Comprehensive Overview of Server Virtualization Technologies

Server virtualization traces its origins to IBM's mainframe z/VM, which could run hundreds of virtual machines, and later to PowerKVM and PowerVM on Power systems, introducing technologies such as vSCSI and NPIV.

The article explains that despite the rise of container technology, virtualization remains indispensable for many scenarios, especially for newcomers to cloud computing.

CPU Virtualization

Virtualization can be classified as full, para (half), or hardware‑assisted. Full virtualization requires binary translation of privileged instructions, para‑virtualization uses hypercalls, and hardware‑assisted virtualization leverages new CPU modes (root/non‑root) to let guest OS run at Ring 0.

Images illustrate each approach.

Virtualization Software Architecture

Four main categories are described:

Hosted (paravirtual) virtualization : a hypervisor runs as an application on a host OS.

Bare‑metal (type‑1) virtualization : the hypervisor runs directly on hardware, handling privileged instructions and scheduling.

OS‑level virtualization : the host OS itself partitions resources, requiring all guests to share the same OS kernel (e.g., Docker, OpenVZ).

Hybrid virtualization : combines a host OS with a kernel‑level driver that acts as a virtual hardware manager, improving performance.

Comparisons indicate that bare‑metal and hybrid models, together with hardware‑assisted features, are the future direction.

Memory Virtualization

Virtual memory managers must present each guest with a consistent view of memory while mapping guest physical addresses to host physical addresses. Techniques include full memory virtualization with shadow page tables, para‑virtual memory where guests register page tables with the VMM, and hardware‑assisted memory virtualization using Extended Page Tables (EPT).

Illustrative diagrams are included.

I/O Virtualization

To avoid I/O bottlenecks, virtualization extends to network and storage devices, providing isolation and reducing overhead. The article details full I/O virtualization (device emulation), para‑virtual I/O (front‑end/back‑end queues), and hardware‑assisted I/O (Intel VT‑d/VT‑c, AMD IOMMU, SR‑IOV). It also describes different NIC models such as ordinary NICs, VMDq, and SR‑IOV.

Intel Hardware Support for Virtualization

VT‑x introduces root and non‑root operation modes, allowing certain privileged instructions to execute directly in the guest. VT‑d enables direct I/O access with DMA remapping and I/O page tables. VT‑c adds connectivity features like VMDq and VMDc, while Trusted Execution Technology (TXT) enhances security.

GPU Virtualization

Three GPU virtualization models are covered: GPU passthrough (one GPU per VM), GPU sharing via a GPU server, and full GPU virtualization where a physical GPU is split into multiple virtual GPUs (vGPUs) that can be concurrently used by several VMs, sharing 3D and video decoding engines.

Images illustrate the architecture of GPU virtualization.

The article concludes with a reminder to follow the “ICT_Architect” public account for more content.