Stress ng shows a difference with cache not in performance but in how the data is available in userspace for built-in prefetch.
Saw a weird behavior, the on-demand governor boots faster (consistently across 3 reboots).
Performance
systemd-analyze
Startup finished in 2.749s (kernel) + 6.095s (userspace) = 8.844s
kodi.target reached after 6.095s in userspace.
Ondemand
systemd-analyze
Startup finished in 2.734s (kernel) + 5.007s (userspace) = 7.741s
kodi.target reached after 5.007s in userspace.
Could be related to the IRQ rebalance in /proc/interrupts
IPI4: 268 243 279 22 169 13 IRQ work interrupts
The above numbers are right at boot.
Can someone else also try the on-demand governor and see how it goes?