[libvirt] <soapbox>Thoughts on Open Source Hypervisor management

Nathan Charles ncharles at gmail.com
Mon Sep 8 22:11:59 UTC 2008

Over the the last few years I've setup and experimented with a large number
of vm technologies.  Everything from x86 based stuff like VMware, kvm, qemu,
xen, to more exotic things like hercules, and z/VM.  So far the solution
I've found to be the most baked is VMware ESX with Virtual Center.  It's
pretty sweet to set up a DRS cluster and have VM's load balance
across available hardware.  However, It's super expensive and is in bed with
windows way to much.  i.e.  Virtual Center and the Virtual Infrastructure
Client only run on windows don't have LDAP auth and only use MSQL.

I've also been working implementing a provisioning solution for our
developers.  We do automated installation of hundreds of machines of which
many are underutilized.  Many are newer machines with some pretty beefy
specs (8 cores and at least 8 gigs of ram).  I've done a lot of thinking in
a vacuum about what would be ideal.  There's a lot out there I don't know so
I hope you give me some grace as I talk.   So I want to be able to pull
these machines into a cluster of hypervisors; a VM fabric aggregated via
software. Or thought of in another way a p2p vm cluster.   In some ways I
think oVirt is the closest and in someways I think simply extending libvirt
would be better and sometimes I think it should be a new project.  So I
thought I'd post this and see what people think.   Since I'm still unsure as
to where the proper place to abstract this is, I'll use the working package
name of genet for the rest of the discussion.

>From my perspective the reason I can't use most of the open source
virtualization solutions for more than local vm's is one of management.
They're great for one machine but as soon as you get to more than one
machine there is a lot of human administrative overhead.  oVirt seems like
it could be really good, but in an existing environment it requires a lot of
substantial changes and requires significant modification to work with an
existing provisioning infrastructure.  So as I imagined a silver bullet I
used biology as a metaphor and used started thinking with the idea of
"every piece contains the image of the whole."   So the here's the 10,000 ft
overview of an idea.

1. cluster administration is done from the command line
2. cluster administration can be performed from any node.
3. a new node can join a cluster on a local subnet with one command.
4. local storage resources are presented to the cluster so there is no need
to have predefined NAS/SAN/iSCSI
5. cluster will load balance vm instances from node to node.
6. a node shouldn't need more than one nic but adding additional nic's
provides failover and load balancing.

I'm kind of ignoring storage at this point, but I think it's the biggest
pieces of the puzzle as long as migration needs network/san storage.  The
ideal scenario would be something really simple like if I have SAN
connectivity use that else I use some sort of cluster file system which adds
in local storage and doesn't require reformating, but that might be wistful

Use cases
1. looking for clusters on the same subnet via avahi
# [yum|apt-get] install genet
# genet list clusters

# rhizome info sumac
sumac info:   8 nodes
  64 Cores
  128 G Ram
  1.5 T storage
   50% Cluster CPU utilized
  75% Memory utilized
  80% Storage utilized

2. a machine joins an existing cluster
# genet join sumac
localhost has successfully joined sumac

Help output
#genet -h
usage: genet [options] command

genet is a simple command line interface for creating and maintaining a
cluster of hypervisors.

   create cluster [cluster name]
   create vm [vm name] (probably integrates with virt-install)
   evacuate [hostname] - migrate all vm instances off of [hostname]
(defaults to localhost)
   list clusters [hostname] - list clusters on local subnet or via hostname
on another network/subnet
   list vm [cluster] - list all vm instances in a cluster defaults to
cluster local node is in.
   console [vmname] - virt-viewer to vmname

So this all begs the the statement, "show me the code" and I think that
perfectly valid.  I may write some code, but I don't want to duplicate
effort and I want to find out what other people are thinking.  I welcome
your thoughts. </soapbox>


Nathan Charles
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listman.redhat.com/archives/libvir-list/attachments/20080908/339a21f5/attachment-0001.htm>

More information about the libvir-list mailing list