"When you commit to a hypervisor,
Although it requires longer-term planning, hypervisor installation is relatively simple. The hypervisor is installed as software similar to that of an OS or critical application. Embedded hypervisors don't require any installation per se since the firmware is already provided in the system hardware itself.
Success with any newly deployed hypervisor depends on several factors. First, all hypervisor installations will impose a performance penalty on the workload. It's a small penalty, and hypervisor efficiency is constantly improving, but administrators must regard its effect on demanding applications, especially I/O-intensive workloads. Workload planning and balancing can offset small penalties, and new hardware platforms can be deployed to support the most demanding tasks.
Workload balance is also reflected in consolidation. Experts underscore the importance of balancing consolidation levels with recovery needs. "If a 4U server fails with 60 virtual machines on it, [then] I have 60 applications that are offline that IT has to deal with," Wolf said. "Even if I have high availability implemented, that's still 60 applications offline for a couple of minutes until those VMs can be restarted."
The ability to support and move multiple workloads without disruptions is fundamental to hypervisors. However, you should also evaluate the following features before you install a hypervisor in your enterprise:
Workload support. Ensure that a hypervisor is fully tested and compatible with the OSes that you intend to virtualize. For example, Citrix XenServer can support virtual machines (VMs) on Windows Server 2000 through 2008, Windows Vista, Windows XP SP3 and a variety of Linux distributions. Other systems may not operate properly as virtual machines.
You should also think about the applications that make up a workload. While many current enterprise applications can support virtualization, some may not perform well in a VM. This isn't as much an issue with processor-intensive programs, but it can happen with some I/O-intensive applications.
To evaluate the hypervisor's scalability and its ability to handle additional workloads as well as oversubscribed physical CPU cores, run multiple workloads on the same processor. Some workloads behave best with a traditional 1:1 ratio of physical cores to VMs. Other workloads such as test and development or virtual desktops may demand access to additional cores.
"You want the hypervisor to be able to intelligently manage oversubscription by as high as 8:1," Wolf said. An older server with four cores might be required to handle up to 32 VMs, he noted.
Processor acceleration requirements. Intel VT and AMD-V processors include hardware assistance for virtualization. Although both approaches differ, each affects the way that memory is handled in an x86 architecture. Hypervisors like Microsoft Hyper-V may require Intel VT or AMD-V processors, so it's important to know that your servers meet system requirements for the intended hypervisor.
Power management and resource optimization support. This is an increasingly important issue for energy-efficient data centers. Hypervisors generally don't support traditional sleep states like standby or hibernate; these features won't work when the hypervisor is enabled.
The hypervisor should support power management and afford a level of control over processor power consumption as server utilization changes. If the hypervisor doesn't handle power management directly, it should interface with power-management capabilities in the OS or other performance-management software.
The hypervisor should also be able to move system resources around -- allocating more or less computing power depending on the unique resource needs of each workload. One well-noted example of this capability is the VMware Distributed Resource Scheduler (DRS) utility.
"[VMware DRS] allows you to automatically balance workloads," said Gary Chen, research manager of enterprise virtualization software at IDC, an IT industry analyst firm headquartered in Framingham, Mass. "They [administrators] can also consolidate and power off servers to save power during off-peak times."
System device support. Some virtual devices are simply not available to a VM -- usually because of missing or inadequate driver support. For example, the hypervisor may not support older network adapters or SCSI controllers under certain OSes. Similarly, USB device support is often absent under VMs in a production environment. Verify hypervisor support if your server uses legacy adapters or controllers as well as USB devices.
Security. The challenge with hypervisors is that a security breach in one workload, or in the hypervisor level itself, can potentially jeopardize all other workloads on a server. This makes security a high priority for administrators who should protect applications such as directory services integration, administrative action logs and role-based access controls.
Resiliency. Virtualization can improve the reliability of a server and its network by adding high availability (HA), fault tolerance, business continuance (BC) and disaster recovery (DR) features. In some cases, resiliency may be a matter of migration -- moving a workload from a failing server to a working server or spinning up a failed instance from storage onto an auxiliary server. In other cases, multiple instances of the same workloads can run redundantly, reducing recovery time when a fault occurs.
Robustness. There are very few features or capabilities that are noticeably absent from modern hypervisors -- most of them are already quite capable and streamlined. "The performance is there for 90% of the market and applications," Chen said.
Virtualization experts like Chen point out several hypervisor capabilities that can be added or improved. Tony Iams, vice president and senior analyst of system software research at Ideas International Ltd., a global IT analysis firm in Rye Brook, N.Y., points to the increasing importance of management tools and middleware. Both of these exploit core abilities of the hypervisor and can sometimes fill in the gaps for what hypervisors may be missing. "You need to look at the add-ons and extensions," Iams said.
Having the ability to over-commit memory -- allocate more memory than what's available --would be noteworthy for memory-intensive virtualization tasks, especially desktop virtualization where many instances can quickly sap memory space. Security features are another facet of hypervisors that could benefit from some improvement, but Wolf suggests the primary need would be better integration with VM management tools.
Virtual machine allocation might also benefit from properly planning an application's I/O load rather than simply examining CPU and RAM requirements. Iams takes an even broader view on this, noting that hypervisors should start the next wave of differentiation by supporting other facets of the data center makeup -- storage solutions and the network, for example.
This was first published in December 2009