| 22 Jul 2024 |
 Atemu | ~25 rather than ~17 | 12:17:12 |
| Atemu | Still a long way off from the 70 it had originally | 12:17:24 |
 woobilicious | I wonder if it's just an L3 cache issue. | 12:19:08 |
| woobilicious | killing frames by flushing the cache too much. | 12:19:29 |
 K900 | That would actually explain things | 12:20:01 |
| K900 | Because I am on X3D | 12:20:06 |
| K900 | So I have a lot more L3 | 12:20:15 |
| Atemu | Remember that we use the same processor ;) | 12:22:48 |
| Atemu | I'd be very surprised if you somehow had more L3 cache than me | 12:23:04 |
| Atemu | So making the daemon SCHED_BATCH actually makes a difference | 12:23:26 |
| Atemu | ~26 at SCHED_OTHER and ~30 with SCHED_RR | 12:23:57 |
| woobilicious | Atemu: is this nixbld or something else? | 12:24:16 |
| Atemu | Yes of course | 12:24:21 |
| Atemu | nix.daemonCPUSchedPolicy = "batch"; | 12:24:37 |
| woobilicious | I would have throught batch would have been worse. | 12:24:57 |
| Atemu | Me too | 12:25:02 |
| Atemu | But, apparently, scheduling priorities are just broken in the kernel | 12:25:16 |
| Atemu | At least for interactive use-cases | 12:25:25 |
| woobilicious | I have my /tmp in tmpfs so I bet that ruins performance lol | 12:26:10 |
| woobilicious | nix.daemonIOSchedClass might be worth tweaking | 12:27:28 |
| Atemu | As long as you have enough memory, it's fine | 12:27:41 |
| Atemu | Though I never understood why it's the default | 12:27:54 |
| K900 | Oh yeah, I have Nix set up to build on /var/tmp | 12:28:19 |
| K900 | Which is not tmpfs | 12:28:21 |
| K900 | Building on tmpfs is definitely going to kill perf | 12:28:33 |
| Atemu | Wait what? | 12:28:46 |
| Atemu | Kill perf in what way? | 12:28:49 |
| K900 | It'll just blow out your cache | 12:29:27 |
| K900 | Because everything on the tmpfs is hot all the time | 12:29:36 |
| Atemu | Ah that's what you mean | 12:30:13 |
