Post AVybsfks89cab3D2qu by flokli@im-in.space
(DIR) More posts by flokli@im-in.space
(DIR) Post #AVyMVcTQfL1ERlbGYS by mic92@toot.matereal.eu
2023-05-24T05:54:24Z
1 likes, 1 repeats
Tvix vs nix evaluation of the hello package on a AMD EPYC 7713P: Summary 'nix eval --impure --expr 'with import <nixpkgs>{}; toString hello'' ran 14.26 ± 0.18 times faster than './target/release/tvix -E 'with import <nixpkgs>{}; toString hello''Full report: https://gist.github.com/Mic92/2edb7d1afa861dffd2f601f1de78cb87
(DIR) Post #AVyOUtHesdwMCXsHp2 by mic92@toot.matereal.eu
2023-05-24T07:33:38Z
1 likes, 0 repeats
I just got told. That nix eval is eager and I should compare it to nix-instantiate instead. New data is in the gist.
(DIR) Post #AVyOYhtPAsMS1xLlQG by flokli@im-in.space
2023-05-24T06:06:23Z
1 likes, 0 repeats
@mic92 As written in https://im-in.space/@flokli/110412679019389276, we shell out to nix-store --add, and benchmarking makes little sense while it's like this.For now we're working on compatibility.
(DIR) Post #AVybschvSeTx92o9LM by flokli@im-in.space
2023-05-24T08:07:43Z
0 likes, 0 repeats
@mic92 can you benchmark stdenv.outPath with tvix from https://cl.tvl.fyi/c/depot/+/8572/33 ?That doesn't shell out to nix-store, but uses the tvix-store model (https://cs.tvl.fyi/depot/-/blob/tvix/store/docs/api.md) under the hood. We still didn't do any performance testing, and there's some known low-hanging fruits, but it's probably a fairer comparison for now.
(DIR) Post #AVybseq9X2Fdl8fkRc by flokli@im-in.space
2023-05-24T09:02:43Z
1 likes, 0 repeats
https://gist.github.com/flokli/0fa74eb334a3c388df0f2cd85f281316
(DIR) Post #AVybsfks89cab3D2qu by flokli@im-in.space
2023-05-24T09:06:36Z
1 likes, 0 repeats
I'm not sure how much of this time tvix currently spends rendering all the warnings about unused variables, and as I said, we didn't do performance-tuning, so there's some low-hanging fruits.