VMM Backends

Four backends behind a common trait. Same code, same protocol, same result format. Pick the right trade-off per workload.

Default

macOS + Linux

libkrun

Embedded VMM via FFI. No separate process. Transparent socket networking (TSI). The fastest path from code to VM.

~337ms end-to-end

Linux only

Firecracker

AWS Lambda's VMM. Minimal device model, battle-tested jailer. ext4 block device rootfs. The production hardening choice.

~344ms end-to-end

Linux only

Cloud Hypervisor

Modern rust-vmm VMM. virtio-fs rootfs via virtiofsd. Snapshot/restore, warm migration, GPU passthrough via VFIO.

~307ms end-to-end

Linux only

QEMU

Full device emulation. Broadest hardware support. GPU passthrough with OVMF firmware boot. The escape hatch for complex workloads.

~555ms end-to-end

Fastest boot Most features

Feature Matrix

libkrun

Firecracker

Cloud Hypervisor

QEMU

Platform

macOS + Linux

Linux only

Boot time (median)

~337ms

~344ms

~307ms

~555ms

Rootfs

virtio-fs (built-in)

ext4 block device

virtio-fs (virtiofsd)

Networking

TSI (proxied sockets)

TAP + NAT + egress filter

Shared directories

Yes

GPU passthrough

Yes (VFIO)

Yes (VFIO + OVMF)

Snapshot/restore

Yes

Snapshot only

Warm migration

Yes

No (no restore)

Persistent VMs

Yes

Port forwarding

TSI port mapping

iptables DNAT

Host sandboxing

hotcell-jailer (14 steps)

Firecracker jailer

CH seccomp + hotcell-ch-worker

KVM isolation only

Guest kernel

Built into libkrunfw

Separate vmlinux

Best for

Dev, macOS, low latency

Production, multi-tenant

GPU compute, migration

Complex hardware, CUDA

Boot times measured on bare metal (AMD Ryzen 9 7950X3D, KVM). Full lifecycle: OCI rootfs assembly, VM boot, command execution, result collection, teardown.

Default

libkrun

An embedded VMM that runs inside the hotcell worker process via FFI. No separate binary to manage, no REST API, no socket coordination. The VM starts when you call krun_start_enter() and the worker process becomes the guest.

Uses Hypervisor.framework on macOS and KVM on Linux. The only backend that runs on macOS, making it the default for development.

+Cross-platform (macOS + Linux)

+No external binary dependencies

+TSI networking (transparent socket proxying)

+14-step hotcell-jailer sandbox on Linux

−No snapshot/restore or migration

−No GPU passthrough

Linux only

Firecracker

The VMM that powers AWS Lambda and Fargate. A separate binary configured via REST API over a Unix socket. Minimal device model with a tiny attack surface. Uses ext4 block device images built from OCI rootfs layers.

Full snapshot/restore support for warm migration between hosts. Pause a running VM, serialize its memory and CPU state, transfer to another host, restore.

+Battle-tested in AWS Lambda (millions of VMs)

+Snapshot/restore and warm migration

+Firecracker's own jailer for additional sandboxing

+TAP networking with egress filtering + rate limits

−No shared directories (ext4 rootfs only)

−No GPU passthrough

Linux only

Cloud Hypervisor

A modern, Rust-based VMM built on the rust-vmm crate ecosystem. Uses virtio-fs via external virtiofsd processes for rootfs and shared directory access. Configured via REST API over a Unix socket.

The most feature-rich backend: snapshot/restore, warm migration, live pre-copy migration plumbing, GPU passthrough via VFIO with firmware boot mode (CLOUDHV.fd).

+GPU passthrough (VFIO) with CUDA compute verified

+Snapshot/restore and warm migration

+Shared directories via virtio-fs

+Live pre-copy migration plumbing

−Requires virtiofsd (external daemon per mount)

−Linux only, larger kernel needed for virtio-net

Linux only

QEMU

The most mature VMM in the ecosystem. Full device emulation, broadest hardware support. Configured via command-line arguments, managed via QMP (JSON over Unix socket).

QEMU is hotcell's escape hatch for workloads that need OVMF firmware boot (required for NVIDIA GPU ROM initialization), complex PCI topologies, or specific device emulation. Uses virtio-fs via virtiofsd for rootfs access.

+GPU passthrough with OVMF firmware (NVIDIA CUDA)

+Broadest hardware and device emulation support

+Most mature migration support (pre-copy, post-copy, RDMA)

+Shared directories via virtio-fs

−Slowest boot (~555ms, larger attack surface)

−Restore path not yet implemented in hotcell

Choosing a Backend

Developing on macOS?

Use libkrun (the default). It's the only backend that supports macOS via Hypervisor.framework.

Multi-tenant production?

Use Firecracker. Battle-tested at AWS scale, minimal attack surface, snapshot/restore for migration.

GPU compute (CUDA)?

Use Cloud Hypervisor or QEMU. Both support VFIO passthrough. QEMU adds OVMF firmware for NVIDIA driver initialization.

Need to share host directories?

Use libkrun, Cloud Hypervisor, or QEMU. All three support virtio-fs shared mounts. Firecracker uses ext4 block devices only.