Four major network IO virtualization models

This article provides a brief overview of the 4 major virtualization IO models used by VMware, Redhat, Citrix, and Microsoft major virtualization vendors (emulation, para-virtualization, pass-through, SR-Iov). Mainly for everyone to go through, the information is relatively large, organizational layout is limited, crossing will look at the point.

Network I/O is not only the most likely bottleneck for physical servers, but also the biggest mishap of virtualization technology today. With the support of hardware virtualization for network I/O, the virtualized network I/O model is evolving, and the I/O performance of virtualization is constantly improving. Today we share the 4 main network I/O models used by VMware, Redhat, Citrix, and Microsoft mainstream virtualization solutions, including model principles, model benefits, model weaknesses, and model support for virtualized migrations.

1, Emulation

principle: Simulation (emulation) is a technology that simulates hardware entirely through a software program. This scenario is used by early virtualization to virtualize network devices. Common simulation software is QEMU, VMware WorkStation, VirtualBox. The emulation network model diagram is as follows:

Virtual network card products from different virtualization vendors are not the same

The VMware emulation type NICs are: E1000 (emulated Intel 82545M Gigabit Network adapter), flexible, vlance (emulation AMC 79c970 PCnet32 10M NIC), LANCE ( There are 3 versions of Vmxnet, namely Vmxnet, VMXNET2, VMXNET3, and no definitive information on VMware has been found to classify the three editions. Individual vmxnet are defined as emulation, VMXNET2, and VMXNET3 as para-virtualization types. VMXNET3 also supports some SR-Iov features).

The Redhat KVM and Citrix Xen type NICs are implemented with QEMU, and the emulation NICs available on KVM and Xen are: RTL8139 (emulation Realtek Link 8139 100M network connection), E1000 (emulation Intel 82545M Gigabit Network adapter).

Microsoft Hyper-V emulation type NICs are: Intel/dec 21140 100M NIC

Advantages: Software simulation does not require hardware support, simulation by CPU calculation, with host physical network card isolation, no platform requirements. Virtual machine operating systems do not need to be modified, the simulation is common network cards (such as: Intel E1000, RTL8139, etc.), the mainstream operating system has already brought these drivers, so by default the virtual machine does not need to install the driver.

Disadvantages: CPU resource consumption is high, especially when the number of virtual machines is many. Network card performance in general, because it is a software simulation, can only simulate common, more simple functions of the network card.

Virtual Machine Migration support: stripped of hardware requirements, using this type of migration is strong. Because both Xen and KVM use QEMU emulation, this type of virtual machine is not much more difficult to mix between Xen and KVM.

2, Para-virtualization

principle: Para-virtualization, also known as para-virtualization, was first used by Citrix's Xen. In the semi-virtualized model, the unified physical hardware resources are managed by hypervisor, and the hypervisor provides the resource invocation interface. The virtual sub-machine communicates with the hypervisor through a specific calling interface, and then the full I/O resource controls the operation. The Para-virtualization model diagram is as follows:

Para-virtualization, also known as semi-virtualized, was first proposed by Xen, and Xen itself started out as virtualization. Para-virtualization model, the virtual Machine network card driver can only be developed by hypervisor vendors, Redhat, VMware, Citrix, Microsoft's major virtual vendors have their own para-virtualization drivers. For example, Redhat KVM called Virtio,vmware VMXNET2, vmxnet 3,citrix Xen called Xen-pv,mircrosoft temporarily not found (welcome friends to add).

Advantages: Personally think is an improved version of the emulation model, but because of the communication between the sub-machine and hypervisor, performance than emulation.

disadvantage: need to modify the virtual machine operating system kernel, add different hypervisor vendors Network driver. For example, Linux (Redhat and Novell) added Mircosoft Para-virtualizaiton network drivers to the distribution, and Microsoft also added support for KVM Virtio and XEN-PV drivers in its own distribution.

Virtual Machine Migration support: Although the para-virtualization scenarios for different virtualization vendors are not the same, the mobility of this type of virtual sub-machine is easier to implement because mainstream operating systems provide support for these scenarios at the same time.

3, pass-through

principle: Hypervisor assigns a PCI device (which can be a network card, USB, CD-ROM) directly to the specified virtual sub-machine for individual access. For security and stability reasons, the use of pass-through is typically combined with Intel vt-d (AMD has similar technology) to ensure that memory access between virtual machines does not conflict through IOMMU. This technique is called Vmdirectpath I/O on VMware and no specific terminology is found in other scenarios.

Advantages: good performance. A separate PCI device is assigned to a virtual sub-machine, and the virtual machine communicates directly with the physical device.

disadvantage: The device can only be used by a virtual sub-machine, configuration is also more complex, first need to hypervisor the specified device through the PCI ID mode assigned to the specified virtual sub-machine, and then the virtual sub-machine to identify the device and then install the driver to use.

Mobility : Migration aspects to be studied, interested friends can be supplemented by perfect.

4. SR-Iov

background:The pass-through model allows virtual machines to use physical devices directly, which makes the network performance of the virtual sub-machine optimal. SR-Iov is mainly used to solve the problem that pass-through can only be accessed by a virtual sub-machine. SR-Iov Standard by PCI-SIG, this standard implementation needs CPU, chipset and PCI device (mainly network card and other I/O resources) to implement the hardware level, SR-Iov is considered to solve the last kilometer of virtualization problem. The model diagram is as follows:

principle: SR-Iov requires network card hardware support, the network card that supports SR-Iov function can be registered as multiple network cards in Hypervior (each network card is independent of interrupt ID, Transceiver queue, QoS management mechanism). Each device can be assigned to a virtual sub-machine by means of pass-through.

Product: Common is the 10GB network adapter based on the Intel 82599 and 82598 chipset. VMware, Redhat, Citrix, and Microsoft have already or are adding support for this feature in Hypervisor. Different vendor virtualization programs are not the same, they are interested in Google can search for more information.

Pros: Needless to say, X86 virtualization is the latest IO virtualization model; The virtual machine not only has good performance, but also combines the hardware function, and proposes a new scheme for the virtual machine IO management.

Cons: Pending

Mobility: SR-Iov requires two levels of hardware and software support, and virtual machines can theoretically not be problematic when migrating between the same NIC hosts. It depends on the implementation of virtualization manufacturers.

Link

Four major network IO virtualization models