Multikernel Linux Tree Released: Runs Multiple Kernels on Bare Metal Without a Hypervisor

The premise is that one server can run a database on one kernel and a GPU training job on another, with each kernel isolated from the other's resources. A host kernel manages the machine's CPUs, memory, and PCI devices. When it starts a new instance, that instance receives its own dedicated slice of hardware and boots directly onto it using Linux's kexec mechanism, running natively rather than under a hypervisor. The host tracks all instances through device tree overlays under /sys/fs/multikernel/ and can shift hardware between running kernels without stopping them.

This approach differs from both virtual machines and containers. A VM places a hypervisor and a host kernel between the application and the hardware. A container skips the hypervisor, but all tenants on the machine share one kernel. The multikernel tree gives each workload a separate kernel with dedicated resources, though the current release supports only x86_64 hardware. Multikernel Technologies says the architecture-specific code is modular enough for other ports later.

The company sells several products based on the tree, including one for running AI agents in their own kernel with direct GPU access, and another that promises to patch a running kernel without taking the system down. The project has been under development since last year.