hardware-counters

Weak scaling: is Cirrus intrinsically hard, or was it the problem size?
==============================================================================
--- dataset characteristics ---
  dataset           n   med rt    <1s  sd log10(eta)  med eta
  -----------------------------------------------------------
  archer2         191     5.16  24.6%         0.1283    0.807
  cirrus-fixed    191     1.05  47.6%         0.2852    0.899
  cirrus-weak     162     3.91  16.7%         0.0719    0.946

--- leave-one-application-out, within each dataset ---
  archer2        intersection  best=gbq  1.0547   const=1.0654   p=0.0001183  BEATS constant
  archer2        full          best=rf   1.0496   const=1.0654   p=3.555e-05  BEATS constant
  cirrus-fixed   intersection  best=rf   1.1863   const=1.1663   p=0.01339  loses to constant
  cirrus-fixed   full          best=rf   1.2804   const=1.1663   p=0.9105  loses to constant
  cirrus-weak    intersection  best=gbq  1.0565   const=1.0617   p=0.3267  BEATS constant
  cirrus-weak    full          best=rf   1.1042   const=1.0617   p=0.2276  loses to constant

--- all models ---
model                       const     gbq     mlp      rf
dataset      features                                    
archer2      full          1.0654  1.0596  1.0752  1.0496
             intersection  1.0654  1.0547  1.0760  1.0549
cirrus-fixed full          1.1663  1.2819  1.3404  1.2804
             intersection  1.1663  1.2146  1.2892  1.1863
cirrus-weak  full          1.0617  1.1066  1.1248  1.1042
             intersection  1.0617  1.0565  1.0807  1.0603

--- cross-platform transfer using WEAK-scaled Cirrus ---
  archer2 -> cirrus-weak:  const 1.1902   rf 1.1822   gbq 1.1764   mlp 1.2107
  cirrus-weak -> archer2:  const 1.1719   rf 1.1433   gbq 1.1473   mlp 1.1392