1

I've had a test CentOS KVM host running for about a week now with 3 CentOS guests. There's 12GB physical RAM with about 7.5GB actually allocated to VMs. These VMs aren't even being used yet as the server's still in the testing stage but I've noticed that swap usage had been climbing over the last 24-48 hours. Now it looks like it's exhausted.

Here's the output of free

# free -m
             total       used       free     shared    buffers     cached
Mem:         11905      11749        155          0         81       4632
-/+ buffers/cache:       7035       4869
Swap:         2047       2047          0

So as you can see the physical memory is all used but it's used as cached memory which I believe is generally fine, as that will be released if an application needs it.

I ran the shell script found in this answer which listed 3 qemu-kvm processes.

The server was provisioned for me with only a 2GB logical volume allocated for swap and usually I like to match physical memory up to 8GB.
Is it worth expanding the swap logical volume or adding a separate swap volume?

Is this common with KVM? It's not something I've seen on other KVM hosts so is there a particular setting I need to adjust?

Any other comments/suggestions?

2

This is normal. Pages used by idle VMs will be swapped and the memory will be used for cache. You could set swappiness to zero which may prevent swap being used like this at the cost of performance (smaller cache).

Here is Red Hat's recommendation (with many YMMVs):

The swap space is calculated as follows:

Calculate the amount of memory needed for the sum of all the guests - In this example: (50 guests * 1GB of memory per guest) = 50GB

Add the guest memory amount to the amount needed for the host OS and for the host minimum swap space - In this example: 50GB guest memory + 4GB host OS + 4GB minimal swap = 58GB

Subtract this amount from the amount of physical RAM there is on the system - In this example 58GB - 32GB = 26GB

The answer is the amount of swap space that needs to be allocated. In this example 26GB

This example does not take into account kernel same-page merging (KSM) which will reduce the amount of memory used.

Here RH is saying 4G swap for host but here they recommend 12G * 0.5 = 6G swap for you.

If you are new to KVM I recommend reading IBM's Best Practices for KVM document.

  • I've upped the swap LV to 6GB and will see how it goes over the next week, I've got the disk space for it so it's no problem. Sounds like KVM's got it under control. – batfastad Nov 16 '13 at 14:27
  • And a further update to this now the machine is in production. Having upped the swap I have not noticed it get exhausted now the machine is in use. So I probably had swap set just slightly lower than was required. I have noticed that swap usage seems to grow and reduce so it looks like KVM is indeed actively managing it. – batfastad Dec 10 '13 at 11:01
  • I thought it was only about overcommitting ? Is SWAP useful on a hypervisor when you allocate less memory on the VMs ? – moutonjr Mar 15 '16 at 6:26
0

In my experience no, it's not worth assigning extra swap. You're going to be using swap (HDD) as RAM, and RAM as cache (shortcut for HDD).... sounds mostly counter-productive. Typically on a system with 8GB RAM I'll assign 1-2GB swap, usually the remainder of a raid partition scheme. Example:

3x64GB SSD (OS/software Raid0) 2x1TB SATA (Data Raid1)

I'll set my 1st SSD as 1GB on /boot, leaving 2x1GB wasted after I make the rest of the drives identical partitioning. Sounds like swap on Raid0, and usually it doesn't get accessed anyways.

As for KVM I'd look into bug reports as well as troubleshooting the individual guest machines. Could there be a memory leak in either?

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.