LINUX PERFORMANCE EXPERIMENT RECORD — Yenra, September 5, 2026
https://yenra.com/high-performance-linux/

DEFINE THE EXPERIMENT
Useful work and success measure:
Input identity/hash, output correctness check, application/build/library versions:
Hardware, topology, RAM, storage, distribution, kernel, virtualization/cgroup limits:
Power/thermal conditions, background work, cache state, worker/library settings:
Hypothesis and the single change being tested:
Trial order, warm-up, repetition count, timing method, and units:

OBSERVATIONS
Elapsed times for every trial; median and spread:
CPU/thread behavior, memory/pressure, storage latency/throughput:
NUMA placement, worker count, nested parallelism:
Correctness, memory use, latency, impact on other jobs:
Conclusion, uncertainty, repeat test, and retained/rolled-back configuration:

FICTIONAL DATA — NOT MEASURED HARDWARE RESULTS
Same fixed job and checked output; comparable execution slots called workers.
Workers | three invented elapsed times (seconds) | median
1       | 39, 40, 41                            | 40
2       | 23, 24, 25                            | 24
4       | 13, 14, 15                            | 14
8       | 12, 13, 14                            | 13

Speedup = one-worker median / candidate median.
Efficiency = speedup / worker count, displayed as a percent.
Use unrounded values for subsequent calculations.
1 worker: 1.00x, 100.0%
2 workers: 1.67x, 83.3%
4 workers: 2.86x, 71.4%
8 workers: 3.08x, 38.5%
4 -> 8 workers: (14-13)/14 = 7.1% less elapsed time.
Allocated worker-seconds: 4*14 = 56 versus 8*13 = 104.
These are allocation-time units, not measured CPU time or energy consumption.