On Sun, Jan 8, 2023 at 2:28 AM Jeff Law jlaw@ventanamicro.com wrote:
On 1/6/23 23:41, David Abdurachmanov wrote:
Summary from multi-year discussions/feedback on this:
- We don't have proper hardware to put into the data center that holds
servers used by Fedora infrastructure.
Right. dev boards are not the solution here. It's got to be something that can be racked and with enough performance to matter.
- Not enough single and multi thread performance to avoid large impact
to Fedora development.
Agreed. Returning to a situation like we had with armv7 isn't acceptable IMHO.
Other than that Fedora/RISCV 37 is the first Fedora version to be built fully natively using 20+ SiFive HiFive Unmatched boards. On a good day we can keep up (if the builds aren't too large, e.g. GCC). We don't really run Bodhi thus once package is built it's immediately available. We run a very minimal setup right now (ideas to expand that).
It's fantastic we've got that far. But clearly it's not viable long term.
Agreed. We have been cooking Fedora/RISCV since 2016, but we really cannot move forward until the proper hardware (and things around it) becomes available.
Another news is that Fedora/RISCV Koji server ( http://fedora.riscv.rocks/koji/ ) just moved into Fedora infra owned server. We are about to start work on Fedora 38/Rawhide.
Excellent. I'll have to update my chroots and containers as the F38 bits start appearing.
I am still tweaking the server configuration, but I should be ready for mass building soonish. I might want to wait for GCC 13 to land in Rawhide, which should happen soon (I think).
2023 is potentially a transition year. Ventana Veyron V1 Development Platform looks good (I assume it has BMC). SOPHGO SG2042 should also show up with 64-core @ 2.0GHz (T-HEAD C910) in early 2023 (?) I am not yet convinced about their upstream support efforts (TBD).
Yes Veyron-v1 will have a BMC and will be rackable. I have no significant insight into the T-HEAD efforts other than they do work a fair amount with VRULL on compiler and related technology.
I noticed that VRULL has been involved with T-HEAD on GCC and potentially kernel side for a few months now. This makes them much more optimistic about their SoC/HW support in general.
If there is away to acquire Veyron V1 Development Platform I would be interested to talk, and figure out what that would take. Such hardware would be a game charger, and I do trust Ventana regarding upstream support :)
I'll be pushing to make systems available to Fedora and the GCC farm, but in general Ventana is more aligned towards Ubuntu. My goal is to have Fedora and Ubuntu on equal footing as quickly as possible.
We have been trying to get stuff into GCC Compiler Farm for years now. There are a couple of boards IIRC. There are additional requirements on the software side (well, distributions) that we couldn't provide for quite some time.
I do know rackable systems will be limited -- there's one particular component needed on the rackable systems that is in very short supply. We've got multiple orders in, but quantities are limited and lead times are absolutely insane.
FYI, I think, the new aarch64 builders are 8 core, 35G RAM and 8G swap. The older machines had 8G/core setup. There seems to be 8 (?) servers with 80 cores, but so far only 40-50 builders are enabled (no overcommitting on CPU as that's not a great idea [based on my own experience]).
I expect Veyron V1 to deliver a decent single and multi thread performance, but we will need a lot of them. Probably 20-25 systems if we assume a similar configuration as aarch64 builders (8-core, 32-64G of RAM, 100-200G for storage). RAM and storage are cheap, but systems will continue to be a problem. If we could somehow get to this level we could stop investing into SBCs as they are stop-gap solutions for builders.
Based on some guesses there isn't a lot of time either. I would love to bootstrap CentOS Stream 10. It would be nice to have Fedora + riscv64 in a good shape before that happens, but probably unrealistic.
Cheers, david