What changed:

• Replaced Boost Accumulators (and headers) with a lightweight, dependency-free implementation using Welford’s online algorithm for stable mean and variance calculation.
• Introduced compile-time clock selection:
  • Prefer std::high_resolution_clock only if it is monotonic (is_steady == true).
  • Fallback to std::steady_clock for guaranteed monotonic behavior.
• Updated confidence interval calculation:
  • Return (mean_ns, 95% CI half-width in ns) instead of a rough estimate.
  • Use z = 1.96 for normal approximation.
• Increased default sample size from 100 → 1000 for better statistical stability.
• Added filtering and safety checks:
  • Filter out unrealistic timer values (> 1 ms).
  • Terminate the program (std::exit) if fewer than 100 valid samples remain after filtering.
• Improved comments and clarified output in p11perftest.cpp:
  • Print: "timer granularity (ns): mean ± (95% confidence interval)".

Why:

• Ensures portable, accurate timing measurements across platforms.
• Eliminates external dependency (Boost) for simpler maintenance.
• Guarantees monotonic clock for reliable interval measurement.
• Provides statistically meaningful confidence interval for precision reporting.
• Adds robustness by filtering out invalid measurements and failing fast when data quality is insufficient.

Test (not bad for a VM):

[dc-user@ech-10-157-158-249 test]$ for i in {1..10}; do ./precision; sleep 1; done
timer granularity (ns): 69.514 +/- 3.91384
timer granularity (ns): 47.406 +/- 1.22714
timer granularity (ns): 48.027 +/- 1.07027
timer granularity (ns): 38.501 +/- 0.378576
timer granularity (ns): 49.323 +/- 0.501975
timer granularity (ns): 52.987 +/- 0.997222
timer granularity (ns): 40.877 +/- 0.705936
timer granularity (ns): 45.405 +/- 1.26509
timer granularity (ns): 44.501 +/- 0.635019
timer granularity (ns): 93.439 +/- 90.7611