Running Kubernetes on a Synology NAS: A VM Beats Native Every Time
February 26th, 2026 We tried running k0s natively on Synology DSM 7, hit every kernel limitation imaginable, then ran it in a VM instead. Here's what we learned.
We run a bare-metal Kubernetes cluster with k0s across ~20 nodes. When we added a Synology DS1823xs+ NAS (8-core AMD Ryzen, 64GB RAM, 25TB storage), it seemed like a natural fit as a worker node. We spent weeks trying to run k0s natively on DSM 7.3.2. It was a mistake.
This post covers what went wrong, why a VM is the right answer, and exactly how to set it up.
The Problem with Native DSM
DSM 7.x ships kernel 4.4.302+. That's a 2016-era kernel running in 2026. For NAS duties it's fine. For Kubernetes, it's missing critical features:
No seccomp filter (CONFIG_SECCOMP_FILTER). Containerd expects to apply seccomp profiles to containers. Without kernel support, every container creation fails. We had to fork containerd and patch it to silently skip seccomp — a maintenance burden we never wanted.
No CFS bandwidth (CONFIG_CFS_BANDWIDTH). The cpu.cfs_period_us cgroup file doesn't exist. Runc writes it unconditionally when a CPU limit is set, then crashes. Another containerd patch.
No pids cgroup (CONFIG_CGROUP_PIDS). Kubelet can't enforce per-pod pid limits. Workaround: enforceNodeAllocatable: [] in the worker profile.
No ip_set_hash_net module. Calico's felix creates hash:net ipsets for IPAM pool matching. Without this kernel module, felix enters a restart loop. Each restart briefly drops SSH connections (felix inserts a default-DROP rule before adding its ACCEPT rules). The node shows as Ready (BGP works independently) but calico-node stays 0/1 forever.
No statx syscall. Added in kernel 4.11, missing in 4.4. Modern Rust binaries (including virtiofsd) call statx unconditionally and fail with ENOSYS.
Root mounted as rprivate. DSM mounts / with MS_PRIVATE. Kubernetes DaemonSets need shared mount propagation. Fix: mount --make-rshared / before every k0s start, in a systemd dropin that survives restarts.
And the worst part: the host crashed under load. Something about running Kubernetes workloads on DSM's kernel was fundamentally unstable. We never fully diagnosed it because we found a better solution.
The Solution: An Ubuntu VM
Synology VMM (Virtual Machine Manager) runs KVM/QEMU with VirtIO paravirtualization. Performance is near-native. We created an Ubuntu 24.04 VM on the NAS and ran k0s inside it.
Every DSM kernel limitation disappeared. calico-node runs 1/1. Containerd uses stock overlayfs. No patches, no forks, no workarounds. The node is just another Ubuntu worker.
Setup
Create the VM in VMM:
- Ubuntu 24.04, 6 vCPUs, 32GB RAM, 200GB VirtIO disk
- UEFI firmware + VGA display (BIOS mode hangs during Ubuntu boot on VMM)
- Default VM Network (bridged to the host's physical NIC)
- Enable autostart
Install Ubuntu with minimal server, OpenSSH server, ext4 filesystem. Standard stuff.
Mount the NAS storage via NFS. The VM needs access to the Synology's 25TB /volume1. Add an NFS export on the DSM host restricted to the VM's IP:
# /etc/exports on DSM
/volume1 10.2.0.25(rw,async,no_wdelay,no_root_squash,sec=sys,anonuid=1025,anongid=100)
Then in the VM's fstab:
10.2.0.73:/volume1 /mnt/volume1 nfs rw,hard,intr,nfsvers=3 0 0
NFS between a VM and its host goes through the virtual bridge. Our measured RTT is ~0.4ms. Not as fast as local disk, but plenty for Longhorn replicas and local-path-provisioner volumes.
Provision with Ansible. We added the VM to our existing Ansible inventory with NFS-specific host vars:
appmana_017_vm:
ansible_host: appmana-017-vm.i.appmana.com
nfs_server: 10.2.0.73
nfs_export: /volume1
nfs_mountpoint: /mnt/volume1
Our worker playbook handles passwordless sudo, SSH keys, sysctl tuning, apt packages (including nfs-common), HWE kernel, swap, and the NFS mount. One command provisions the entire node.
Join the cluster with k0sctl. The VM entry in k0sctl.yaml is three lines:
- ssh:
address: 10.2.0.25
user: administrator
port: 22
role: worker
No dataDir, no workerProfile, no installFlags, no hooks. Just a standard worker. k0sctl apply took 50 seconds.
Configure storage. After the node joins:
- Patch the Longhorn node to use
/mnt/volume1/.../longhorninstead of/var/lib/longhorn/ - Update local-path-provisioner's
nodePathMapwith the NFS-backed path
The VM now contributes 17.8TB of Longhorn-managed storage to the cluster.
What About virtiofs?
We tried. virtiofs provides near-native file I/O for sharing host directories into VMs, which would have been ideal. But on DSM 7.x:
- The Rust virtiofsd requires
statx(kernel 4.11+). DSM has 4.4. Dead end. - The C virtiofsd from QEMU 6.2 works on kernel 4.4, but virtiofs requires QEMU shared memory (
memory-backend-memfdwithshare=on), which conflicts with VMM's memory configuration. - Libvirt hooks on Synology's libvirt version strip
qemu:commandlinenamespace blocks, so you can't inject extra QEMU args through hooks. - A QEMU wrapper script can rewrite the args, but it's fragile and breaks VMM's capability probing.
After multiple attempts, we gave up on virtiofs and used NFS. It's boring and it works.
Synology VMM API
A nice surprise: VMM exposes a full REST API. You can power on, shut down, and manage VMs programmatically:
# Authenticate
SID=$(curl -s -k "https://localhost:5001/webapi/entry.cgi?\
api=SYNO.API.Auth&version=6&method=login&account=admin&passwd=<pass>" \
| python3 -c "import sys,json; print(json.load(sys.stdin)['data']['sid'])")
# Power on a VM
curl -s -k "https://localhost:5001/webapi/entry.cgi?\
api=SYNO.Virtualization.API.Guest.Action&version=1&method=poweron\
&guest_id=<vm-uuid>&_sid=$SID"
This is useful for automation — no need to open the DSM web UI.
Results
| Metric | Native DSM | Ubuntu VM |
|---|---|---|
| calico-node | 0/1 (restart loop) | 1/1 |
| containerd | Patched fork | Stock |
| Kernel workarounds | 6+ | 0 |
| k0sctl.yaml lines | ~190 | 6 |
| Host stability | Crashed under load | Stable |
| Setup time | Weeks | Hours |
| Maintenance burden | High (patches, hooks, dropins) | None (standard Ubuntu) |
Lessons
- Don't fight the kernel. If the kernel is too old, run a newer one. A VM is cheaper than maintaining kernel workaround patches.
- VMM is good. Synology's KVM implementation is solid. VirtIO performance is near-native. The REST API is a bonus.
- NFS is fine. For a VM accessing storage on its own host, NFS over a virtual bridge is simple and fast enough. Don't over-engineer with virtiofs when NFS works.
- Keep it boring. The VM approach uses zero Synology-specific code. It's just Ubuntu. When the next DSM update ships, nothing breaks.
Deploy in minutes
Get started with your agent
Paste this into your coding agent. With these instructions, it can configure your Unity project and get you streaming in minutes.
For this session, read appmana.com/llms.txt for instructions on how to add your Unity project to be built and streamed automatically on AppMana.