VisionFive 2 Lite runners: a RISC-V matrix cell without the vector extension issue #95
open andymccall opened this · 0 comments
andymccall
Andy has two VisionFive 2 Lite boards to put in as runners. They're worth it as a RISC-V matrix cell and for redundancy, not as queue relief.
What they are: StarFive JH7110 (4× SiFive U74 at about 1.25 GHz). There's no hardware virtualisation (no H extension), so no microVM: they'd use the syscall tier like narvi-1. There's no vector extension either.
Why bother:
A real matrix cell (issue #89, issue #90). narvi-1's SpacemiT K1 has RVV 1.0. A JH7110 doesn't, so running riscv64 jobs there tests that our riscv64 builds don't quietly assume vector instructions.
Redundancy: riscv64 checks would no longer hang on one board.
What they won't do:
Help the x86 queue. They take riscv64 images only.
Run fast. narvi rates 6 against Earth's 82, and a U74 is slower again; the riscv64 job has a 3-hour timeout.
Memory: the riscv64 sandbox asks for 4 GiB, so only the 8 GB model fits it with room for the runner. A 4 GB board would need a smaller sandbox, or to take only lighter jobs.
Setup, once: Ubuntu or Debian riscv64; a user runner with a delegated cgroup, like narvi-1; labels riscv64 and vf2; the same rootfs pin as narvi.
Andy has two VisionFive 2 Lite boards to put in as runners. They're worth it as a RISC-V matrix cell and for redundancy, not as queue relief.
What they are: StarFive JH7110 (4× SiFive U74 at about 1.25 GHz). There's no hardware virtualisation (no H extension), so no microVM: they'd use the syscall tier like narvi-1. There's no vector extension either.
Why bother:
What they won't do:
Memory: the riscv64 sandbox asks for 4 GiB, so only the 8 GB model fits it with room for the runner. A 4 GB board would need a smaller sandbox, or to take only lighter jobs.
Setup, once: Ubuntu or Debian riscv64; a user runner with a delegated cgroup, like narvi-1; labels
riscv64andvf2; the same rootfs pin as narvi.