hardware-counters

Baseline ladder: what each layer of measurement actually buys
==============================================================================
191 merged configurations, 8 applications: comd, gromacs, hpcg, hpl, lulesh, minife, openfoam, stream
f_peak = 2.25e+09 Hz (ARCHER2 EPYC 7742)
Protocol: leave-one-application-out, error = max(pred/act, act/pred)

log-log correlation r(PAPI_TOT_CYC, runtime) = 0.9929 (r^2 = 0.9858)
eta = t_analytic/runtime: median 0.807, IQR 0.681-0.839, sd(log10 eta) 0.128

--- the ladder ---
rung                             measurement cost         median      p90    d.med    d.p90          p
0  median training runtime       nothing                   7.588   29.191      nan      nan        nan
1  configuration only (RF)       free metadata             7.826   41.944   -0.238  -12.753      0.034
2  t_analytic, eta = 1           1 counter, on target      1.238    1.888    6.588   40.056    1.2e-31
3  t_analytic / constant eta     1 counter, on target      1.065    1.532    0.173    0.356    4.1e-32
4  t_analytic / RF(counters)     5 counter sets, on target    1.054    1.390    0.011    0.142    0.00022

d.med and d.p90 are the reduction in error contributed by that rung
relative to the rung above it, so positive is an improvement.

--- attribution of the total gain ---
total gain, rung 0 to rung 4: 6.533 error factor
  1  configuration only (RF)       contributes  -0.238  ( -3.6% of the total)
  2  t_analytic, eta = 1           contributes   6.588  (100.8% of the total)
  3  t_analytic / constant eta     contributes   0.173  (  2.6% of the total)
  4  t_analytic / RF(counters)     contributes   0.011  (  0.2% of the total)

Rung 2 is a measurement, not a prediction. Everything from rung 2
downwards requires an instrumented run of the target configuration.

--- per-application median error ---
          0  median training runtime  1  configuration only (RF)  2  t_analytic, eta = 1  3  t_analytic / constant eta  4  t_analytic / RF(counters)
app                                                                                                                                                 
comd                           3.381                       3.309                   1.210                         1.046                         1.026
gromacs                       46.586                      50.688                   1.185                         1.051                         1.007
hpcg                          13.332                      20.714                   1.229                         1.048                         1.117
hpl                            4.637                       3.702                   1.230                         1.078                         1.035
lulesh                         1.431                       1.552                   1.443                         1.175                         1.048
minife                        12.096                      22.007                   1.667                         1.372                         1.061
openfoam                       3.588                       3.351                   1.172                         1.068                         1.075
stream                         1.853                       2.112                   1.202                         1.051                         1.095