| 23 Aug 2024 |
 roberth | In reply to @joerg:thalheim.io
Robert Hensing (roberth): Do you have access to macOS or do you need something like a flamegraph for nix 2.18 vs 2.24? not a true x86_64 one | 17:28:10 |
| roberth | I'm fairly confident that 2.18 is faster because we added a cache we're not taking much (or any?) advantage of yet. The reason is we're moving to a setup that will be faster than 2.18 but that's taking longer than expected to get there. I think we should look into disabling that for now, but I haven't found the time to confirm this yet, so caveat. | 17:34:13 |
 Mic92 | @roberthensing:matrix.org: i could try to get you ssh access to one in 1-2 weeks. What also works very well is this: https://github.com/Mic92/dotfiles/blob/main/.github/workflows/os-ondemand.yaml | 18:12:06 |
| Mic92 | It's also a nice devenv for windows | 18:12:28 |
| Mic92 | GitHub still does x86 for Mac. So that's probably the easiest option | 18:13:11 |
| Mic92 | If you can point me to the cache I.e the commit, I can check if disabling it performs better | 18:14:53 |
 emily | I think that GitHub Actions thing is against the Actions terms of service FWIW (your judgement call to make: just making it known) | 18:21:43 |
 jade_ | i would personally recommend a certain "osx-kvm" which is at least only a license violation of macOS purely on your own computer | 18:23:31 |
| roberth | In reply to @joerg:thalheim.io
If you can point me to the cache I.e the commit, I can check if disabling it performs better Not sure if useful but it'd be
- originally https://github.com/NixOS/nix/pull/9485
- however, performance was improved in a later PR
- and as you know may be improved again by https://github.com/NixOS/nix/pull/11330
So I don't know how representative the delta of that PR would be anymore
| 18:24:22 |
| emily | well, also probably a more serious copyright violation to obtain the things you need to use it unless you already own a Mac… | 18:24:50 |
| emily | but sure, YMMV | 18:24:51 |
| emily | I can still be pinged for "running commands that print out numbers" purposes | 18:25:18 |
|  @9999years:matrix.org left the room. | 18:25:19 |
| emily | the patch Robert Hensing (roberth) asked me to benchmark looked pretty good | 18:25:24 |
| emily | dunno if it restores things to 2.18 perf though | 18:25:30 |
 niksnut | Note that the Git-based tarball cache only got merged in 2.21 or so, so it doesn't explain the performance regression in 2.19 | 18:25:52 |
| roberth | yeah I think that's the question Mic92 is trying to help with | 18:25:57 |
| jade_ | In reply to @niksnut:matrix.org
Note that the Git-based tarball cache only got merged in 2.21 or so, so it doesn't explain the performance regression in 2.19 oh that's caused in part by an ill advised migration to string_view in Value that we dropped from lix due to performance regression | 18:27:46 |
| roberth | I think for a more productive, actionable measurement, we could try using the store as a "tarball cache" instead. That seems like a viable stop gap to improve performance until we have lazy source semantics | 18:28:22 |
| niksnut | In reply to @jade_:matrix.org
oh that's caused in part by an ill advised migration to string_view in Value that we dropped from lix due to performance regression How does that explain the minutes-long slowdown on macOS and some docker containers? | 18:29:38 |
