hardware-counters
counter_value_summary.txt
Do the hardware counters earn their keep?
==============================================================================
191 merged configurations, 8 applications
Leave-one-application-out; error = max(pred/act, act/pred)
constant (no features) median 1.0654 p90 1.5324
config only (free) [RF] n_feat= 3 median 1.1040 p90 1.7210
config only (free) [Ridge] n_feat= 3 median 1.1456 p90 1.5485
+ analytic cycles (1 counter) [RF] n_feat= 4 median 1.1075 p90 1.3981
+ analytic cycles (1 counter) [Ridge] n_feat= 4 median 1.1252 p90 1.3480
+ instruction ratios (1 set) [RF] n_feat= 8 median 1.0796 p90 1.3448
+ instruction ratios (1 set) [Ridge] n_feat= 8 median 1.1556 p90 1.3663
+ all counters (5 sets) [RF] n_feat=21 median 1.0554 p90 1.3910
+ all counters (5 sets) [Ridge] n_feat=21 median 1.2926 p90 2.3679
--- Does adding counters beat config-only? (paired Wilcoxon) ---
+ analytic cycles (1 counter) [RF] delta=-0.0035 wins 118/191 p=0.0001899
+ instruction ratios (1 set) [RF] delta=+0.0244 wins 131/191 p=3.919e-09
+ all counters (5 sets) [RF] delta=+0.0486 wins 148/191 p=8.307e-14
--- vs the no-feature constant ---
config only (free) [RF] p=0.001425 (WORSE)
config only (free) [Ridge] p=0.009181 (WORSE)
+ analytic cycles (1 counter) [RF] p=0.9666 (WORSE)
+ analytic cycles (1 counter) [Ridge] p=0.1284 (WORSE)
+ instruction ratios (1 set) [RF] p=0.123 (WORSE)
+ instruction ratios (1 set) [Ridge] p=0.01305 (WORSE)
+ all counters (5 sets) [RF] p=0.000247 (better)
+ all counters (5 sets) [Ridge] p=3.961e-12 (WORSE)
--- per-application median, RF ---
constant (no features) config only (free) [RF] + analytic cycles (1 counter) [RF] + instruction ratios (1 set) [RF] + all counters (5 sets) [RF]
app
comd 1.046 1.090 1.121 1.057 1.026
gromacs 1.051 1.103 1.014 1.009 1.006
hpcg 1.048 1.147 1.102 1.118 1.117
hpl 1.078 1.092 1.054 1.061 1.043
lulesh 1.175 1.113 1.177 1.071 1.049
minife 1.372 1.273 1.242 1.150 1.067
openfoam 1.068 1.102 1.042 1.080 1.073
stream 1.051 1.069 1.190 1.115 1.099