Tuesday, 9 September 2014

Squeezing more from your Linux VMs

There is a way to squeeze a little more from Linux VM's.

The theory

Linux has a number of ways of sharing it's storage IO among different processes, the norm seems to be to use the Completely Fair Queuing (CFQ) scheduler which helps to prevent a single process from using more than it's fair share of storage IO. This is usually helpful in an environment where this one Linux box is the only OS using a storage device but when it is running on an hypervisor the hypervisor is also busy trying to dish out fair access to the storage so we are effectively doubling up calculating fair access. There are other schedulers that we can choose from and for VM's it seems to make the most sense to use either the noop or the deadline schedulers. They are both more simple to calculate then CFQ.

The practice

If you are using the device sda for your drive use the following to check what scheduler you are using :
# cat /sys/block/sda/queue/scheduler
which will return something like
noop anticipatory deadline [cfq]
Which indicates that the scheduler for this device is CFQ.

To change it on the fly you just: 
echo noop > /sys/block/sda/queue/scheduler
If you want to make the change work across reboots you will need to add elevator=noop to your kernel boot parameters, on a Debian system you edit /etc/default/grub and add "elevator=noop" to the GRUB_CMDLINE_LINUX line then run:
# update-grub
To update the grub configuration.

Sources: kb.vmware.com/kb/2011861http://serverfault.com/questions/360718/kvm-low-io-performance