Re: Guest performance is reduced after live migration

[Marcelo Tosatti <mtosatti@redhat.com>]

On Wed, Jan 02, 2013 at 11:56:11PM +0000, Mark Petersen wrote:
> I don't think it's related to huge pages...
> 
> I was using phoronix-test-suite to run benchmarks.  The 'batch/compilation' group shows the slowdown for all tests, the 'batch/computation' show some performance degradation, but not nearly as significant.

Huge pages in the host, for the qemu-kvm process, i mean.
Without huge pages backing guest memory in the host, 4k EPT TLB entries
will be used instead of 2MB EPT TLB entries.

> You could probably easily test this way without phoronix -  Start a VM with almost nothing running.  Download mainline Linux kernel, compile.  This takes about 45 seconds in my case (72GB memory, 16 virtual CPUs, idle physical host running this VM.)  Run as many times as you want, still takes ~45 seconds.
> 
> Migrate to a new idle host, kernel compile now takes ~90 seconds, wait 3 hours (should give khugepaged a change to do its thing I imagine),

Please verify its the case (by checking how much memory is backed by
hugepages).

http://www.mjmwired.net/kernel/Documentation/vm/transhuge.txt
"Monitoring Usage".


> kernel compiles still take 90 seconds.  Reboot virtual machine (run 'shutdown -r now', reboot, whatever.)  First compile will take ~45 seconds after reboot.  You don't even need to reset/destroy/shutdown the VM, just a reboot in the guest fixes the issue.

What is the qemu command line?

> I'm going to test more with qemu-kvm 1.3 tomorrow as I have a new/dedicated lab setup and recently built the 1.3 code base.  I'd be happy to run any test that would help in diagnosing the real issue here, I'm just not sure how to best diagnose this issue.
> 
> Thanks,
> Mark
>  
> -----Original Message-----
> 
> Can you describe more details of the test you are performing? 
> 
> If transparent hugepages are being used then there is the possibility that there has been no time for khugepaged to back guest memory with huge pages, in the destination (don't recall the interface for retrieving number of hugepages for a given process, probably somewhere in /proc/pid/).
> 
> On Wed, Dec 19, 2012 at 12:43:37AM +0000, Mark Petersen wrote:
> > Hello KVM,
> > 
> > I'm seeing something similar to this (http://thread.gmane.org/gmane.comp.emulators.kvm.devel/100592) as well when doing live migrations on Ubuntu 12.04 (Host and Guest) with a backported libvirt 1.0 and qemu-kvm 1.2 (improved performance for live migrations on guests with large memory guests is great!)  The default libvirt  0.9.8 and qemu-kvm 1.0 have the same issue.
> > 
> > Kernel is 3.2.0-34-generic and eglicb 2.15 on both host/guest.  I'm seeing similar issues with both virtio and ide bus.  Hugetblfs is not used, but transparent hugepages are.  Host machines are dual core Xeon E5-2660 processors.  I tried disabling EPT but that doesn't seem to make a difference so I don't think it's a requirement to reproduce.
> > 
> > If I use Ubuntu 10.04 guest with eglibc 2.11 and any of these kernels I don't seem to have the issue:
> > 
> > linux-image-2.6.32-32-server - 2.6.32-32.62 
> > linux-image-2.6.32-38-server - 2.6.32-38.83 
> > linux-image-2.6.32-43-server - 2.6.32-43.97 
> > linux-image-2.6.35-32-server - 2.6.35-32.68~lucid1 
> > linux-image-2.6.38-16-server - 2.6.38-16.67~lucid1 
> > linux-image-3.0.0-26-server  - 3.0.0-26.43~lucid1
> > linux-image-3.2-5 - mainline 3.2.5 kernel
> > 
> > I'm guess it's a libc issue (or at least a libc change causing the issue) as it doesn't seem to a be kernel related.
> > 
> > I'll try other distributions as a guest (probably Debian/Ubuntu) with newer libc's and see if I can pinpoint the issue to a libc version.  Any other ideas?
> > 
> > Shared disk backend is clvm/LV via FC to EMC SAN, not sure what else might be relevant.
> > 
> > Thanks,
> > Mark
> > 
> > 
> > ______________________________________________
> > 
> > See http://www.peak6.com/email_disclaimer/ for terms and conditions 
> > related to this email
> > --
> > To unsubscribe from this list: send the line "unsubscribe kvm" in the 
> > body of a message to majordomo@vger.kernel.org More majordomo info at  
> > http://vger.kernel.org/majordomo-info.html