Why keelOS
One machine. Nothing has to stop.
Home servers and small offices run everything on one box: the NAS, the router, a Windows desktop, a few Linux servers. keelOS was built for exactly that box, and the same image runs the rows of hosts in a data centre, managed from one place. Each point below was measured on real hardware; the linked articles (in Chinese) have the details.
What happens to the VMs when the host restarts
| System | Ordinary VMs | NAS VM with its disk controller passed through |
|---|---|---|
| Unraid 7 | Shut down, or hibernated by the guest OS (needs the guest agent) | None: the host itself is the NAS |
| TrueNAS SCALE | Shut down and started again | None: the host itself is the NAS |
| Proxmox VE 9 | Shut down and started again; suspend to disk only by hand | Shut down and started again |
| Hyper-V (Windows Server, DDA) | Saved to disk, restored after the restart | Shut down (Save is not allowed with passed-through devices) |
| ESXi 8 | Can be suspended to disk, resumed after the restart | Shut down and started again (cannot be suspended) |
| keelOS | Paused for about 25 s in memory; TCP connections stay up | The same: paused for about 25 s, the controller keeps running |
Suspend and save write the VM's memory to disk and read it back after the restart; this takes longer the more memory the VM has, and none of the platforms above allows it for a VM with passed-through devices (sources below). Data centres move VMs to a second machine before a host restart. At home there is no second machine.
Sources: Unraid 7.0 release notes · Proxmox forum: hibernate on host reboot · Proxmox forum: snapshots with passthrough · Microsoft: Discrete Device Assignment · VMware: DirectPath I/O · Broadcom KB 428701
Host updates do not stop the VMs
A small hypervisor holds guest memory and vCPU state and stays in memory while the FreeBSD host is upgraded and restarted. Guests pause for 24–26 s with TCP connections intact. 100 host restarts in a row without a failure; 17 VMs resumed together after one restart.
Passthrough that survives the restart
The NAS VM gets the whole disk controller or NVMe drive and sees real disks: SMART, write cache, real errors. The hypervisor owns the IOMMU, so DMA keeps working while the host restarts: 14 restarts under full write load, no timeouts, ZFS scrub with zero errors. SR-IOV VFs for the network.
Windows 11 just installs
A virtual TPM 2.0 and Secure Boot pass the installer's check without registry tricks. BitLocker works, and the TPM state survives host restarts.
Memory that is not wasted
On our development host, 49% of 44 GB of guest memory was zero pages after a day. keelOS reclaims them and lends them to the host as disk cache, and gives them back the moment a guest writes. No overcommit: the memory can always be returned.
Clocks that stay right
A pause is invisible to the guest's clock: after dozens of host restarts FreeBSD and Linux guests are within ±0.4 s. Windows Server 2003 no longer runs 3.6% slow.
From Windows XP to Windows 11
keelOS added IDE, PCnet and UHCI devices and SeaBIOS to bhyve, so XP and Server 2003 install with the drivers they ship with. Nothing to add to the guest.
Your router in a VM
OpenWrt, ImmortalWrt, iStoreOS, OPNsense, pfSense CE, VyOS and RouterOS CHR all run, most with an SR-IOV VF of the network card. keelOS stops the classic mistake of bridging the office LAN into a loop.
ZFS underneath
Each VM is one ZFS dataset tree: configuration, firmware variables and disks snapshot together. Cold migration to another keelOS host with zfs send works, driven from the management VM; migration with memory is planned.
Light by design
One read-only system image, one management program with no database, a web interface, command line and HTTPS API. RDP straight to any VM's screen, with nothing installed in the guest.