Offer to host RISC-V CI runner #12263

Activity

1. danog

   added

   Feature

   Status: Needs Triage

   on Sep 21, 2023
2. mvorisek commented on Sep 22, 2023

   

   mvorisek

   on Sep 22, 2023

   Contributor

   More actions

   Wouldn't QEMU for RISC testing be better using standard GH (or any Docker based CI)?
3. danog commented on Sep 22, 2023

   

   danog

   on Sep 22, 2023

   Last edited by danog

   ContributorAuthor

   More actions

   QEMU is super slow compared to real hardware, for example here are some basic benchmarks compiling php-src, running on my visionfive2 board and on QEMU on an AMD EPYC 7401P:

   ./configure --enable-debug --without-pcre-jit --without-sqlite3 --without-pdo-sqlite

   starfive (1m14s)

   real  1m14.050s
   user  0m56.293s
   sys  0m18.814s
   

   QEMU (6m45s)

   real	6m 45.44s
   user	5m 52.75s
   sys	0m 49.56s
   

   make -j4

   starfive (7m52s)

   real	7m52.868s
   user	27m15.208s
   sys	2m27.981s
   

   QEMU: still in progress after 30m, aborted...

   AFAIK, all CIs currently used to test other php-src on arm64 and s390x run on native hardware.
   Woodpecker is a great FOSS CI system based on Docker meant for self-hosting, unlike GHA self-hosted runners which are an absolute mess to work with (and currently don't support riscv64, unless we were to use a reverse-engineered Golang implementation of GHA runners like https://github.com/ChristopherHX/github-act-runner).
4. NattyNarwhal commented on Sep 22, 2023

   

   NattyNarwhal

   on Sep 22, 2023

   Member

   More actions

   I’d be worried about things like QEMU not properly emulating atomics, memory model, etc. Things that aren’t obvious, but can trip you up badly.