The value virtualization brings to datacenters depends on what problems it can solve for the IT technologist. These are defined to be virtualization's usage models. Virtualization usage models today are focused on legacy consolidation, flexible provisioning, test/development, dynamic load balancing, and disaster recovery. As the technology penetrates the mainstream datacenter, new usage models will likely emerge.

The performance discipline¹ presented in this paper provides a basic framework fitting multiple usage models. However, we focus specifically on two usage models due to their historic appeal to datacenter managers: legacy consolidation and flexible provisioning

Legacy consolidation refers to transferring application and OS stacks from multiple servers to a fewer number of (typically more powerful) servers. The old "legacy" servers are often running older OSs because upgrading to a newer version has some manpower impacts. For example, a legacy application may need some porting work, but the resources to do this cannot be justified. In this paper an old (legacy) server is classified to be 3+ years old. The utilization level of these legacy systems is often quite low, 5-15%. Such low utilization of an older platform means that a contemporary server could provide the processing capability for running many multiples of the legacy applications simultaneously. Market research indicates that legacy consolidation projects with virtualization tend to mix different application types on the new consolidated server; however, IT managers are avoiding mixing OS types [3]. The final consolidated server typically has some mixture of Web servers, e-mail servers, database servers, and/or other types of applications running simultaneously, but all are running the same type of OS (e.g., Linux* or Windows*) and often the same version of the OS [3].

Flexible provisioning is essentially forward-looking consolidation. As new servers are requested from the IT department, rather than deploying a new physical server for each request, the deployment is a virtual slice of an existing server. For example, if a two-processor server is needed to support a new Web server application, perhaps two virtual processors of a multi-processor server are deployed in support of the request. This drives up the utilization of the servers, thereby improving efficiency. The virtual provisioning is often much faster, as hardware may already exist to support the request, eliminating the procurement cycle.

These two usage models, legacy consolidation and flexible provisioning, constitute the main driver behind the adoption of virtualization in mainstream datacenters. The focus of this methodology is to measure the performance of virtual servers in a way that is related to these two usage models. Other usage models, like test and development, are also popular but not addressed in this paper.

Having the usage model defined is important for developing a relevant performance benchmark. The usage model provides boundary conditions to the testing. For example, the usage model indicates what types of applications should be used, whether the application and OS stacks are heterogeneous or homogeneous, what is the configuration of a typical system-under-test (SUT), and what is the appropriate size and number of the virtual machines (VM).