leetbit
Kernel Workbenchactive

The primary environment for filesystem and kernel work — a single Linux workstation hosting the VM fleet, tracing, and benchmark tooling that the VFS articles and atlas-fs are built and tested on.

Since
2024-11
Location
Home lab · Rack A
Components
7
Last verified
Jul 15, 2026
Host OS
Debian 13 (trixie)
Kernel
6.9.x (custom .config)
HardwareOne workstation; the VM fleet runs on it.2
workbench-01Workstationactive
Primary development host — builds kernels and runs the hypervisor.
CPUAMD Ryzen 9 7950X (16C / 32T)
Memory128 GB DDR5-5200 ECC
Boot / build2× 2 TB NVMe (mdraid RAID1)
Bulk8 TB HDD (results, archives)
GPUnone — headless
Since2024-11
nas-01Storage nodeactive
Off-box archive for benchmark output and VM images.
Capacity8 TB usable (ZFS mirror)
Link2.5 GbE
Since2025-02
Software3
KVM / libvirtHypervisoractive
Runs the VM fleet; virtio devices; snapshot-per-experiment.
Guestsvfs-lab · fuse-lab · bench
TracingObservabilityactive
Kernel-side measurement inside the guests.
Toolsftrace · bpftrace · perf
Benchmark harnessToolingactive
fio plus a custom lookup/latency harness; results are versioned.
Network1
lab-vlanSegment
Isolated VLAN for the workbench and its storage; no inbound.
VLAN20
Speed2.5 GbE
Storage1
build-arrayVolume
Fast scratch for kernel trees and VM disks.
LayoutRAID1 · 2× NVMe
Filesystemext4 (build) · xfs (images)

The Kernel Workbench is one physical machine that behaves like a small lab. It builds custom kernels, hosts a fleet of purpose-built virtual machines, and captures traces and benchmarks — all on hardware that can be reset to a known state between runs. It exists so that the claims made in the VFS articles and in atlas-fs are things that were observed here, not recalled.

Logical layout#

Everything hangs off a single host. The workstation builds kernels and runs the hypervisor; each experiment gets its own VM from a snapshot, so a run never inherits state from the one before it.

Schematic: workbench-01 hosts a KVM hypervisor running vfs-lab, fuse-lab, and bench VMs, with results written to nas-01.
Logical layout — one host, a snapshot-per-experiment VM fleet, results archived off-box.

How it is used#

Kernel changes are built on workbench-01 and booted inside vfs-lab, where ftrace and bpftrace are already wired up. FUSE work runs in fuse-lab, whose only job is to be a clean mount target for atlas-fs. Anything that produces numbers runs in bench, which is pinned to an isolated cpuset with no other tenants during a measurement.

Reproducibility
Each VM boots from a named snapshot and is discarded afterward. The kernel .config, the VM definitions, and the harness scripts are versioned; a run records the commit it was taken at, so a result can be traced back to an exact environment.

The machine and its fleet#

The inventory above is the source of truth for what is actually installed; this page is refreshed whenever the environment changes.

Relations

where else this lives in the system
status activeupdated Jul 18, 2026verified Jul 15, 2026