mirror of
https://relay.ngit.dev/npub15qydau2hjma6ngxkl2cyar74wzyjshvl65za5k5rl69264ar2exs5cyejr/ngit-grasp.git
synced 2026-10-30 11:07:33 +00:00
DanConwayDev
484b58664a
Keep compilation, lint and tests in one job and require at least four vCPUs. Controlled production runs on de1 used the same source revision (ed0ed6fd), 8 GiB RAM, 32 GiB scratch and cached third-party Rust objects. The sequence was two CPUs, four CPUs, two CPUs again, then one CPU. Step seconds (two CPU runs / four CPUs / one CPU): compile: 125.48-145.90 / 81.98 / 248.89 lint: 61.58-63.19 / 42.89 / 121.97 tests: 503.73-506.24 / 331.55 / 946.21 Whole-guest elapsed time fell from 827.85-839.34s to 570.90s with four CPUs. Compilation benefits proportionally as much as tests; splitting it into a smaller guest would slow it and add setup/output transfer. This favors latency, not maximum CPU-slot throughput: four CPUs reserve about 37% more CPU-seconds across the whole job, while measured consumed CPU time rises from 714.17-723.93s to 778.79s. Current production admission is also constrained by scratch reservations. More CPUs are not a promise of more simultaneous jobs; remeasure if that bottleneck changes. Keep the measured 4 GiB RAM floor. Whole-guest observed peaks were 2804.4-2839.3 MiB at two CPUs and 2960.1 MiB at four. Earlier two-CPU runs peaked at 2717.7-2868.4 MiB, and an actual 4 GiB guest atf1c867e1passed at 2656.4 MiB. The four-CPU measurements used 8 GiB guests, so the 4 GiB minimum includes headroom but is not an exact-cap validation for four CPUs. Sampling includes the guest OS; all runs had swap disabled and zero OOM kills. No five-CPU measurement was made. Validation: all four controlled runs succeeded and the native workflow parser accepts the new requirement. Other jobs and commands are unchanged. Production evidence (four CPUs, including usage header): https://blossom.primal.net/f50ae89366d38e7d0c113cad626a2f8da17010f3e60e43fa778f000dd01157f0.txt Assisted-by: GPT-6