Oversubscribed CPU Limits Build Speed | Generated by AI
Here’s the verdict based on live system state:
System: 20 CPU cores, 235GB RAM (218GB available)
Current build (MAX_JOBS=32):
- 32 active clang-22 processes, all CPU-saturated (50-90% each)
- Progress: 1158/2525 objects (~46%)
- Load average: 34.4 — already 1.7x oversubscribed
- RAM usage: ~17GB of 235GB (plenty of headroom)
Can you set MAX_JOBS=64?
Technically yes, but it will NOT help and will likely hurt:
- CPU bottleneck: 20 cores × 100% = 20 full cores of work. With 32 jobs you’re already at ~170% utilization. Adding 32 more jobs just means more context switching between the same 20 cores.
- RAM is fine: 64 jobs would use ~12-50GB, well under 235GB.
- Diminishing returns: The build is CPU-bound, not memory-bound. You’d see load average jump to ~60+, but actual compilation rate would stay flat or drop due to scheduler thrashing.
- Risk: Occasional OOM kills from transient spikes, or hipcc failures from resource starvation.
Recommendation: Keep it at 32. With 20 cores, the sweet spot is probably 20-24 jobs. Going above that just adds contention. If you want faster builds, the bottleneck is the number of CPU cores, not job count.
Generated by AI. Curating and sharing still takes effort.
If you find it useful, feel free to
donate.
WeChat: @lzwjava
·
X: @lzwjava ·
Say hi 👋