我正在尝试设置Slurm-我只有一个登录节点(称为ctm-login-01(和一个计算节点(称称为
ctm-deep-01。我的计算节点有几个CPU和3个GPU。
我的计算节点一直处于drain
状态,我一辈子都不知道从哪里开始。。。
登录节点
sinfo
ctm-login-01:~$ sinfo
PARTITION AVAIL TIMELIMIT NODES STATE NODELIST
debug* up infinite 1 drain ctm-deep-01
原因是什么?
sinfo-R
ctm-login-01:~$ sinfo -R
REASON USER TIMESTAMP NODELIST
gres/gpu count repor slurm 2020-12-11T15:56:55 ctm-deep-01
事实上,我一直在/var/log/slurm-llnl/slurmctld.log
:中收到这些错误消息
/var/log/slulm llnl/slurmctld.log
[2020-12-11T16:17:39.857] gres/gpu: state for ctm-deep-01
[2020-12-11T16:17:39.857] gres_cnt found:0 configured:3 avail:3 alloc:0
[2020-12-11T16:17:39.857] gres_bit_alloc:NULL
[2020-12-11T16:17:39.857] gres_used:(null)
[2020-12-11T16:17:39.857] error: _slurm_rpc_node_registration node=ctm-deep-01: Invalid argument
(请注意,我已将slurm.conf
Debug设置为verbose
,还将DebugFlags=Gres
设置为了解有关GPU的更多详细信息。(
这些是我在所有节点中的配置文件及其一些内容。。。
/etc/slurm-llnl/*文件
ctm-login-01:/etc/slurm-llnl$ ls
cgroup.conf cgroup_allowed_devices_file.conf gres.conf plugstack.conf plugstack.conf.d slurm.conf
ctm-login-01:/etc/slurm-llnl$ tail slurm.conf
#SuspendTime=
#
#
# COMPUTE NODES
GresTypes=gpu
NodeName=ctm-deep-01 Gres=gpu:3 CPUs=24 Sockets=1 CoresPerSocket=12 ThreadsPerCore=2 State=UNKNOWN
PartitionName=debug Nodes=ctm-deep-01 Default=YES MaxTime=INFINITE State=UP
# default
SallocDefaultCommand="srun --gres=gpu:1 $SHELL"
ctm-deep-01:/etc/slurm-llnl$ cat gres.conf
NodeName=ctm-login-01 Name=gpu File=/dev/nvidia0 CPUs=0-23
NodeName=ctm-login-01 Name=gpu File=/dev/nvidia1 CPUs=0-23
NodeName=ctm-login-01 Name=gpu File=/dev/nvidia2 CPUs=0-23
ctm-login-01:/etc/slurm-llnl$ cat cgroup.conf
CgroupAutomount=yes
CgroupReleaseAgentDir="/etc/slurm-llnl/cgroup"
ConstrainCores=yes
ConstrainDevices=yes
ConstrainRAMSpace=yes
#TaskAffinity=yes
ctm-login-01:/etc/slurm-llnl$ cat cgroup_allowed_devices_file.conf
/dev/null
/dev/urandom
/dev/zero
/dev/sda*
/dev/cpu/*/*
/dev/pts/*
/dev/nvidia*
计算节点
我的计算节点中的日志如下。
/var/log/slurm llnl/slurmd.log
ctm-deep-01:~$ sudo tail /var/log/slurm-llnl/slurmd.log
[2020-12-11T15:54:35.787] Munge credential signature plugin unloaded
[2020-12-11T15:54:35.788] Slurmd shutdown completing
[2020-12-11T15:55:53.433] Message aggregation disabled
[2020-12-11T15:55:53.436] topology NONE plugin loaded
[2020-12-11T15:55:53.436] route default plugin loaded
[2020-12-11T15:55:53.440] task affinity plugin loaded with CPU mask 0000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000ffffff
[2020-12-11T15:55:53.440] Munge credential signature plugin loaded
[2020-12-11T15:55:53.441] slurmd version 19.05.5 started
[2020-12-11T15:55:53.442] slurmd started on Fri, 11 Dec 2020 15:55:53 +0000
[2020-12-11T15:55:53.443] CPUs=24 Boards=1 Sockets=1 Cores=12 Threads=2 Memory=128754 TmpDisk=936355 Uptime=26 CPUSpecList=(null) FeaturesAvail=(null) FeaturesActive=(null)
CPU掩码的亲和力看起来很奇怪。。。
请注意,我已经呼叫了sudo nvidia-smi --persistence-mode=1
。还要注意,前面提到的gres.conf
文件似乎是正确的:
nvidia smi topo-m
ctm-deep-01:/etc/slurm-llnl$ sudo nvidia-smi topo -m
GPU0 GPU1 GPU2 CPU Affinity NUMA Affinity
GPU0 X SYS SYS 0-23 N/A
GPU1 SYS X PHB 0-23 N/A
GPU2 SYS PHB X 0-23 N/A
我应该从其他日志或配置中获得线索吗?谢谢
这都是因为打字错误!
ctm-deep-01:/etc/slurm-llnl$ cat gres.conf
NodeName=ctm-login-01 Name=gpu File=/dev/nvidia0 CPUs=0-23
NodeName=ctm-login-01 Name=gpu File=/dev/nvidia1 CPUs=0-23
NodeName=ctm-login-01 Name=gpu File=/dev/nvidia2 CPUs=0-23
显然,应该是NodeName=ctm-deep-01
,它是我的计算节点!天啊。。。
在添加新的GPU计算节点后,我在/var/log/slurmctld
中看到了同样的情况。
[2023-08-10T09:50:19.072] error: _slurm_rpc_node_registration node=c0236: Invalid argument
事实证明,在新节点完全启动后,我将GPU添加到了/etc/slurm/gres.conf
中:
NodeName=c[0236-0237] Name=gpu Type=ampere File=/dev/nvidia0 CPUs=[0-63]
NodeName=c[0236-0237] Name=gpu Type=ampere File=/dev/nvidia1 CPUs=[64-127]
一个sudo systemctl restart slurmctld && sudo scontrol reconfigure
是不够的。
我在节点上重新启动了slurmd
,slurmctld
日志指示如下,之后invalid argument
停止记录
[2023-08-10T09:51:58.765] gres/gpu: count changed for node c0236 from 0 to 2
根据[https://slurm.schedmd.com/gres.conf.html][SchedMD的gres.conf]文档中,gres.conf
slumd应该在使用reconfigure
时重读该文件。也许是因为我们还在18.08,直到后来才加上。
至少,对于我们的新节点,我们需要在添加Gres条目后重新启动slurmd。