11 Oct 2024 |
Tristan Ross | In reply to @rick:matrix.ciphernetics.nl Isn't that compression on the builders? No | 05:11:24 |
Tristan Ross | https://github.com/NixOS/hydra/issues/553 | 05:11:35 |
Mindavi | Ah, like that | 05:14:43 |
Mindavi | Ok now it makes sense again :) | 05:14:55 |
vcunat | I'm fairly certainly that the compression of results on the central machine is what's limiting the throughput almost all the time, as the farm is currently. | 06:11:34 |
emily | an EPYC with better thermals seems like it'd be a good investment, then? | 06:12:06 |
vcunat | We have some bigger aarch64 in the farm already, so alternatively we could try that "for free". | 06:13:08 |
emily | wouldn't that run into the atomics and platform purity problems wrt the evaluator? | 06:13:59 |
vcunat | Right. It might be risky in some ways. | 06:15:46 |
emily | I take it setting up a separate Hydra instance to test it would be onerous – we'd need to just switch over? | 06:17:32 |
vcunat | I don't see a little down-time as an issue (for hydra.nixos.org) | 06:20:38 |
emily | I more meant it sounds like a pain to revert if it goes wrong | 06:22:16 |
emily | (but maybe it's not) | 06:22:32 |
vcunat | No need to wipe the old machine too soon. And it seems unlikely that we'd want to revert completely. | 06:23:14 |
tomberek | In reply to @emilazy:matrix.org if the Foundation will pay for a month then it seems worth it to just do it? Not too hard. We're spending similar amounts on meetups/pizza. Infra team should have a discretionary budget (no?), otherwise throw a submission into this repo similar to this: https://github.com/NixOS/foundation/issues | 06:39:11 |
emily | I didn't get the impression that the infra team has a budget at all | 06:39:33 |
emily | though I don't know how the finances work at all :) | 06:39:44 |
emily | In reply to @vcunat:matrix.org No need to wipe the old machine too soon. And it seems unlikely that we'd want to revert completely. I think we should do whatever you think is best for Hydra out of trying an existing AArch64 machine and getting an EPYC to try. | 06:40:24 |