In the past few weeks I had the chance to have a closer look at the current release 2.2. The reason is that I’m working on a project using RHEL6 clients as virtual Desktops. For a proof-of-concept I’ve set up a test environment in the lab. Due to the lack of time I was not able to test every single feature.
After reading some docs, it was amazingly easily and quickly installed.
The tests have been made on the following hardware:
- 2 Xeon servers with a total of 8 cores and 24Gbyte on each host. OS is RHEL5.6. Those hosts are called RHEV-H (H like Hypervisor)
- 1 Windows 2008R2 server (64bit) with 4Gbyte RAM as RHEV-M (M like management)
- 1 dedicated NetApp filer serving storage over NFS
- 1 RHEL 5.6 Server providing “cheap” NFS storage
- 3 different networks connected to 3 NICs
- 1 RHEL 5.6 “thin” client with spice-xpi as client
- 1 Windows 2008R2 Server with spice-activex as client (actually on the management server)
VM’s (All 64bit):
- 20 Virtual desktops with RHEL 6.0 clients using one core and 2Gbyte or RAM each
- 2 Virtual servers with RHEL 6.0 server using 2 cores and 2Gbyte of RAM each
- 2 Virtual servers with RHEL 5.6 server using 2 cores and 2Gbyte of RAM each
The Management interface is the central point of administration. It is not always as intuitive as expected. If you are new to RHEV, you can get confused. A big plus is the search functionality. One can search for storage, VMs, hosts and any other item in the environment.
The user portal is very simple, lean and clean. Users see the machines that are assigned to them, they can power on and off the machines and can connect them to powered VMs, and the client appears on the desktop.
The user portal resides on the management server and is just the connection broker. As soon as a user connects to its desktop or server, the connection will be made directly to the spiced qemu-kvm running on one of the hypervisors.
At the moment there are two different clients supported:
- spice-activex with Internet Explorer 7+ on Windows XP, Vista and 7
- spice-xpi with Firefox for RHEL5 and RHEL6
If you do not fear the work, you probably can get the spice-xpi client also running on Fedora 13 and other Linux distributions.
The client communicates with the VM over SPICE (Simplified Protocol for Independent Computing Environments). It can forward USB devices as well as sound and other multimedia stuff. For more informations about the protocol, please have a look at http://www.spice-space.org/home.html.
The first step to set up a RHEV environment is to define and set up the storage.
There are 3 types of storage:
- ISO, where you upload your install-media for your virtual machines.
- Data, where the VM images are stored.
- Export, where I have no clue (yet) what it is good for.
Usually the storage is a NFS or iSCSI server, FC storage is also supported. The storage server creates snapshots of volumes for backing up the stuff.
You can either use RHEV-H as an “embedded” hypervisor which comes as a stripped-down RHEL5 distribution (the ISO is just about 75Mbyte IIRC) or you can set up RHEL5 and use it as a hypervisor.
First I tried the RHEV-H, it is manual work to install it. You need to burn a CDROM put it in your server and enter some stuff like network config etc. In the short time I did not tried to provision RHEV-H with cobbler. According to the documentation one can pass boot parameters to partially automate the installation.
Because I have a cobbler server handy, I decided to use RHEL5 as hypervisors. Just be sure you subscribe the system to the channels “rhel-x86_64-server-vt-5″ and rhel-x86_64-rhev-mgmt-agent-5. You also need to allow root to log in via SSH. You can change this later after the hosts is registered on the RHEV-M server.
After the setup of the RHEL5 servers you just need to tell RHEV-M that there is a new host available. Tell RHEV-M about the root password and all the needed additional software gets installed automatically.
That’s a pretty lean procedure to install and set up hosts. Hopefully there will soon be a way to fully automate this, w/o using MS Powershell.
The technologies on the hosts are the well known stable and mature KVM and qemu-kvm. There are also parts of o-virt being used.
I do not know how many networks you can add, but I think enough even for large environments. The network configuration part in RHEV-M shows a list of available Ethernet ports on each host. Just assign them a static network and IP address and you’re done. Make sure you define the network you want to add on all hosts in a particular datacenter.
Ensure that firewalls between the users and VM’s networks are configured accordingly to avoid connection problems.
Sizing and performance
- Memory overcommitment
Thanks to KSM (Kernel Samepage Merging) one can quite overcommit the memory. Ensure you have a LOT of swap on your hosts. Why? It takes some time until KSM kicks in and frees memory pages. During the collection of those pages, swap space is needed to prevent getting a visit by the OOM killer.
I once faced a complete crash when putting one host in maintenance mode and all VMs have been migrated to the second host. That’s exactly why you should have LOTs of swap spaces to survive when using memory over-commitment. During the period of KSM searching for same pages, performance is degraded.
Under normal circumstances, this will never happen since the virtual hosts are load balanced. This means VM’s are distributed to have the optimal performance.
- CPU overcommitment
This very depends on your users computing needs. For normal desktops and servers you can overcommit at least 200% of the CPUs and the performance is still fine. For CPU intensive workload it is not recommended to over commit CPUs.
- Number of hosts
Its recommended to start with at least three hosts to avoid temporary performance penalties if you need to put a host into maintenance mode. (See also “Memory overcommitment”).
The performance is surprisingly good, the user experience is nearly the same as on physical desktops. For servers, KVM is already known for its good performance.
Unfortunately you can not hot-add or hot-remove virtual CPU’s, NICs and storage. Despite of being supported by the underlying technologies used, your VM must be powered off to add or remove resources.
Another drawback is the lack of different storage classes, this makes it actually impossible to use the product in large enterprise environments. Since it makes no sense to back up swap drives, it is waste of backup space.
It is also not possible to have different storage for different needs. Lets say a bulk RAID5 consisting of cheap 2Tbyte SATA disks for operating systems and a RAID10 consisting of fast 15k SAS or FC disks for data and swap. It is also not possible to life-migrate VMs from one storage to another.
As of today, SPICE is only useful on LAN networks. Depending on the work load it can demand lots of bandwidth available. For usual office workload, 1 Mbps per connection is fine. A WAN optimized version is under development and should be released “later this year”.
The two supported clients (spice-xpi and spice-activex) can be problematic when the clients are not in an environment controlled by you. Why? Lots of enterprises are preventing its users to install Active X applications and Firefox Add-On’s for security reasons. It would be better to have some kind of “fat client” which does not need to be installed, or a Java Applet which does the job.
At the moment, it is still required to have a Windows 2008R2 server installed and use MS technologies such as MSSQL, DotNET, Powershell and Active Directory. Because you can not manage RHEV-M with a satellite, one needs to download and install updates manually.
RHEV-M is not only the management server, it is also the connection broker for clients connecting. RHEV-M is a single point of failure. Maybe one can build a MS-Cluster with it, no clue. It is also not possible to self-host RHEV-M on the RHEV-H hosts. Already connected clients should not be disconnected in the case of RHEV-M failing (to be tested).
The same comment talks about the WAN problems to be solved in near future.
Frankly: From my point of view, RHEV is not yet “enterprise ready”, this is because of the drawbacks mentioned above, especially the storage shortcomings and RHEV-M being a SPoF.
For smaller environments up to lets say ~50 servers or ~200 desktops it good enough.
Nevertheless: I think RHEV has a huge potential in the future. Compared to VMware ESX, the KVM’s technology is much better and scales better. At the moment VMware’s USP is its management software, not the hypervisor used. As soon as Red Hat ironed out the major drawbacks, I’ll expect a boost for RHEV.
Red Hat is keen to improve the product, I guess a lot of improvements will be annouced in 2011.