08-Feb-2013 19:12, Jacob Carlborg пишет:
On 2013-02-08 16:08, Brian Schott wrote:
http://hackerpilot.github.com/experimental/std_lexer/images/times3.png
Your suggestions seem to have worked. We're below 20ms on my machine for
the datetime module.
DMD is still consistently faster, :(
Keep in mind the D runtime start-up cost. In the end on small files DMD
always wins because of slim C-runtime.
To estimate runtime startup lag I've compared run times of
int main(){ return 0; } in C (gcc):
------------------------
Total time (ms): 31637.7
Repetitions : 35000
Sample mode : 0.8 (23649 ocurrences)
Median time : 0.873
Avg time : 0.903935
Std dev. : 0.204107
Minimum : 0.552
Maximum : 9.057
95% conf.int. : [0.503892, 1.30398] e = 0.400043
99% conf.int. : [0.378189, 1.42968] e = 0.525746
EstimatedAvg95%: [0.901796, 0.906073] e = 0.00213832
EstimatedAvg99%: [0.901125, 0.906745] e = 0.00281023
void main(){} in D (ldc):
------------------------
Total time (ms): 15429.4
Repetitions : 7000
Sample mode : 2.1 (1785 ocurrences)
Median time : 2.128
Avg time : 2.2042
Std dev. : 0.466834
Minimum : 1.286
Maximum : 19.933
95% conf.int. : [1.28922, 3.11918] e = 0.914978
99% conf.int. : [1.00171, 3.40668] e = 1.20248
EstimatedAvg95%: [2.19326, 2.21514] e = 0.0109361
EstimatedAvg99%: [2.18983, 2.21857] e = 0.0143724
dmitry@dmitry-VirtualBox ~ $
I think that the mode could serve as an indicator of the most probable
outcome.
For me it's around 0.8 vs 2.1 ms. D runtime rides on top of C one + plus
the same OS costs to launch a process, so from now on let's keep in mind
these extra ~ 1.3 ms of bias in favor of C-based lexer.
--
Dmitry Olshansky
