[libvirt-users] using shared storage with libvirt/KVM?

Mon May 3 20:19:43 UTC 2010

On 05/03/2010 10:57 AM, David Ehle wrote:

> I've spent a few days googling and reading documentation, but I'm looking for
> clarification and advise on setting up KVM/libvirt with shared storage.

I, too, have been looking at these kinds of setups in the last few 
weeks, and I've found myself confused about the stack more than once. 
The documentation is very scattered and incomplete in key areas (as you 
pointed out with the RHEL docs on shared storage), which makes it tough 
on a new user.  Here's what I can share, I hope this encourages others 
to join in and help us both out.

> I am trying to figure out the best way to set things up so that I can run
> several (eventually) production linux guests (mostly debian stable) on the 2
> KVM host systems and be able to migrate them when maintenance is required on
> their hosts.

  From the small testing I've done, making an iscsi based storage pool 
and using it to store the VMs between the two KVM hosts.

> A common opinion seems to be that using LVs to hold disk images gives possibly
> the best IO performance, followed by raw, then qcow2.
>
> Would you agree or disagree with this? What evidence can you provide?

That's probably right, but I think those are more relevant when you are 
using non-shared storage, like images on your local disk or perhaps on a 
shared mount.

> I have succesfully made an iscsi device available to libvirt/virsh/virt-manager
> via an XML file + pool-define + pool-start. However the documentation states
> that while you can create pools through libvirt, volumes have to be
> pre-allocated:
> http://libvirt.org/storage.html
> "Volumes must be pre-allocated on the iSCSI server, and cannot be created via
> the libvirt APIs."
>
> I'm very unclear on what this means in general, and specifically how you
> preallocate the the Volumes.

I believe this means you will need to create the LUNs on the storage and 
present them to libvirt, but libvirt cannot talk back to the storage 
system and create the targets/LUNs for you as you try to spin up new VMs.

> If you are using shared storage (via NFS or iSCSI) does that also mean you MUST
> use a file based image rather than an LVM LV?

No, you can use iscsi and treat the storage as a raw disk.

> Redhat provides pretty good documentation on doing shared storage/live
> migration for NFS:
> http://www.redhat.com/docs/en-US/Red_Hat_Enterprise_Linux/5.4/html/Virtualization_Guide/chap-Virtualization-KVM_live_migration.html
> But unfortunately the section on Shared Storage with iSCSI is a bit lacking:
> "9.1. Using iSCSI for storing guests
> This section covers using iSCSI-based devices to store virtualized guests." and
> thats it.

I know!  How frustrating!  Even the RHEL6 beta document is missing those 
sections...

Here's what I've documented for my initial test run.  I hope this helps 
you get started, please note that I changed the IPs and iqn's to prevent 
sharing private data on the list.

1. Create a LUN on your shared storage system & get the iqn.

iqn.long.long.long:a6c98a7e-0275-c020-efae-a386dfd46b84

2. From the KVM host, scan for the newly created target

# iscsiadm -m discovery -t sendtargets -p <shared storage FQDN>

1.2.3.4:3260,2 iqn.long.long.long:a6c98a7e-0275-c020-efae-a386dfd46b84

There it is, *84.

3. Write the XML config describing the storage we will add to the 
storage pool, using the iqn from step 2.

<pool type="iscsi">
    <name>lun1</name>
    <source>
      <host name="storage hostname"/>
      <device 
path="iqn.long.long.long:a6c98a7e-0275-c020-efae-a386dfd46b84"/>
    </source>
    <target>
      <path>/dev/disk/by-path</path>
    </target>
</pool>

4. Import that XML to define the pool:

# virsh pool-define /tmp/lun1.xml
Pool lun1 defined from lun1.xml

5. Verify:
# virsh pool-info lun1
Name:           lun1
UUID:           701a1401-547a-08da-5f14-befea84778d9
State:          inactive

6. Start the pool so we can use it

# virsh pool-start lun1
Pool lun1 started

7. List the volume names in the new pool so we'll know what to pass 
along to virt-install to use it:

# virsh vol-list lun1
Name                 Path
-----------------------------------------
10.0.0.0 
/dev/disk/by-path/ip-1.2.3.4:3260-iscsi-iqn.long.long.long:a6c98a7e-0275-c020-efae-a386dfd46b84-lun-0

virsh # vol-info --pool lun1 10.0.0.0
Name:           10.0.0.0
Type:           block
Capacity:       20.00 GB
Allocation:     20.00 GB

8. Install the VM with virt-install and use "--disk 
vol=lun1/10.0.0.0,size=20" to access the storage

[root at rvrt010d ~]# virt-install <all of your options>
--disk vol=lun1/10.0.0.0,size=20

After that's up and running, your VM will use the shared storage as if 
it is a local disk.

9.  If you repeat steps 2-6 on your other KVM host, the storage will be 
shared between both hosts.  Now you are setup for migration, either 
offline or live.

# virsh migrate VM qemu+ssh://other.kvm.host/system

> For maybe 6-10 guests, should I simply just be using NFS? Is its performance
> that much worse than iSCSI for this task?

This totally depends on your network and hardware, but here's how a dd 
test over NFS and iSCSI compared in my setup:

iSCSI
run     throughput
01       24.5 MB/s
02       25.7 MB/s
03       52.4 MB/s
04       56.5 MB/s
05       57.7 MB/s
06       56.2 MB/s
07       51.5 MB/s
08       57.8 MB/s
09       57.3 MB/s
10       55.8 MB/s

NFS
run     throughput
01       34.5 MB/s
02       35.2 MB/s
03       36.2 MB/s
04       41.2 MB/s
05       39.2 MB/s
06       39.2 MB/s
07       36.2 MB/s
08       34.8 MB/s
09       36.7 MB/s
10       40.4 MB/s

I hope this helps, and I hope that someone corrects any errors I may 
have made.  :)

Tom