[Trisquel-devel] success on arm64 Supermicro ARS-110M-NR

Tue Jul 9 11:14:54 UTC 2024

Hi Simon and bill-auger and maybe Luis and Rubén,

I also have good news! I just installed Trisquel on my new arm64 server 
and it looks good. I hope I can use this server to help Trisquel's 
development in some way.

On 2024-06-16 10:21, bill-auger wrote:
> congrats - AFAIK, this is the first powerful non-x86 computer that is 
> known to
> run a libre distro; so it is a significant landmark

Talos II (2019) is also powerful too...? FSF has one I think.

> if you would like to run an experiment, i would be interested to know 
> ho much
> time it takes to compile linux on that box, or some other heavy build 
> like
> firefox or blender

(Please scroll down to see some test/benchmark results of linux and 
gcc-14.)
Firefox has too many build-deps, so I didn't want to build it.

I recently acquired a budget DIY system that is similar to Simon's 
Supermicro 1U server. The motherboard is Asrock ALTRAD8UD-1L2T. The CPU 
is an engineering sample of Ampere Altra Max I got from ebay, it has 128 
cores and claims that it boosts to 3.0GHz (although it could only reach 
2.8GHz `cat /sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_cur_freq`, so 
slower than Simon's). 8 sticks of common 16GB DDR4 ECC RDIMM.

Unlike Simon, I used the prebuilt Trisquel ISO:

https://repo.jing.rocks/trisquel-iso/trisquel-netinst_11.0.1_arm64.iso

See, that's the benefit of hosting a mirror for Trisquel :)

The installation was so smooth that I started to suspect something would 
go wrong, given my experience with Rockchip-based SBCs. But luckily, 
things just worked. I used the "expert install" option and formatted the 
rootfs as btrfs, with linux-generic-64k-hwe-11.0-edge for the kernel, 
and no desktop environment (don't have a mouse).

Time for some numbers:

Building trisquel's linux 5.15 using sbuild took about 17 minutes, with 
the following command:

sbuild -d aramo --no-arch-all --no-run-lintian --no-run-autopkgtest 
--no-run-piuparts linux_5.15.0-113.123+11.0trisquel30.dsc

(found a bug in trisquel's debootstrap, will suggest a fix on gitlab 
later.)

Since the dsc also produces many unimportant deb/udeb packages, runs 
many debhelper programs, I also attempted to `make` using the source 
tarball only, to get a better result. allyesconfig failed to build from 
source (FTBFS), defconfig took 57 seconds.

Then, to establish a base line for test/benchmark results, I downloaded 
linux 6.1.94 from Debian and ran `make` on my Ampere Altra and my AMD 
EPYC, defconfig and allyesconfig, separately:

(command) make distclean && make defconfig && time make -j128

arm64 defconfig: real 1m19.278s, user 97m42.929s, sys 20m5.660s
amd64 defconfig: real 0m54.487s, user 24m19.703s, sys 39m33.731s
arm64 allyesconfig: real 13m3.830s, user 427m32.619s, sys 56m36.050s
amd64 allyesconfig: real 11m58.768s, user 311m59.228s, sys 475m12.674s

Next, gcc 14.1, which should be a more accurate comparison because linux 
has too different configs on different arch:

(command) ../configure --enable-languages=c,c++ --enable-threads=posix 
--enable-default-pie --enable-default-ssp --with-system-zlib 
--disable-multilib
then `time make -j128` as usual:

arm64: real 26m5.756s, user 340m30.708s, sys 11m3.072s
amd64: real 23m5.812s, user 278m49.675s, sys 26m57.191s

`time make -k -j128 check`:

arm64: real 16m24.069s, user 1563m3.285s, sys 375m6.749s
amd64: real 14m20.548s, user 753m53.327s, sys 672m37.120s

So...if Luiz or Rubén is reading this, I would be happy to host a VM for 
Trisquel to speed up jenkins builds, seeing that Trisquel is using a 
ROCK 3A for the build farm. It should be 20~30 times faster. I plan to 
use Proxmox (should be ported to arm64 first) or libvirtd to manage VMs, 
and upgrade RAM to 512GB around this September. I'm also doing more 
testings since I got an engineering sample CPU instead of a retail one, 
to see if it's stable or has any CPU bug.

-- 
Jing Luo
About me: https://jing.rocks/about/
GPG Fingerprint: 4E09 8D19 00AA 3F72 1899 2614 09B3 316E 13A1 1EFC
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 228 bytes
Desc: OpenPGP digital signature
URL: <http://listas.trisquel.info/pipermail/trisquel-devel/attachments/20240709/a93eacb1/attachment.sig>