In-Depth
Getting Real About Server Virtualization
Some manufacturers are buying so much of the advertising in magazines that articles critical of certain products never see the light of day.
When you sift back through the tons of marketing material from the tech industry in 2008, a disproportionate amount seems to have been generated by the server virtualization people -- specifically, VMware. To hear the story according to the vendor marketing departments, the analyst community, and the press, server virtualization is the ultimate cure-all for what ails IT.
When you think about it, server virtualization is like plastic surgery. Metaphorically, you have a couple of average-looking server administrators, one male, one female, who go to the neighborhood plastic surgeon to get work done and emerge looking like supermodels.
The punchline: The surgically-altered techies meet at Starbucks, fall in love, get married, and produce ugly offspring. Why? Because plastic surgery, like virtualization, only changes the presentation layer -- it doesn't change the underlying genetics.
Server virtualization is simply an abstraction layer that obfuscates the real issues of systems integration and application misbehavior like so many subprime mortgage derivatives. The best you can say is what Ed Walsh, CEO of Virtual Iron, recently told me in an interview, "Virtualization is [just] another tool and one that can be used effectively and affordably to accomplish discrete, well-defined objectives." It is by no means a panacea.
I ran into a problem last year when speaking at a conference (sponsored in part by VMware). I had a slide in my deck that noted (rather graciously, in my opinion) that there was nothing intrinsically right or wrong with server virtualization, but that the products needed to be used judiciously. That word "judiciously" roiled a VMware representative in the audience and created a bit of a scene after the show.
"How dare he say that our product should be deployed judiciously!" exclaimed the representative, annoyed that his company's sponsorship of the event did not merit a more positive endorsement of his wares. My effort to explain what judiciously means did not assuage his ire. "Deployment after careful consideration and by a conservative process of limited testing and measurement of results" was apparently not what VMware wanted customers to do. "We are not just a product," said the fellow, "we are a movement."
That comment sent a chill up my spine.
From my perspective, there’s been a dearth of criticism of the hype around server virtualization. The tales told by paid analysts of widespread adoption were beginning to grate. Moreover, they did not ring true in the polling that I had done.
For one thing, with the exception of testing and development labs in companies -- which were using the software layer to build up and tear down virtual machines (VMs) in order to do more testing with less equipment -- the preponderance of those deploying server virtualization were using it mainly to consolidate file servers and low-traffic Web sites. I wasn't seeing too many SQL Server or Oracle databases, or even Exchange Mail installations, being migrated to VMs.
There are ways to consolidate file servers -- via global namespaces or more simply by making their contents subfolders of a master file server -- that didn't require server virtualization at all. As for low-traffic (or even high-traffic) Web server consolidation, there were products such as Plesk that could be licensed to provide a solid management front end for many sites operating from the same box. Again, no virtualization needed. Where are the articles about this?
To get the 20:1 or 60:1 server consolidation ratios that VMware was touting, you needed a lot more than the basic software. You were looking at a substantial investment in additional wares and training from VMware that significantly altered the cost-benefit analysis proffered by the vendor. To create a 20-socket VMware environment, with all the bells and whistles for high availability, you needed to buy a couple of expensive servers and add in about $75K in software for both. This was a far cry from VMware’s model: take an underutilized server you already own and throw on the $3K hypervisor kit to consolidate 20 or so physical boxes.
In addition, there are risks and vulnerabilities of server consolidation. Stacking up a lot of virtual machines atop x86 chip extent code using a hypervisor that brokered machine resources to each app sounds cool enough, but we have been finding that a lot of things could go wrong. For one, if an application made an "illegal" resource request that wasn't caught by the hypervisor (as many Microsoft and third-party apps are keen to do), the stack would become as unstable as a tower of playing cards.
Though instability may not be the fault of the hypervisor but of a "misbehaving application," the result would be the same. Lacking the VM isolation that has been perfected in the magical world of mainframe logical partitioning (LPAR) technology over two decades, not only the misbehaving app, but all apps in VMs stacked in the system would fail. How's that for high availability?
Instability can also come from below rather than above. Think about thin provisioning in some storage arrays. Thin provisioning is a high-tech shell game that takes space from the applications that own it and reallocates it under the covers to other users. Vendors promoting the technology claim that this optimizes resource allocation efficiency and removes the burden to manage capacity that is the bane of existence for increasingly short-handed IT departments.
While there is certainly some truth in these assertions, consider what would happen if the application that owns the space ever does the equivalent of a "margin call" -- the result of a spike in activity in a database, for example. Things can get very weird very fast.
Here's a scenario discussed in the Q&A session of an event I attended this year. When a margin call occurs, first, the application, starved of its storage resource, will abend. If that application resides in a VM, chances are better than even that the entire VM stack will fail, as likely will the server OS. Then, in the words of the CTO for the thin-provisioning storage vendor who was being grilled in the session, "Smoke will probably start billowing out of the sides of the server."
