[systemd] Add value to LimitNOFILE due to performance problems

When k3s is installed on an OS with default high ulimits, performance
issues can be observed. This was discovered on CoreOS where the default
value is 1073741816. Symptoms include very slow file operations such
as installing a Rook/Ceph cluster will take ~6 hours instead of ~10 minutes.

A google search for 'container LimitNOFILE' will show that most major
projects set this already, including the (unused) containerd systemd unit
found in this repository at /vendor/github.com/containerd/containerd/containerd.service

k3OS is not affected becuasse the default there is already 1048576.

See description in coreos/fedora-coreos-tracker#329
This commit is contained in:
Adam Farden 2020-05-03 09:27:54 +02:00
parent 62a50429ca
commit df10b51c91
2 changed files with 6 additions and 2 deletions

View File

@ -10,7 +10,9 @@ ExecStartPre=-/sbin/modprobe overlay
ExecStart=/usr/local/bin/k3s server {{ extra_server_args | default("") }}
KillMode=process
Delegate=yes
LimitNOFILE=infinity
# Having non-zero Limit*s causes performance problems due to accounting overhead
# in the kernel. We recommend using cgroups to do container-local accounting.
LimitNOFILE=1048576
LimitNPROC=infinity
LimitCORE=infinity
TasksMax=infinity

View File

@ -10,7 +10,9 @@ ExecStartPre=-/sbin/modprobe overlay
ExecStart=/usr/local/bin/k3s agent --server https://{{ master_ip }}:6443 --token {{ hostvars[groups['master'][0]]['token'] }}
KillMode=process
Delegate=yes
LimitNOFILE=infinity
# Having non-zero Limit*s causes performance problems due to accounting overhead
# in the kernel. We recommend using cgroups to do container-local accounting.
LimitNOFILE=1048576
LimitNPROC=infinity
LimitCORE=infinity
TasksMax=infinity