[lfs-dev] kvm performance

Bruce Dubbs bruce.dubbs at gmail.com
Mon Sep 23 09:10:13 PDT 2013


Stuart Stegall wrote:
> I use virtio with the cache set to none on KVM.  I get 57s on Host (16
> threads, 64GB RAM, single SSD storage, using -j4) and 69s in a VM (4
> threads, 8GB RAM, virtio is on the same SSD formated ext2, using -j4).  Not
> sure if these metrics are helpful for anyone or not, but there you go.

The results are useful, but for many of the messages here, we are 
comparing apples and oranges.  I always use -j1 for timing purposes.  To 
check though, I went to the host machine and ran two tests.  I am just 
timing the make portion of binutils and I delete and recreate the 
binutils directory each time:

Host -j1: 119s
      -j2:  63s

KVM (all -j1 -- only one processor in virtual system)

file system    cache               time (s)
virtio         default=writeback   225
virtio         unsafe              211
virtio         none                213
std            default=writeback   216
std            none                215

What I am seeing is that the file system and cache in kvm are not making 
a significant difference.

I did find these URLs:

http://dtrace.org/blogs/brendan/2013/01/11/virtualization-performance-zones-kvm-xen/

https://www.duo.uio.no/handle/10852/34900

On the pdf of the above, it measured KVM I/O at about 60% of the host. 
See pages 56-58.  That seems to be consistent with my results.

For LFS, the primary activity is building: compilation and linking. 
These operations are pretty demanding of a system for both disk I/O and 
CPU.  From what I can tell, vmware seems to do a lot better than kvm for 
this type of activity.  For many activities, e.g. rendering a web page, 
I suspect that kvm is OK.

   -- Bruce



More information about the lfs-dev mailing list