Emerging Hardware Support For I/O Virtualization; Intel Vt-D And Amd Iommu - HP BL680c - ProLiant - G5 Introduction Manual

At this point in time, the PCI-SIG has finalized specifications for a 150-watt PCI Express slot in order
to provide a standardized solution that can be used by designers/manufacturers of add-in cards with
large power requirements, including high-end graphics engines in particular. Since the PCI Express
connector itself is not designed to deliver more than 75 watts of total power, the additional power
must be delivered using a specific external connection between the system board and the card itself.
HP is beginning to implement a limited number of 75-watt slots into some of the ProLiant server line.
As for the 150-watt slot design, which is being driven primarily by the needs of high performance
graphics engines used in workstations, HP will provide this in server designs that require it.

Emerging hardware support for I/O Virtualization

As use of Virtual Machine technology in server environments continues to grow and as Virtual
Machine Manager (VMM) software evolves, the hardware industry continues to look for ways to
enable VMM software, helping to make it more efficient, more reliable and more secure.
One of the core technical challenges that all VMMs have to solve is the virtualization of I/O such that
each individual virtual machine on the server can access the real (physical) I/O devices of the
underlying server hardware. Any one of three different methodologies – emulation, paravirtualization
or direct assignment – can be employed to accomplish this task. Each of these methods also has
benefits as well as drawbacks.
Emulation and paravirtualization are somewhat similar in that they both involve the creation of a
software layer, or device model, that emulates each I/O device within a virtual machine. Guest OSs
in each virtual machine then communicate with this device model and not directly with the physical
I/O device in the server. That task is left to the VMM, which has to both translate the guest OS's I/O
requests into actual I/O operations on the physical device as well as manage and schedule the
requests coming from the various virtual machines that have been installed.
Both emulation and paravirtualization do a good job of insulating the various virtual machines from
the server hardware as well as from each other. The price for this, of course, is significantly reduced
performance and potential I/O latency issues as a result of the extra layers of software emulation (the
device models), translation and arbitration that have been introduced into the system.
Direct assignment is designed to solve these issues by eliminating the need for a device emulation
layer in software. As its name implies, with this method each physical I/O device is directly assigned
to a given virtual machine, allowing the virtual machine (and typically only that virtual machine) to
perform I/O operations directly to the device. The result is significantly improved I/O performance in
VM environments. Because of its simpler architecture and improved performance characteristics,
direct assignment is becoming the direction in which VMM software is trending for the future.
Still, direct assignment is not without its challenges. I/O devices that use DMA can have problems
since the guest VM sees a virtualized subset of system memory while the devices themselves
communicate directly with the host physical memory addresses when transferring data. Today's
VMM's have to deal with this mismatch by employing software mechanisms to perform this translation
between the guest VM and the real physical addresses.
In order to address these types of issues regarding I/O virtualization, both Intel and AMD have
announced enabling technologies to better support I/O virtualization in VMM environments.

Intel VT-d and AMD IOMMU

Intel Virtualization Technology for Directed I/O (VT-d) and AMD I/O Virtualization Technology
(IOMMU) are new technologies that will be incorporated into processor chipsets in the near future.
Although they each have their own unique programmatic interfaces, both VT-d and IOMMU are
designed to accomplish the same task, to provide hardware-based functionality that VMMs and OSs
can utilize to provide better performing and more secure I/O virtualization for Virtual Machine
9