|
Virtualization is a key strategy for simplifying deployment of IT resources and maximizing their utilization. Specifically, virtualization refers to the concept of abstracting physical resources such as compute cycles, data storage, and network bandwidth, and then provisioning and sharing these resources amongst multiple applications. For example, a single server may be "virtualized" to allow multiple operating system (OS) images to run concurrently; the amount of storage available to a user on a Storage Area Network (SAN) may be dynamically adjusted on the fly; and the amount of bandwidth allocated to a given application may be boosted or reduced as required. Virtualization began as a niche market but is rapidly gaining acceptance as the preferred way to manage and provision system resources within a network.
The benefits of virtualization are well understood. System administrators are able to capture underutilized resources and re-allocate them to constrained applications. Resources can be dynamically allocated and load-balanced as the characteristics of traffic and applications change over time. Hardware can be transparently replaced or upgraded with a minimum of downtime. Utilizing existing resources more efficiently in this way leads to reduced infrastructure cost, better utilization of IT assets, lower power consumption, reduced cooling requirements, and inevitably lower total cost of ownership.
Today, server virtualization is primarily handled through specialized software provided by vendors such as VMware. The success of this technology has led system architects to think about ways to increase virtualization's effectiveness by extending the same concepts to the hardware level. Companies such as Intel, with its Virtual Technology (a.k.a Vanderpool), and AMD, with "Pacifica", are implementing virtualization-specific features in their CPUs. Following along the same path, input/output (I/O) architectures are now being redesigned to support this powerful concept from end to end, with the introduction of hardware-based I/O virtualization, commonly referred to as IOV.
Neterion's Hyperframe? IOV architecture offers the industry's first comprehensive set of I/O virtualization features implemented at the hardware level. Neterion's Xframe? line of 10 Gigabit Ethernet adapters is built on the principles of the Hyperframe architecture.
The Need for IOV
The basic principle of IOV is to share an I/O (or pool of I/O) sub-systems among various compute systems ? different images of an Operating System in a server, or different blades in a chassis, each running one or multiple images of an OS. An example of an I/O sub-system is a network interface card (NIC). Implementing an IOV architecture for the network interface means abstracting one or several NICs and making them appear like a multitude of independent interfaces to the virtualized compute systems.
One of the primary benefits of IOV is the reduction in Total Cost of Ownership (TCO). IOV enables consolidation of I/O resources, which both reduces equipment and implementation costs, while allowing for better centralized and more efficient network management.
As an example, in today's architecture a chassis housing 10 blades would require a single 1Gigabit Ethernet adapter per-blade (see Figure 1). Using IOV technology it becomes possible to replace multiple gigabit adapters with one 10 Gigabit Ethernet adapter, which would be shared simultaneously by all the blades in the chassis. The 10 Gbps adapter is seen as 10 independent, "virtual" interfaces, with bandwidth that can be allocated dynamically. In addition to reducing the network complexity (consequently reducing the cost of managing it), IOV makes the system more efficient, as it allows sharing the unused bandwidth amongst the blades, instead of it being left idle. In the old, fixed model, each blade with a 1Gigabit adapter is physically limited to no more than 1Gbps, whereas blades sharing the 10 Gigabit bandwidth can dynamically provision and share the entire pipe.
IOV provides the same number of paths as an adapter-per-system implementation but does so without fixing the per-blade or per-image bandwidth. Additionally, if the requirements of an application change, it is possible to dynamically scale bandwidth allocation. For example, 3 Gbps could be made available during busy times to a blade that is normally allocated 1Gbps. Furthermore, this allocation can be groomed to provide guaranteed bandwidth under all operating conditions to accommodate service level agreements at the flow level.
|
