Loading from squashfs with zstd compression slower than ext4

[Jonathan Liu]

Hi Simon,

On Thu, 1 Jun 2023 at 14:10, Jonathan Liu <net147 at gmail.com> wrote:
> On Thu, 1 Jun 2023 at 11:27, Jonathan Liu <net147 at gmail.com> wrote:
> > On Thu, 1 Jun 2023 at 03:13, Simon Glass <sjg at chromium.org> wrote:
> > > On Wed, 31 May 2023 at 06:52, Jonathan Liu <net147 at gmail.com> wrote:
> > > >
> > > > Hi All,
> > > >
> > > > I noticed that loading from squashfs with zstd compression is
> > > > significantly slower than ext4 on RK3399 using U-Boot 2022.01. The
> > > > filesystem is located on eMMC.
> > > >
> > > > I have CONFIG_BLOCK_CACHE=y in .config and am using the generic load command.
> > > >
> > > > ext4:
> > > > 15682578 bytes read in 366 ms (40.9 MiB/s)
> > > >
> > > > SquashFS with zstd compression:
> > > > 15685001 bytes read in 2339 ms (6.4 MiB/s)
> > > >
> > > > Any suggestions on how to speed up loading from SquashFS with zstd compression?
> >
> > >
> > > Could it just be that zstd is slow? Have you tried lz4?
> > >
>
> >
> > I tried LZ4 but get "Error: unknown compression type."
> > It seems the SquashFS decompressor only supports lzo, gzip and zstd -
> > https://source.denx.de/u-boot/u-boot/-/blob/master/fs/squashfs/sqfs_decompressor.c
> >
> > ext4:
> > 15682578 bytes read in 366 ms (40.9 MiB/s)
> >
> > SquashFS with lzo compression (~129 MiB squashfs):
> > 15686657 bytes read in 1032 ms (14.5 MiB/s)
> >
> > SquashFS with gzip compression (~116 MiB squashfs):
> > 15686657 bytes read in 1494 ms (10 MiB/s)
> >
> > SquashFS with zstd compression (~107 MiB squashfs):
> > 15685001 bytes read in 2339 ms (6.4 MiB/s)
>
> SquashFS with zstd compression seems to be even slower after zstd was
> updated to 1.5.2 in U-Boot.
>
> SquashFS with zstd compression (~107 MiB squashfs), with zstd 1.5.2
> update cherry-picked from
> https://source.denx.de/u-boot/u-boot/-/commit/4b9b25d943ff95c7421cab261333fc29852fe3b1:
> 15686657 bytes read in 2726 ms (5.5 MiB/s)

If I use unsquashfs on the same board to decompress the same file
using a single processor, it is much faster:
# sync; echo 3 > /proc/sys/vm/drop_caches; time unsquashfs -p 1
/dev/mmcblk1p1 /boot/test-file
Parallel unsquashfs: Using 1 processor
1 inodes (120 blocks) to write

[=========================================================================================================|]
120/120 100%

created 1 files
created 2 directories
created 0 symlinks
created 0 devices
created 0 fifos
real    0m 0.19s
user    0m 0.16s
sys     0m 0.12s

Regards,
Jonathan