> selling point was th[at] systemd is faster,
SystemD has certainly been touted that way, I personally have not found
it so in practice across a variety of hardware.

There is a catch, one that *any* init system needs to take into
consideration. Whilst a modern CPU can cope with multi-tasking most disk
systems cannot. I have observed the highest occurrences of disk
thrashing during system boot. Starting disk hungry processes in parallel
just slows down the whole system as the cumulative disk head seek times
go through the roof. You can often get better performance using the old
style sequential sysvinit scripts under these conditions.

Once every system runs from flash disk this should not be a problem,
until then, parallelism is not necessarily the best way. The init system
should be devised to proactively measure and handle this.

I used to use 'make' to launch parallel processes on a cluster dependant
upon CPU load, what would be the best way to take CPU and IO (disk,
network) into account. Suggestions anyone ?