Home > Articles > Cisco Network Technology > General Networking > Cisco Data Center Virtualization Server Architectures

Cisco Data Center Virtualization Server Architectures

Chapter Description

This chapter examines processor, memory, and I/O subsystems with particular reference to servers built according to the IA-32, often generically called x86 architecture. In particular, it describes the most recent generation of Intel processors compatible with the IA-32 architecture; i.e., the Intel microarchitecture.

Chipset Virtualization Support

In addition to the virtualization support provided inside Nehalem, other improvements have been implemented at the chipset/motherboard level to better support virtualization. These improvements are important to increase the I/O performance in the presence of a hypervisor (in Intel parlance, the hypervisor is referred to as VMM: Virtual Machine Monitor).

Intel® VT-d for Direct I/O

Servers use an Input/Output Memory Management Unit (IOMMU) to connect a DMA-capable I/O bus (e.g., PCIe) to the main memory. Like a traditional MMU (Memory Management Unit), which translates CPU-visible virtual addresses to physical addresses, the IOMMU takes care of mapping device-visible virtual addresses to physical addresses. These units also provide memory protection from misbehaving devices.

A general requirement for I/O virtualization is the ability to isolate and restrict device accesses to the resources owned by the partition managing the device.

In 2008, Intel published a specification for IOMMU technology: Virtualization Technology for Directed I/O, abbreviated VT-d.

Intel® VT for Directed I/O provides VMM software with the following capabilities:

  • I/O device assignment: For flexibly assigning I/O devices to VMs and extending the protection and isolation properties of VMs for I/O operations.
  • DMA remapping: For supporting independent address translations for Direct Memory Accesses (DMA) from devices.
  • Interrupt remapping: For supporting isolation and routing of interrupts from devices and external interrupt controllers to appropriate VMs.
  • Reliability: For recording and reporting to system software DMA and interrupt errors that may otherwise corrupt memory or impact VM isolation.

Intel® VT-c for Connectivity

Intel® Virtualization Technology for Connectivity (Intel® VT-c) is a collection of I/O virtualization technologies that enables lower CPU utilization, reduced system latency, and improved networking and I/O throughput.

Intel® VT-c consists of platform-level technologies and initiatives that work together to deliver next-generation virtualized I/O:

  • Virtual Machine Device Queues (VMDq) dramatically improves traffic management within the server, helping to enable better I/O performance from large data flows while decreasing the processing burden on the software-based Virtual Machine Monitor (VMM).
  • Virtual Machine Direct Connect (VMDc) provides near native-performance by providing dedicated I/O to virtual machines, bypassing the software virtual switch in the hypervisor completely. It also improves data isolation among virtual machines, and provides flexibility and mobility by facilitating live virtual machine migration.


In virtual environments, the hypervisor manages network I/O activities for all the VMs (Virtual Machines). With the constant increase in the number of VMs, the I/O load increases and the hypervisor requires more CPU cycles to sort data packets in network interface queues and route them to the correct VM, reducing CPU capacity available for applications.

Intel® Virtual Machine Device Queues (VMDq) reduces the burden on the hypervisor while improving network I/O by adding hardware support in the chipset. In particular, multiple network interface queues and sorting intelligence are added to the silicon, as shown in Figure 2-45.

As data packets arrive at the network adapter, a Layer 2 classifier/sorter in the network controller sorts and determines which VM each packet is destined for based on MAC addresses and VLAN tags. It then places the packet in a receive queue assigned to that VM. The hypervisor's layer 2 software switches merely routes the packets to the respective VM instead of performing the heavy lifting work of sorting data.

As packets are transmitted from the virtual machines toward the adapters, the hypervisor layer places the transmit data packets in their respective queues. To prevent head-of-line blocking and ensure each queue is fairly serviced, the network controller transmits queued packets to the wire in a round-robin fashion, thereby guaranteeing some measure of Quality of Service (QoS) to the VMs.


To take full advantage of VMDq, the VMMs needs to be modified to support one queue per Virtual Machine. For example, VMware® has introduced in its hypervisor a feature called NetQueue that takes advantage of the frame, sorting capability of VMDq. The combination of NetQueue and VMDq offloads the work that ESX has to do to route packets to virtual machines; therefore, it frees up CPU and reduces latency (see Figure 2-46).

Figure 2-46

Figure 2-46 VMM NetQueue


VMQ is Microsoft's Hyper-V® queuing technology that makes use of the VMDq capabilities of the Intel Ethernet controller to enable data packets to be delivered to the VM with minimal handling in software.


Virtual Machine Direct Connect (VMDc) enables direct networking I/O assignment to individual virtual machines (VMs). This capability improves overall networking performance and data isolation among VMs and enables live VM migration.

VMDc complies with the Single Root I/O Virtualization (SR-IOV) standard; see "SR-IOV" in Chapter 3, page 83.

The latest Intel® Ethernet server controllers support SR-IOV to virtualize the physical I/O port into multiple virtual I/O ports called Virtual Functions (VFs).

Dividing physical devices into multiple VFs allows physical I/O devices to deliver near-native I/O performance for VMs. This capability can increase the number of VMs supported per physical host, driving up server consolidation.


When VMDq and VMDc are combined with VMware® VMDirectPath, the ESX/vSphere hypervisor is bypassed in a way similar to a "kernel bypass", the software switch inside the hypervisor is not used, and data is directly communicated from the adapter to the vNICs (virtual NICs) and vice versa. See "VN-Link" in Chapter 3, page 90, and in particular Figure 3-20.