesla File=/dev/nvidia1
slurm 腳本中添加gres 指定gpu 資源
#SBATCH --gres=gpu:tesla:2