The X Factor: Virtualization Belongs in the System, Not in the Software
Published: April 10, 2007
by Timothy Prickett Morgan
While server virtualization has been a hot topic for a number of years now, and industry juggernaut VMware is poised to break $1 billion in sales in 2007 for its various virtualization hypervisors and related products, this is still a nascent market--at least on X64 server platforms. Which means that it is not too late to rethink how hypervisors are being implemented.
The variety of virtualization technologies on servers allow a single physical machine to be sliced up into multiple virtual machines. There are a number of approaches. One allows a virtual or logical partition to support a whole operating system; IBM's LPARs, Hewlett-Packard's vPars, Sun Microsystems' LDoms, and VMware's ESX Server all do this. Another approach presents a sandbox for applications and end users that looks like a virtual machine even though it really isn't. Sun's Solaris containers, SWsoft's Virtuozzo, and BSD jails have a single operating system kernel and file system and present virtual private server spaces to applications that look and act like separate operating systems. In short, one puts the virtualization hypervisor below the operating system kernel and file system, and one puts it above it. To further confuse matters, there are other partitioning approaches, such as with VMware's GSX Server and VMware Server and Microsoft's Virtual Server 2005, which allow an operating system such as Linux or Windows to provide a shell that a virtual machine hypervisor can be loaded into that can in turn support multiple and incompatible operating systems inside virtual machines. The open source Xen hypervisor from XenSource can be embedded inside of an operating system in the latter style, or it can act as a hypervisor on a barebones server that can in turn support many partitions with operating systems side by side.
There is much talk about the merits and shortcomings of these different products and competing approaches. There are similar arguments about the eventual commoditization of hypervisors, the high cost of virtualization software, the lack of file formats and open interfaces into hypervisors, and numerous other issues. People seem to be grappling with the complexity of this technology, which is being used as a means to speed up the deployment of servers, operating systems, and applications as well as making systems more resilient. They resent the cost, too. Virtualizing a server can cost twice as much as the price of a physical server itself, which is just a bit crazy, regardless of the flexibility that these technologies offer. Companies wanting to increase the utilization of their servers through virtualization have been resistant to this pricing. Even as XenSource and Virtual Iron have brought more aggressively priced virtualization technologies to market and even as Intel and Advanced Micro Devices have added hardware features to make it possible to support unmodified operating systems on X64 iron atop of these hypervisors, there is this feeling that the price is too high.
That is why virtualization technologies should be pushed down further into the iron and sold in volume. In short, there should be some way to make these technologies a low-cost part of the system, not an optional and relatively expensive add-on that companies deploy begrudgingly. Server virtualization should also be transparent, invisible. No one wants to know a hypervisor is there.
It is tempting to think that market forces alone will be enough to coerce a standard for server virtualization. VMware may be the money maker today, but the Xen hypervisor is open source, not closed source, and will be compatible with Microsoft's future Viridian hypervisor for Longhorn Server (if not a Xen hypervisor itself if Microsoft's homegrown development doesn't pan out). IBM is experimenting with a Xen hypervisor for its Power processors, and Sun is providing support on its Opteron-based systems running Solaris. It is feasible that a Xen hypervisor could end up on future Sparc chips, too. While having a Xen hypervisor will not allow applications to run across X64, Power, or Sparc platforms--this is not a Java virtual machine, after all--it would mean that the tools that plug in and manage Xen hypervisors could span the vast majority of commercial systems in the world if this comes to pass.
While such cross-platform compatibility will be useful and commendable--and VMware could be coaxed to do something similar, as can the providers of virtual private servers like SWsoft--this again is not the point. People do not want to think about a hypervisor any more than they think about a motherboard BIOS or a RAID 5 disk controller chip.
A mezzanine card of some sort that plugs into a server board and can virtualize CPUs and I/O--all transparently--would seem to be ideal. A hypervisor is itself a single point of failure in a system, and smart IT shops today would demand that an electronically packaged hypervisor--perhaps implemented as a set of chips and ROM, like a BIOS--would come in redundant pairs with fault tolerant lockstepping to remove this risk. Such virtualization chips and ROMs could be on each motherboard, manufactured in very high volumes like processor, memory, and I/O chips in systems. Since variety is the spice of life, if a standard for a hypervisor chip/ROM interface was created for motherboards, customers could pick and choose the hypervisors they want, and change them if they were unhappy. And because not all customers want to virtualize, IT shops should not be charged for this virtualization hardware; it should be made in an on-demand fashion, activated with a key for a nominal fee, complete with physical-to-virtual conversion tools and virtual-to-physical tools to undo the virtualization if customers decide to do that, too.
All of the vendors of server virtualization hypervisors have agreed the hypervisor will be commoditized eventually, but the truth is, the hypervisors are not yet commodities and they sure are not priced that way. It is time to accelerate this commoditization by pushing virtualization down into the iron, where it rightfully belongs.
Gartner Says It Was "All Over" the Virtualization Effect
IDC Chops Server Forecasts Thanks to Virtualization, Multicore Chips
The X Factor: Virtual Server Sprawl
How Low Can You Go?
Post this story to del.icio.us
Post this story to Digg
Post this story to Slashdot