virsh snapshot-create too slow (kvm, qcow2, savevm)

Reported by Martin Kopta on 2011-03-24
This bug affects 5 people
Affects Status Importance Assigned to Milestone
qemu-kvm (Ubuntu)

Bug Description

# time virsh snapshot-create 1

* Taking snapshot of a running KVM virtual machine

Domain snapshot 1300983161 created
real 4m46.994s
user 0m0.000s
sys 0m0.010s

Expected result
* Snapshot taken after few seconds instead of minutes.

* Ubuntu Natty Narwhal upgraded from Lucid and Meerkat, fully updated.

* Stock natty packages of libvirt and qemu installed (libvirt-bin 0.8.8-1ubuntu5; libvirt0 0.8.8-1ubuntu5; qemu-common 0.14.0+noroms-0ubuntu3; qemu-kvm 0.14.0+noroms-0ubuntu3).

* Virtual machine disk format is qcow2 (debian 5 installed)
image: /storage/debian.qcow2
file format: qcow2
virtual size: 10G (10737418240 bytes)
disk size: 1.2G
cluster_size: 65536
Snapshot list:
1 snap01 48M 2011-03-24 09:46:33 00:00:58.899
2 1300979368 58M 2011-03-24 11:09:28 00:01:03.589
3 1300983161 57M 2011-03-24 12:12:41 00:00:51.905

* qcow2 disk is stored on ext4 filesystem, without RAID or LVM or any special setup.

* running guest VM takes about 40M RAM from inside, from outside 576M are given to that machine

* host has fast dual-core pentium cpu with virtualization support, around 8G of RAM and 7200rpm harddrive (dd from urandom to file gives about 20M/s)

* running processes: sshd, atd (empty), crond (empty), libvirtd, tmux, bash, rsyslogd, upstart-socket-bridge, udevd, dnsmasq, iotop (python)

* networking is done by bridging and bonding

Detail description

* Under root, command 'virsh create-snapshot 1' is issued on booted and running KVM machine with debian inside.

* After about four minutes, the process is done.

* 'iotop' shows two 'kvm' processes reading/writing to disk. First one has IO around 1500 K/s, second one has around 400 K/s. That takes about three minutes. Then first process grabs about 3 M/s of IO and suddenly dissapears (1-2 sec). Then second process does about 7.5 M/s of IO for around a 1-2 minutes.

* Snapshot is successfuly created and is usable for reverting or extracting.

* Pretty much the same behaviour occurs when command 'savevm' is issued directly from qemu monitor, without using libvirt at all (actually, virsh snapshot-create just calls 'savevm' to the monitor socket).

* This behaviour was observed on lucid, meerkat, natty and even with git version of libvirt (f44bfb7fb978c9313ce050a1c4149bf04aa0a670). Also slowsave packages from https://bugs.launchpad.net/ubuntu/+source/libvirt/+bug/524447 gave this issue.

Thank you for helping to solve this issue!

ProblemType: Bug
DistroRelease: Ubuntu 11.04
Package: libvirt-bin 0.8.8-1ubuntu5
ProcVersionSignature: Ubuntu 2.6.38-7.38-server 2.6.38
Uname: Linux 2.6.38-7-server x86_64
Architecture: amd64
Date: Thu Mar 24 12:19:41 2011
InstallationMedia: Ubuntu-Server 10.04.2 LTS "Lucid Lynx" - Release amd64 (20110211.1)
SourcePackage: libvirt
UpgradeStatus: No upgrade log present (probably fresh install)

Martin Kopta (martin-kopta) wrote :
description: updated
tags: added: kvm libvirt qcow2 qemu savevm snapshot virsh
removed: apport-bug
Changed in libvirt (Ubuntu):
status: New → Confirmed
importance: Undecided → Medium
affects: libvirt (Ubuntu) → qemu-kvm (Ubuntu)
Serge Hallyn (serge-hallyn) wrote :

Yup, I can definately reproduce this.

Serge Hallyn (serge-hallyn) wrote :

The current upstream qemu.git from git://git.savannah.nongnu.org/qemu.git
also has the slow savevm. However, it's loadvm takes only a few seconds.

savevm _is_ slow, because it's writing to a qcow2 file with full (meta)data allocation which is terrible slow since 0.13 (and 0.12.5) unless you use cache=unsafe. It's the same slowdown as observed with default cache mode when performing an operating system install into a freshly created qcow2 - it may take several hours. To verify, run `iostat -dkx 5' and see how busy (the last column) your disk is during the save - I suspect it'll be about 100%.

Serge Hallyn (serge-hallyn) wrote :

Confirmed that doing

  kvm -drive file=lucid.img,cache=unsafe,index=0,boot=on -m 512M -smp 2 -vnc :1 -monitor stdio

and doing 'savevm savevm5'

takes about 2 seconds.

So, for fast savevm, 'cache=unsafe' is the workaround. Shoudl this bug then be marked invalid, or 'wontfix'?

Martin Kopta (martin-kopta) wrote :

I confirm that without 'cache' option, I have got from iostat those result while doing 'savevm'

Device: sda
rrqm/s: 0.00
wrqm/s: 316.00
r/s: 0.00
w/s: 94.80
rkB/s: 0.00
wkB/s: 1541.60
avgrq-sz: 32.52
avgqu-sz: 0.98
await: 10.32
svctm: 10.10
%util: 95.76

I also confirm, that when option 'cache=unsafe' is used, snapshot (from qemu monitor) is done as quickly as it should (few seconds).

I am not sure if this is a solution or workaround or just a closer description of a bug.

http://libvirt.org/formatdomain.html#elementsDisks describes option 'cache'. When I use that (cache="none") it spits out:

error: Failed to create domain from vm.xml
error: internal error process exited while connecting to monitor: kvm: -drive file=/home/dum8d0g/vms/deb.qcow2,if=none,id=drive-ide0-0-0,boot=on,format=qcow2,cache=none: could not open disk image /home/dum8d0g/vms/deb.qcow2: Invalid argument

When that option is removed, domain is created successfuly. I guess I have another bugreport to fill.

So, for me, the issue is somehow solved from the qemu side. I think, this could be marked as wontfix.

Kevin Wolf (kwolf-redhat) wrote :

In qemu 0.14 cache=writeback and cache=none are expected to perform well. The default cache=writethrough is a very conservative setting which is slow by design. I'm pretty sure that it has always been slow, even before 0.12.5.

I think that the specific problem with savevm may be related to the VM state being saved in too small chunks. With cache=writethrough this will hurt most.

edison (sudison) wrote :

I had posted a patch to fix the issue before:(http://patchwork.ozlabs.org/patch/64346/), saving memory state is time consuming, which may takes several minutes.

Serge Hallyn (serge-hallyn) wrote :


if you want to push such a patch, please do it through upstream, since it is actually a new feature.

I'm going to mark this 'wontfix' (as I thought I had done before), rather than invalid, though the latter still sounds accurate as well.

Changed in qemu-kvm (Ubuntu):
status: Confirmed → Won't Fix
Cinquero (cinquero) wrote :

Cool. Writes about 9 times the data of the actual snapshot size.

To post a comment you must log in.
This report contains Public information  Edit
Everyone can see this information.

Other bug subscribers