CS Notes

Recent writing

See 657 more →

Virtualisation

337 words, 2 min read
Last updated on
🌟 Edit This Page!   🗓️ History

Xinyang YU
sai

Abstract

  • Virtualisation is a software emulation of hardware. It allows you to have multiple OS running inside a host OS

Type 1 hypervisor

Type 2 hypervisor

Firecracker

  • Firecracker runs microVMs, which are extremely lightweight virtual machines that run on top of KVM, a type 1 hypervisor. Each microVM has its own guest kernel, providing hardware-level isolation that’s much stronger than container (OS-level) isolation
  • Because they boot extremely fast and use very little memory, they’re used by AWS Lambda and AWS Fargate to safely run multi-tenant workloads at massive scale

Virtualisation limitation

Unlike Docker or gVisor, it doesn’t share the host kernel or oversubscribe CPU and memory, each microVM gets fixed, dedicated resources for predictable performance and strong security.

This minimalism also means it lacks PCIe virtualisation, making GPU passthrough (including NVIDIA MIG slicing) difficult to implement. While gVisor offers software-based isolation using a user-space kernel (runsc), Firecracker enforces true hardware isolation via KVM, trading flexibility (like GPU and overcommit support) for tighter security and deterministic behavior.

gVisor

  • A sandboxing runtime that improves container security by running applications inside a user-space kernel called Sentry, which intercepts and emulates system callsi nstead of letting them reach the host kernel directly.
  • Its runtime, runsc, replaces Docker’s default runc to provide this isolation transparently. This design greatly reduces the attack surface and protects the host from compromised containers, though it comes with some performance overhead.
  • gVisor sits between Docker and full virtualization tools like Firecracker, offering stronger isolation than regular containers but less than hardware-based microVMs.

NVProxy

  • gVisor’s GPU bridge that safely forwards CUDA calls from the sandbox to the host’s NVIDIA driver, which can slightly slow GPU communication.
  • Despite this potential limitation, NVProxy still supports most CUDA-based AI workloads effectively, making gVisor suitable for secure AI use cases where slight speed trade-offs are acceptable.

Virtualisation Tools

Vagrant

  • A piece of software to manage VMs with ease
  1. Install VirtualBox
  2. Install Vagrant
brew tap hashicorp/tap
brew install hashicorp/tap/hashicorp-vagrant
vagrant plugin install vagrant-share

QEMU

  • Powerful open-source machine emulator and virtualizer
brew install qemu

Mentioned by

Graph View