Reducing initramfs size and speed up the generation

[Benjamin Drung]

On Sun, 2023-07-09 at 15:29 -0700, Steve Langasek wrote:
> On Sun, Jul 09, 2023 at 04:28:42AM +0200, Heinrich Schuchardt wrote:
> > > > Benjamin Drung <bdrung at ubuntu.com> schrieb am Sa., 8. Juli 2023, 02:19:
> 
> > > > > Hi all,
> 
> > > > > a year ago we changed the default compression and level for the
> > > > > initramfs to zstd -1. This fixed the very slow creation times on
> > > > > development boards (see bug #1958148), but that leads to bigger
> > > > > initramfs sizes that triggered other bugs (like bug #1842320).
> > > > > Big initramfs sizes can also fill up small sized /boot partitions easily
> > > > > (grooming the 850 initramfs-tools bugs revealed several such reports).
> 
> > > > > Using xz -9 would give very good compression, but it takes very long
> > > > > (especially on slow development boards) and a lot of memory (good luck
> > > > > on Raspberry Pis with small memory like Pi Zeros).
> 
> > > > > I propose following approach to address the drawback: Create cpio
> > > > > archives (compressed with xz -9) for the kernel modules and firmware
> > > > > files when building the kernel/firmware Debian package. Then ship those
> > > > > cpio archives in the package (or in a separate binary package). Then the
> > > > > CPU load it put on the builders. The cpio archives would contain the
> > > > > modules for MODULES=most.
> 
> > > > > mkinitramfs will then look for those cpio archives and uses those in
> > > > > case they are present. Such a initramfs would look like this:
> 
> > > > > * AMD/Intel microcode cpio archive (on amd64)
> > > > > * main cpio archive compressed with zstd -1
> > > > > * kernel modules from the Debian package compressed with xz -9
> > > > > * firmware files from the Debian package compressed with xz -9
> 
> > > > > After working on initramfs-tools as part my day job, my fingers were
> > > > > itching and I had to create a quick and dirty draft in my free night
> > > > > time. You can find the result of the last two hours in [1]. This draft
> > > > > has a mkinitramfs-kernel script that creates a cpio archive containing
> > > > > the kernel modules and firmware (that needs to be split later on).
> 
> > > > > The lunar test result on my AMD Ryzen 7 5700G look promising: Building
> > > > > 6.2.0-24-generic-modules-most.cpio.xz takes around 90 seconds and is
> > > > > 54.9 MiB in size. Creating the initramfs speeds up from around 8.7
> > > > > seconds to 3.5 seconds (saves 60 %). The size reduces from 133.1 MiB to
> > > > > 80.7 MiB (saves 39.4 %). So the boot needs 52.4 MiB less, but
> > > > > /lib/modules need 54.9 MiB for the cpio archive.
> 
> > > > > The drawback is that building the kernel would take longer, the package
> > > > > takes more space on the archive and mirrors, and downloading them could
> > > > > take longer on slow connections.
> 
> > > > > Implementing my proposal would be relative easy for initramfs-tools, but
> > > > > would mean some work for the kernel team.
> 
> > > > > What do you think?
> 
> > Will the user still be able to add further modules and will machine specific
> > configuration files (e.g. for booting from iSCSI) still be included into the
> > initrd?
> 
> I think a robust implementation of this on the initramfs-tools side looks
> like:
> 
>  - identify all the contents that belong in the initramfs
>  - among those contents, find all zstd-compressed files, if any, and store
>    them in an uncompressed initramfs
>  - put the rest of the contents in a compressed initramfs
> 
> This would be compatible with kernel packages whether they are changed to
> ship zstd-compressed modules or not and allow for a smooth transition.

This is exactly what I tried to implement over the weekend. The result
looked promising. The creation time and initramfs size decreased by a
few percent. Booting that initramfs failed and today I figured out why.
Now I have a new draft version that works, but is slower to compress,
because it moves the compressed files around after running depmod.
Fiddling with the list of files passed to cpio should give back the
speed, but that is a task for later.

So without further ado here is the result so far:

| test        | size / b   | uncompressed / b |
|-------------|------------|------------------|
| dt original | 139679329  | 481805575        |
| dt zstd -19 | 133992932  | 195432237        |
|-------------|------------|------------------|
| vm original |  68421993  | 229445821        |
| vm zstd -19 |  62147141  | 111007617        |
| vm xz -9    |  58592562  | 107451029        |

or converted to MiB:

| test        | size / MiB | uncomp. / MiB |
|-------------|------------|---------------|
| dt original | 133.2      |  459.4        |
| dt zstd -19 | 127.8      |  186.3        |
|-------------|------------|---------------|
| vm original |  65.3      |  218.8        |
| vm zstd -19 |  59.3      |  105.9        |
| vm xz -9    |  55.9      |  102.5        |

Legend:

dt: AMD Ryzen 7 5700G on lunar
vm: AMD Ryzen 7 5700G in mantic VM

So the size of the initramfs decreases slightly, but the uncompressed
size is less than half the size. So booting a Pi Zero 2 with 512 MiB
memory could be doable again (or any other system with small memory).

How to measure the boot speed in an accurate way? The number provided by
systemd-analyze fluctuate a lot. Take the following numbers with a grain
of salt. I averaged them over three measurement and excluded boots that
took over 4 second:

| test        | kernel | userspace | sum    |
|-------------|--------|-----------|--------|
| vm original | 1.619s | 1.931s    | 3.550s |
| vm zstd -19 | 1.360s | 1.965s    | 3.325s |
| vm xz -9    | 1.280s | 1.940s    | 3.220s |

That result hints that the boot speed could increase by some percent.

If you want to test it yourself, you can find initramfs-tools
0.142ubuntu7bd2 for mantic in my PPA:
https://launchpad.net/~bdrung/+archive/ubuntu/ppa

-- 
Benjamin Drung
Debian & Ubuntu Developer