efficiency question in TBB

[anine]

When I am testing efficiency of the TBB programs. Carry out the same TBB programs 10,000 times.
The first several efficiency is always worse. what kinds of reason could cause this?

test method :
task_scheduler_init init;

for(int i = 0; i < testcount; i++){
tick_count t0 = tick_count::now();
parallel_for(blocked_range(0, size), ApplyFoo(source,dest,minus), auto_partitioner()); //for auto
tick_count t1 = tick_count::now();
sum = sum + (t1-t0).seconds();
fprintf(fp, " %d : %.12f\n",i, (t1-t0).seconds());
} // end for

where
class ApplyFoo : minus = source - dest;
testcount = 10000 times
size = 640*480

efficency:

1 time : 0.011139 <-------
2 time : 0.001736
3 time : 0.001742
4 time : 0.001273
5 time : 0.001272
.....
average : 0.001284

Thank you.

[robert_jay_gould]

Quoting - anine

When I am testing efficiency of the TBB programs. Carry out the same TBB programs 10,000 times.
The first several efficiency is always worse. what kinds of reason could cause thisThe first time your loading up your cache

This is because during the first attempt your program is loading instructions, cleaning up, getting memory and preparing the cache, and in the following cases everything already in place and its getting optimized by the cache, so it works smoothly.

When you need to compare results of two operations, a good option is to mixup the tests, so they thwart any extra optimizations, that you won't see in an actually complicated program.