arm64 as Tier 1 for FreeBSD 13

[Mark Millard]

On 2020-Jan-29, at 19:21, Klaus Küchemann via freebsd-arm <freebsd-arm at freebsd.org> wrote:

> From my personal view the RPI-thing is overrated ,
> it’s just another 4-core.. and if you ever made e.g.
> a -j16 buildworld buildkernel in 90 minutes or so  on aarch64 : you won’t 
> expect the big revolution in computing from this RPI-gadget. :-)
> 

I'm not sure that the RPi4B support has progressed to the point
that benchmarking that is limited by cpu/memory performance is
expected to be near where it would be in the long run if the
RPi4B ends up well supported.

So I did an single-threaded experiment comparing a Ubuntu RPi4B
configuration that was using:

over_voltage=6
arm_freq=2000

(I only used those for Ubuntu because they were already
set up for other reasons) to whatever results currently
for FreeBSD on the same RPi4B. The context is
CPU/RAM-caches/RAM limited, not an I/O benchmark at all.

For a problem size were Ln caches are effective, I got
an example rates of progress ratio for Ubuntu/FreeBSD of
around 2.8.

For a problem size where the RAM access pattern makes
the Ln caches not very effective, I got an example rates
of progress ratio for Ubuntu/FreeBSD of around 1.2.

The RPi4 is rather constrained by its RAM subsystem,
but not so much where the Ln caches are effective.
So, using the 2.8, I get a quick ball-park estimate for
the CPU running at something like 700 MHz for FreeBSD
currently. (I was using a head -r356426 based context.)

If the RPi4B ends up well supported, the usual RPi4B
default of 1500 MHz is likely and FreeBSD might support
use of the something analogous to setting over_voltage
and arm_freq, as was done for Ubuntu.

Until things progress to such a point, I'd be cautious
interpreting real-world performance experiments as a
source of estimates for future performance.


===
Mark Millard
marklmi at yahoo.com
( dsl-only.net went
away in early 2018-Mar)