On 02/18/2014 09:06 PM, Wangyufei (James) wrote:

Hello, When I call virDomainManagedSave, libvirtd will fork a child process libvirt_iohelper to write save file, and the system cache will increase soon just like this:

Swap: 0M total, 0M used, 0M free, 1668M cached Swap: 0M total, 0M used, 0M free, 1715M cached

Yep, that's what the kernel does if you don't use 'virsh managedsave --bypass-cache'.

But I have cgroup to control the memory of libvirtd below 50 M, so in this case, libvirt_iohelper will be killed, and virDomainManagedSave failed.

So my question is: 1. Is it ok for libvirtd to occupy large amount of cache when save? Can we flush cache in time? 2. In my situation, is it any good idea to help me out? Increase the memory limit value? Pass through the cache? Flush cache? Or any one better? Which one do you preffer?

Try using the --bypass-cache flag (VIR_DOMAIN_SAVE_BYPASS_CACHE), which tells libvirt to use O_DIRECT which in turn avoids filling the kernel file system cache (won't work on tmpfs, which is why it is not default). -- Eric Blake eblake redhat com +1-919-301-3266 Libvirt virtualization library http://libvirt.org