| 24 Jun 2026 |
 Vladimír Čunát | pytest I suspect | 13:47:58 |
 hexa | scikit-bio is my guess | 13:48:16 |
| Vladimír Čunát | I mean, all the culprits seem to be running pytest tests right now. | 13:48:37 |
| Vladimír Čunát | So perhaps it's what pytest does in usual setups? | 13:49:07 |
| hexa | yeah, but some are doing openmp | 13:49:11 |
| hexa | pytest-xdist in nixpkgs respects NIX_BUILD_CORES | 13:49:27 |
| hexa | also, how is this anything new? | 13:49:44 |
| hexa | we still schedule the same max number to elated-minsky as before | 13:49:51 |
 K900 | I don't think it is | 13:49:55 |
| hexa | but now we stop after cpu some exceeds 70% :) | 13:49:59 |
| K900 | We're just seeing more occupancy than before | 13:50:01 |
| K900 | So we hit that more often | 13:50:07 |
| hexa | I think we need to fix some packages in response | 13:50:14 |
| hexa | find out how to limit libgomp parallelism during tests | 13:50:27 |
| hexa | 53.40% libgomp.so.1.0.0 [.] gomp_team_barrier_wait_end
42.08% libgomp.so.1.0.0 [.] gomp_barrier_wait_end
3.01% libopenblas.so.0.3 [.] inner_thread
0.28% libopenblas.so.0.3 [.] inner_thread
0.19% libopenblas.so.0.3 [.] inner_thread
0.14% .perf-wrapped [.] queue_event
0.03% .perf-wrapped [.] perf_sample__init
0.03% libc.so.6 [.] _int_malloc
0.02% .perf-wrapped [.] hist_entry__cmp_impl.constprop.2
0.02% .perf-wrapped [.] evlist__event2evsel
0.02% libopenblas.so.0.3 [.] inner_thread
| 13:51:03 |
| Vladimír Čunát | I agree it does not look related to the new queue-runner. | 13:51:19 |
| hexa | something OMP_NUM_THREADS | 13:51:33 |
| hexa | also I noticed that nodejs is not marked big-parallel | 13:52:26 |
| hexa | I'm retry to repro scikit-bio now | 13:52:58 |
| Vladimír Čunát | Nested parallelism is hard. | 13:53:23 |
| hexa | patching libgomp to respect NIX_BUILD_CORES? 🤔 | 13:53:59 |
| Vladimír Čunát | We have jobs, then make-level parallelism or equivalent and underneath you might have OpenMP. | 13:53:59 |
| Vladimír Čunát | You can still get a square of what you wanted. | 13:54:17 |
| hexa | but openmp cores are on top of the actual build cores already | 13:54:19 |
| Vladimír Čunát | I mean, for elated-minsky in particular, 2^2 isn't bad, so at least some improvement. | 13:54:50 |
| hexa | oh, yeah | 13:56:59 |
| hexa | here we are. 64 threads at 83% test progress | 13:57:10 |
| hexa | https://github.com/NixOS/nixpkgs/pull/534961 | 14:09:35 |
| hexa | propagatedPreCheck when | 14:09:58 |
| hexa | mace-torch is another omp consumer | 14:27:05 |
