What's broken on ARM64?
Wookey
wookey@wookware.org
Who is Wookey?
- Debian Developer, ARM porter
Embedded/multiarch/crossbuild/bootstrap - Linaro, seconded from ARM
Ramble through state
- Supposed to be BOF-y
- Built, working, optimised, missing
- Hardware availability
- Benchmarking, what next
Mostly Working!
- Nearly everything builds
I'm a build engineer
If it builds, it's done.
So we are DONE :-)
Mostly Done
- Things people care about work
- Some things still need porting
But How well do they work?
- Some things may not work at all...
- Popular things are optimised
- Many things are not
- Need feedback from real users
Some History
- Started 2010 - toolchains
- 'Fast' Model 2011
- Qemu 2013
- Hardware March 2014
5 bootstraps
- Crossbuilt internal bootstrap 2011
- Crossbuilt ubuntu bootstrap 2012
- Ubuntu rebootstrap 2013/14
- Debian ports rebootstrap 2014
- Debian official rebootstrap 2014/15
Now
- 96% built (11331 packages)
- 303 failed
- 377 not for us
Languages
- Optimised
- C, C++, Java (8&9), Python, Perl, Ocaml, Javascript (v8), Haskell (ghc), Lisp (SBCL) (2 months)
- Ported
- Lua, R, Rust, Golang, Julia (2 months), Perl6 (4 months), Javascript, Pascal (fpc) (at FOSDEM),
- Missing
- Mono, libphobos (D), Luajit, ?
Still missing
- Mono (23)
- libphobos (D-libraries) (11)
Mono/C#
- Blocks 23 packages
- Done but not released
- Does anyone care enough?
- Coreclr instead? (not in Debian yet)
Not missing
- Nodejs - libv8
- golang - (base support in v1.5)
- Lua (but no luajit)
- Ocaml (native in 4.02, 2015.07)
- SBCL (2015.10)
- Perl6 (2015.10)
- Julia (2016.01)
- fpc (2016.02)
Optimisation
- C fallback
- Assembler, Intrinsics, Neon
Assembler in packages
https://wiki.linaro.org/Platform/DevPlatform/ArmSoftwareList http://performance.linaro.org/
1000 bits of assembler:
Removing often better than 'fixing' (e.g. alsalib)
What's Optimised
- GCC (compilers)
- LLVM (no linker) (compiler)
- Java 8 & 9
- openSSL
- libV8 (Javascript jit)
- fftw3 (FFT library) Neon support
- gnu-mp (gmp) (Feb 2013)
- hadoop (crc, using HWCAP)
- ceph (crc, using HWCAP)
- Kernel (raid6, crypto)
- Xen, (loads less code on arm64)
Not Optimised
- Ionmonkey (Mozilla JIT)
- Golang (being worked on)
- OpenCV
- ffmpeg/libav (parts)
- Openoffice
- Lua (luajit)
- R?
- ATLAS?
- What else?
Blocking Packages (Debian)
- mono (23), libphobos (11)
- openais (6)
- pocl (6)
- dietlibc (5)
- ffcall (5)
- insighttoolkit (3)
- timblserver (2)
- umview (3)
- kexec-tools (2)
Never Built
('Auto Not for us')
- For other arches
- nvidia-support
- nvramtool
- powerpc-utils
- raxml
- Probably should be fixed
- openafs
- scm
- qtemu
- darktable
- lots more
I have no hardware!
Online
- OVH (Runabove.com) ThunderX
- OBS (Open Build System)
- Linaro build farm
- Debian porter boxes
I have no hardware!
Hardware
- HP Moonshot
- ARM juno
- APM C1
- Softiron 3000
- Gigabyte MP30
- Hikey (96boards)
- Dragonboard 410c(96boards)
- Cello/Husky
- Pine64
Applied Micro X-gene
$1500
Juno (ARM A57)
$6000
Moonshot (HP)
$10000+
Softiron 3000 (AMD A1100)
$2500
Gigabyte MP30 (X-gene SOC)
€950
Hikey (96boards, Hisilicon/Lemaker)
$75 (1G) $99(2G)
Dragonboard 410c (96boards, Qualcomm)
$75
Cello (AMD A1100
April 2016? $300
Pine64 (Allwinner A64)
May 2016? $15(0.5GB) - $29 (2G)
Benchmarking
Benchmarking across architectures is difficult.
What are good tests?
Sometimes it's obvious (botch)
| Arch | Build Time |
|---|---|
| amd64: | 37m |
| arm64 (generic ocaml): | 4hrs 52m |
| arm64 (native ocaml): | 1hr 15 |
Benchmarking
Find good tests?
Broadly equivalent platforms
Look for changes over time