Permalink
48e00e3 Sep 23, 2016
@ivuk @stgraber @naisanza @cryptofuture
80 lines (50 sloc) 4.5 KB

Introduction

So you've made it past trying out LXD live online, or on a server scavenged from random parts. You like what you see, and now you want to try doing some serious work with LXD.

With the vanilla installation of Ubuntu Server 16.04, some modifications to the server configuration will be needed, to avoid common pitfalls when using containers that require tens of thousands of file operations.

Common errors that may be encountered

Failed to allocate directory watch: Too many open files

<Error> <Error>: Too many open files

failed to open stream: Too many open files in...

Server Changes

/etc/security/limits.conf

Domain Type Item Value Default Description
* soft nofile 1048576 unset maximum number of open files
* hard nofile 1048576 unset maximum number of open files
root soft nofile 1048576 unset maximum number of open files
root hard nofile 1048576 unset maximum number of open files
* soft memlock unlimited unset maximum locked-in-memory address space (KB)
* hard memlock unlimited unset maximum locked-in-memory address space (KB)

/etc/sysctl.conf

Parameter Value Default Description
fs.inotify.max_queued_events 1048576 16384 This specifies an upper limit on the number of events that can be queued to the corresponding inotify instance. 1
fs.inotify.max_user_instances 1048576 128 This specifies an upper limit on the number of inotify instances that can be created per real user ID. 1
fs.inotify.max_user_watches 1048576 8192 This specifies an upper limit on the number of watches that can be created per real user ID. 1
vm.max_map_count 262144 65530 This file contains the maximum number of memory map areas a process may have. Memory map areas are used as a side-effect of calling malloc, directly by mmap and mprotect, and also when loading shared libraries.

Then, reboot the server.

Network Bandwidth Tweaking

If you have at least 1GbE NIC on your lxd host with a lot of local activity (container - container connections, or host - container connections), or you have 1GbE or better internet connection on your lxd host it worth play with txqueuelen. These settings work even better with 10GbE NIC.

Server Changes

txqueuelen

You need to change txqueuelen of your real NIC to 10000 (not sure about the best possible value for you), and change and change lxdbr0 interface txqueuelen to 10000.
In Debian-based distros you can change txqueuelen permanently in /etc/network/interfaces
You can add for ex.: up ip link set eth0 txqueuelen 10000 to your interface configuration to set txqueuelen value on boot.
You could set it txqueuelen temporary (for test purpose) with ifconfig interfacename# txqueuelen 10000

/etc/sysctl.conf

You also need to increase net.core.netdev_max_backlog value.
You can add net.core.netdev_max_backlog = 182757 to /etc/sysctl.conf to set it permanently (after reboot) You set netdev_max_backlog temporary (for test purpose) with echo 182757 > /proc/sys/net/core/netdev_max_backlog Note: You can find this value too high, most people prefer set netdev_max_backlog = net.ipv4.tcp_mem min. value. For example I use this values net.ipv4.tcp_mem = 182757 243679 365514

Containers changes

You also need to change txqueuelen value for all you ethernet interfaces in containers.
In Debian-based distros you can change txqueuelen permanently in /etc/network/interfaces
You can add for ex.: up ip link set eth0 txqueuelen 10000 to your interface configuration to set txqueuelen value on boot.

Notes regarding this change

10000 txqueuelen value commonly used with 10GbE NICs. Basically small txqueuelen values used with slow devices with a high latency, and higher with devices with low latency. I personally have like 3-5% improvement with these settings for local (host with container, container vs container) and internet connections. Good thing about txqueuelen value tweak, the more containers you use, the more you can be can benefit from this tweak. And you can always temporary set this values and check this tweak in your environment without lxd host reboot.