Re: [Caml-list] Linear Scan Register Allocator for ocamlopt/ocamlnat

[Benedikt Meurer <benedikt.meurer@googlemail.com>]

On Aug 25, 2011, at 12:21 , Gerd Stolpmann wrote:

>> Reran the benchmark with 32bit ocaml on the MBP (Early 2011), results are available at:
>> 
>> http://ps.informatik.uni-siegen.de/~meurer/tmp/linscan-i7-i386-timings.pdf
> 
> Hard to interpret. I have the impression that the compile times for
> graph coloring on i386 are lower than on amd64 - has maybe to do with
> the number of registers - but those for linear scanning are about the
> same. (I'm comparing here with the amd64 numbers taken on the same
> machine.)
> 
> The graphs for the runtimes are on a logarithmic scale. Do you also have
> raw numbers? 

Of course.

These are the raw numbers for the i7 (i386):

command                  opt_gc     opt_ls       opt_gc/ls
----------------------------------------------------------
binarytree 16             1.096      1.102           0.995
fasta 100                 0.008      0.008           1.000
fasta 250000              0.044      0.044           1.000
fasta 25000000            2.767      2.737           1.011
mandelbrot 1000           0.224      0.224           1.000
mandelbrot 4000           3.583      3.570           1.004
meteor 2098               0.578      0.586           0.986
nbody 3000000             1.142      1.168           0.978
spectral 3000             4.963      5.243           0.947
almabench                 4.665      4.687           0.995
almabench.unsafe          4.629      4.702           0.984
bdd                       0.279      0.285           0.979
boyer                     0.566      0.572           0.990
fft                       0.421      0.422           0.998
nucleic                   0.499      0.513           0.973
quicksort                 0.150      0.149           1.007
quicksort.unsafe          0.121      0.121           1.000
sorts                     1.889      1.947           0.970

And the raw numbers for the P4:

command                  opt_gc     opt_ls       opt_gc/ls
----------------------------------------------------------
binarytree 16             2.796      2.780           1.006
fasta 100                 0.024      0.024           1.000
fasta 250000              0.404      0.408           0.990
fasta 25000000           33.350     33.402           0.998
mandelbrot 1000           0.892      0.884           1.009
mandelbrot 4000          14.220     14.040           1.013
meteor 2098               1.300      1.312           0.991
nbody 3000000             4.524      4.840           0.935
spectral 3000            51.979     53.919           0.964
almabench                17.261     17.721           0.974
almabench.unsafe         17.317     18.441           0.939
bdd                       1.256      1.248           1.006
boyer                     1.668      1.672           0.998
fft                       3.064      3.080           0.995
nucleic                   2.188      2.268           0.965
quicksort                 0.340      0.340           1.000
quicksort.unsafe          0.272      0.272           1.000
sorts                     6.164      6.180           0.997

> Gerd

Benedikt