Of course, the thin provisioning storage vendors claim that their capacity-forecasting algorithms are so good that this could never happen. That doesn't reassure me. Given that there are no standards for forecasting storage capacity and the algorithms developed by the vendors themselves are proprietary and inaccessible for review, I have to wonder about the meaningfulness of vendor reassurances. Were I an IT consumer, I would be very concerned about this situation, and would at least consider it in my judicious review of server virtualization efficacy.
Given that most of the three-letter storage vendors are planning to implement thin provisioning on their 2009 wares in response to the success of 3PAR, Compellent and others in the array business who are attracting some of their customers, I would be very concerned.
We have been testing server virtualization wares in my test lab for most of the year, including VMware ESX, Virtual Iron, and Hyper-V from Microsoft. We haven't had the opportunity to look at Citrix Zen or other products that are coming along now, so we don't have quite enough data for a formal bake-off, but what we have learned is telling.
I bounced a few of our thoughts off of Virtual Iron's Walsh, in part because I have known him for many years during his tenure as a senior manager for both de-duplication pioneer Avamar Technologies and storage switch vendor CNT.
Walsh confirmed that business was good at Virtual Iron, bragging about a record quarter with 30 percent growth. He said that business was shifting from a model in which customers start with a couple of "sockets," then add 50 more, to one in which they buy lots of sockets for building virtual machines right from the start. "There is no indication of a slow down," he said, providing survey findings that showed the drivers of adoption from the consumer's point of view.
In the survey, conducted by Tech Validate (sample size undisclosed), the major reason cited by consumers for buying virtualization products was to consolidate servers. This is not surprising to Walsh, who observed that in the medium-to-large company space that he describes as his market, there is a widespread view that multi-core servers are not being used to their full potential.
"Most applications don't use the additional cores (except for Linux environments) and every box has spare capacity, even low-end, four-core systems. Basically, our virtualization solution is riding the trends of the chip makers, Intel and AMD, by providing the means to use the additional core processors and the extent code on the chips themselves."
His survey data sees customers preferring Virtual Iron to VMware because of price. Says Walsh, Virtual Iron has an MSRP of about $799, about a quarter of the price of a comparable VMware kit, and a full functionality set that includes high availability and disaster recovery failover out of the box. Kudos were also given to Virtual Iron for quality of service and support by the surveyed customers. Altogether, this accounts for the finding that 93 percent would buy from Virtual Iron again.
What Virtual Iron doesn't have is a proprietary file system or VMware's one-and-a-half week off-site training requirement, Walsh observed, which have turned off many companies interested in exploring server virtualization but wary of hidden costs. Our testing agrees with Walsh's claims that Virtual Iron's installation is more intuitive than VMware’s, and the access to online video instruction and real-time support is a big plus.
Management is key to understanding how resources are being used in the virtual world, and Virtual Iron provides VI Center to report and track resource utilization. This facility contributes to the ease of use praise that is also cited in the Tech Validate survey.
Virtual Iron has been attacked by competitors because of its dependency on hardware to deliver hypervisor functionality. Walsh says that that used to be an issue but lost its luster as more and more companies deployed chipsets that featured X86 extent code. Now, the attack has become a plus, in his view.
"VMware existed before the extent code was on the chips, so they did things like binary translation in software, supporting older technology," Walsh claimed. "Now it is VMware that is working to catch up to chip enhancements that we already support."
Walsh said that the key elements to look at when examining server virtualization products are provisions to "isolate applications that are acting badly … to move other virtual machines away so that a catastrophic failure does not occur." He added that failover and rebalance capabilities are also paramount. "Commodity hardware fails whether it is virtualized or not, so if the application is critical, you want it to failover smoothly and alert you to what has happened."
Says Walsh, VMware doesn't provide this functionality out of the box. You need substantial software upgrades, including Site Recovery Manager, "which I’ve been told is actually an EMC failover product under the covers." He says that Virtual Iron does not provide disaster failover across a WAN but works well with the same EMC product that VMware calls Site Recovery Manager, as well as failover wrapper products from numerous vendors, including CA XOsoft, DoubleTake, and NeverFail Group.
He also cautioned consumers about the limitations of HA and failover capabilities in Microsoft Hyper-V, noting that Hyper-V provides baseline virtualization capabilities (like basic VMware) but none of the HA or disaster recovery capabilities that Virtual Iron delivers out of the box. "With Virtual Iron, our customers realize payback on their investment within six months versus Microsoft, which will not deliver the desired feature set until 2010."
The bottom line: server virtualization, used judiciously, may be a useful tool in server consolidation and hardware optimization efforts. However, these values must be carefully weighed against alternatives that may not require adding new software to the stack at all. When it is appropriate to virtualize a server, consider the options carefully. Brand name is becoming less important than functionality and ease of use. At the end of the day, application stability and performance count more than all the marketecture in the world.
Your opinions are welcome: jtoigo@toigopartners.com.