Hi,
interesting detail, thanks for letting us know.
On Wed, May 25, 2016 at 10:49:56AM -0300, Daniel Bolgheroni wrote:
> Hi,
>
> I've been following armv7 recent changes and these days I was able to
> compare kernel compile times on 2 of the supported boards.
>
> The first is an omap BeagleBone Black with a NFS-mounted /usr/src. It's
> pretty stable and I had no issues so far with what is supported.
>
> The second is a sunxi CubieBoard 2 with the Allwinner A20 SoC. This one
> has a 1.8" hdd instead.
>
> BeagleBone Black:
> (...)
> ld -T ldscript --warn-common -nopie -S -o bsd ${SYSTEM_HEAD} vers.o ${OBJS}
> text data bss dec hex
> 3684900 104816 470508 4260224 410180
> 59m00.17s real 51m31.67s user 3m43.93s system
>
> CubieBoard 2:
> (...)
> ld -T ldscript --warn-common -nopie -S -o bsd ${SYSTEM_HEAD} vers.o ${OBJS}
> text data bss dec hex
> 3684128 104816 470080 4259024 40fcd0
> 269m38.01s real 252m33.07s user 12m29.56s system
>
> The sizes are slightly different because one source tree is maybe a
> dozen commits ahead of the other, so no significant changes.
>
> The cb2 host kernel has a small workaround published by Patrick Wildt on
> bugs@, to only flush a TLB for a specified VA, not the entire TLB. It
> prevents a flood of pmap_fault_fixup messages on console. I know there
> are things to be done here, although I don't know if the slow cb2 times
> compared to bbb are related to this.
>
> http://marc.info/?l=openbsd-bugs&m=145840460330909&w=2
>
> I was expecting cb2 to be faster than bbb.
The A7 used in the A8 is probably a tiny bit slower than an A8, but
that's only a tiny difference. The most probable reason is that the
frequency the CPU is running on is different. We don't do frequency
scaling on those machines, so we boot up with whatever u-boot has set
us up with. So if the BBB was clocked higher, that would be reason.
The workaround diff is a big hack which probably adds to the slowness,
though I have no numbers on that.
>
> Related to sxie(4) on cb2, it's common to lose the ssh session with
> packet_write_poll when there is no activity. I was able to compile the
> full kernel without problem, though.
>
> packet_write_poll: Connection to 192.168.1.220 port 22: Host is down
>
> It seems the imx are much more popular on arm@ because they are faster,
> but these boards are tough to find in Brazil.
That's not really the main reason. The reason is that the i.MX6 is the
least crap of all available. There are a few other nice SoCs, but they
are only available as expensive development kits. ARMv8 might bring us
some proper server hardware, but that needs a brand new port.
Patrick
>
> Just a little feedback.
>
> --
> db
>
