innodb_read_ahead_threshold (linear read-ahead) does not work

We upgraded an MariaDB Galera Cluster from Version 10.5.13 to 10.9.3. Everything is up and running, from a client's point of view.

However we see a regression in I/O access when running mariadb-dump. A custom script (dumping and compressing every table in separate file) ran for about 45 minutes with MariaDB 10.5.13, the time increased to about two and a half hour with MariaDB 10.9.3.

A simple cp can do about 500MB/s on the machines, with mariadb-dump we get about 25MB/s - limited by disk I/O (where mariadbd is in I/O wait / state "D"). MariaDB 10.5.13 did something about 100MB/s - limited by compression, so could probably do a lot more.

This happens with all available I/O methods, tried with aio, uring and native. Also tried different filesystems (ext4 & btrfs) which does not make a difference. Same results with a cluster node and standalone (non-Galera) installation.

No idea if I/O for regular SQL queries regressed as well. The machines have a reasonable amount of RAM, so no delay is noticeable there. Running mariabackup for Galera state transfer reaches expected transfer rates.