Server Virtualization

Description

• Until fairly recently, data center operators typically installed at least one physical server per application. When taking into account testing/development, staging, and disaster recovery 3 to 5 servers per application may have been typical.¹ The traditional one workload, one box approach means that most servers run at a low "utilization rate" – the fraction of total computing resources engaged in useful work. A 2012 New York Times article cited two sources that estimated the average server utilization rate to be 6 to 12%.² Another study stated that the one workload, one box approach leads to 90% of all x86 servers running at less than 10% utilization with a typical server running at less than 5% utilization.³
• Server virtualization offers a way to consolidate servers by allowing you to run multiple different workloads on one physical host server. A "virtual server" is a software implementation that executes programs like a real server. Multiple virtual servers can work simultaneously on one physical host server. Therefore, instead of operating many servers at low utilization, virtualization combines the processing power onto fewer servers that operate at higher total utilization. (See Figure 1 below.)

  

• Virtualization improves scalability, reduces downtime, and enables faster deployments. In addition, it speeds up disaster recovery efforts because virtual servers can restart applications much more rapidly than physical servers. With virtualization, you can move entire systems from one physical server to another in just a few seconds to optimize workloads or to perform maintenance without causing downtime. Some virtualization solutions also have built-in resiliency features, such as high availability, load balancing and failover capabilities.⁴
• Due to these benefits, virtualization has become commonplace in large data centers. A 2011 survey of over 500 large enterprise data centers found that 92% use virtualization to some degree.⁵ Of those, the ratio of virtual servers to physical host server averaged 6.3 to 1 and 39% of all servers were virtual.
• However, virtualization is less common in small data centers. A 2012 NRDC paper entitled Small Server Rooms, Big Energy Savings⁶ included an informal survey of 30 small businesses (ranging from 3 to 750 employees) and found that only 37% used virtualization.

Savings and Costs

• Virtualization enables you to use fewer servers, thus decreasing electricity consumption and waste heat. One watt-hour of energy savings at the server level results in roughly 1.9 watt-hours of facility-level energy savings by reducing energy waste in the power infrastructure (power distribution unit, UPS, building transformers) and reducing energy needed to cool the waste heat produced by the server.⁷
• Virtualization enables the repurposing and decommissioning of some existing servers. According to the Uptime Institute, decommissioning a single 1U rack server can annually save $500 in energy, $500 in operating system licenses, and $1,500 in hardware maintenance costs.⁸
• In a 2007 study, the University of Santa Cruz used virtualization to run 54 virtual servers on 8 physical hosts, reducing peak demand by 20 kW and saving $22,000 in energy annually.⁹
• In a 2010 study, Southwestern Illinois College performed a detailed 3-year total cost of ownership analysis for a 35 server upgrade with and without virtualization (see table below).¹⁰ A system with 35 virtual servers on 4 physical host servers saved over $280,000 in total savings, from savings of:
  • Nearly $150,000 in direct costs: the costs for VM Ware and SAN were offset by substantial savings in server hardware and networking.
  • Over $130,000 in indirect costs: including substantial savings in the "data center" (power, cooling), server provisioning, and procurement.
• A 2009 IDC study examined the annual savings at a Landmark Healthcare virtualization effort where 63 physical servers were replaced with 63 virtual servers operating on 3 physical host servers.¹¹ Even with $3,600 in virtualization software costs, the effort saved $60,000 annually due to savings in server hardware, backup system software, and operating system licenses.
• A Cisco study found that a virtualized server costs about $2,000 to deploy compared to $7,000 for a standard physical server with 2 CPUs due to reduced labor and hardware costs.¹²
• Note that some utility companies offer rebates that can help offset the costs of virtualization projects. Please review the EPA ENERGY STAR Utility Guide for Designing Incentive Programs Focused on Data Center Efficiency Measures for a list of utilities that offer data center focused rebates.

Considerations¹³

• Take inventory. It is easy to lose track of servers as time goes by, as mergers and acquisitions take place, as people procure servers without going through central IT, or as servers are retired or repurposed. The first step toward virtualization is identifying all servers in the organization. Take an inventory of computing resources and their associated application workloads.
• Identify servers that should not be considered for virtualization. Servers that have privacy, security, or regulatory restrictions, or require ultra-high service levels, tend not to be the best candidates for virtualization or consolidation.
• Group remaining servers by workloads. Group the server inventory into pools of workloads that can work off of the same physical server at the same time. "Development" workloads and "Production" workloads would most likely not work well together. As previously mentioned, some servers may not be able to share resources due to technical, security, privacy, or regulatory concerns. Answering the following questions allows you to pool workloads into groups and determine the computing resources required for each group:
  • Who owns the server?
  • What is the purpose of the server?
  • What is the service level requirement?
  • What is the location of the server?
  • Is there a requirement for this workload to remain on a dedicated physical resource?
  For example, the exercise may yield a list of 100 "Innovation" workloads that all reside in the same location. You then can determine the resource requirements for the entire group. In this case, you may need 5 to 10 servers to support the resource requirements of those 100 workloads.
• Examine memory. Virtual host servers tax memory more than CPU utilization. A recent Information Week survey of large data centers indicated that memory is more important than the CPU when evaluating servers in a virtualized environment.¹⁴ The survey found that the median memory capacity was 48GB for a server, with nearly one quarter of large data centers buying systems with more than 128 GB of RAM.
• Look into shared storage. In general, a virtual server system will require networked storage as physical host servers must be able to address the same storage at the same time. Shared storage can come in many different forms: Internet Small Computing System Interface (iSCSI), Network File System (NFS), and Fibre Channel.¹⁵
• Estimate processor resources required. Although not the resource that limits virtualization, processor requirements can be estimated by multiplying the reported utilization (% of processor utilized) by the total processor capacity (maximum processor frequency * number of processors * number of cores). This produces a GHz requirement that can be applied to target servers.
• Keep in mind that application workload can depend on the time of day, week, month or longer if the business cycle dictates. For example, some applications are used primarily during business hours, while others peak during off hours. Servers should be sampled at least once per hour over a 30-day period to develop a complete workload profile. This technique ensures that peak- and sustained-load times are captured in addition to normal usage.
• Combine workloads that have complementary characteristics. For example: Workload A consumes 10% processor capacity during business hours and very little at night. Workload B consumes 10% processor capacity during the night and very little during business hours. Combining the two workloads delivers the best possible results.
• Keep an eye on cooling. For example, consolidating higher workloads into a smaller footprint is likely to create "hot spots" in your data center. You will need to work with HVAC experts to ensure an adequate and efficient cooling configuration.
• Track all virtual machines (VMs). Easy to deploy, VMs can get out of control if you are not careful. If new VMs are created for specific projects, be sure they are not forgotten once those projects are completed.