# Faster, smolBSD! Boot, Boot!

### Paper Summary

Virtualisation technology has gone a long way. Virtualisation began as early
as the 1960s with IBM CP/CMS, and was already distributed as Open Source back
then. On modern commodity hardware however, the technology really hit us in the
mid-2000s with Intel's VT-x and AMD-V. Many different implementations compete on
these platforms alone, and thankfully this is also true for our beloved BSD
systems.

While the technology kept improving over time, both on the hardware and on
the software side, raw performance has long been an issue and important criteria
for comparing solutions. Different approaches have been released, like the
emulation of existing components, the development of dedicated drivers with
paravirtualization, or hybrid systems. All in all, these different approaches
have been fairly complex to develop or operate.

But then, "containerization" started gaining real prominence in 2014. What we
already knew as jails not only exploded in popularity, but also changed the way
software was written, scaled, and distributed. Entire infrastructures can be
deployed and adjusted within seconds, with new services starting in milliseconds
in complete isolation from their environment.

This last property has notably been made possible with the introduction of
lightweight virtual machines, or "microVMs", as implemented by Firecracker or in
QEMU with the "microvm" virtual platform. Far from the complexity of prior
approaches but building on its innovations, it offers a minimalist machine type
without PCI nor ACPI support, allowing more barriers to be breached in
performance.

Indeed, in the 2020s, we have witnessed a drag race for boot time between
FreeBSD and NetBSD! This paper focuses on the progress made on the side of
NetBSD, which saw the introduction of a number of different drivers, fixes,
improvements and new features, each responsible for their respective gains.
Overall, this has allowed reducing boot time from over 300 to just under 10
milliseconds!

In no particular order, this included: boot-time benchmarking with tslog(4),
generic Xen PVH support, removing calibration loops and DELAY(), PCI-less
operation with boot-time MMIO parameters, system console over VirtIO, or
reducing the kernel's size with the MICROVM kernel configuration. We are happy
to report that all of these improvements have made their way into the coming
NetBSD release, in its version 11!

The latest addition to this pile is the possibility to start the NetBSD kernel
with an in-memory root filesystem, when booting in the generic Xen PVH mode
introduced here. While unfortunately, this will likely not be adopted in time
for the coming release, it finally allowed us to break the 10 milliseconds
barrier on the reference hardware.

Last but not least, and in line with the traditions from NetBSD, the whole
solution is portable and can easily be deployed on other systems as well, like
Linux or macOS. The smolBSD project, hosted on GitHub, offers a web interface
and convenience scripts and tools, managing such NetBSD Virtual Machines from a
web browser.

This presentation will summarize the different improvements relevant for the
faster boot times, especially since Emile Heitor's presentation at BSDCan 2024.
It will also explain the patch bringing support for ramdisks in the generic Xen
PVH mode. Finally, the smolBSD project will be presented, with demos exploring
different possibilities of the system.