Why cpu time of FC-MRV of nqueen csp fluctuates

[ThuriyaThwin]

@DonatoMeoli @norvig Please explain me why CPU time of FC-MRV algorithm with N-Queen CSP does not scale with board size. Here are the result of average run benchmark on i7 ram 8gb.

N=4 Time=0.695
N=8 Time=0.785
N=12 Time=0.895
N=16 Time=0.077
N=20 Time=0.092
N=28 Time=0.0795.

In this result, time falls down when N=16 and N=28. Could you give me the reason?

[dmeoli]

This is expected measurement variance, not a bug. Three factors make FC+MRV wall-clock time non-monotonic in N: (1) MRV ties are broken randomly, so different runs explore different search trees; (2) backtracking cost for n-queens is inherently non-monotonic — some board sizes are 'easier' instances than larger ones; (3) at sub-second timings, warm-up/GC/measurement noise dominates. For a stable comparison, take the median over many runs and/or count constraint checks instead of wall-clock time (see #1112). Closing as answered.