Re: Teaching new tricks to an old dog (C++ -->Ada)

[CTips <ctips@bestweb.net>]

Georg Bauhaus wrote:


> Done. You might be disappointed by the results. In short:
> < 1%, < 3%; and < 8% for a different program, really.

I'm not disappointed at all, in fact the numbers are way worse than I 
expected, given that you used:

>    VERY_LARGE_NUMBER: constant := 100_000_000;

You're using 800MB of memory, and unless you have at least 1GB, you're 
probably swapping in and out. Also you're probably trashing your TLB. 
You're *definitely* thrashing all levels of cache.

The other point is that the branch prediction table is primed (i.e. it 
will always go to the dont-fail case). In a real program, the 
branch-predictor would mispredict more often.

If, after all that, the effects of the check are at all visible, it 
means that the checks are probably significant.

Try again with smaller numbers that yield arrays 1/8 the size of your L1 
and 1/4 the size of your L2. E.g., If you've got a 64 KB  L1 and a 1MB 
L2, make VERY_LARGE_NUMBER 1000 and 64000. Call this benchmark 
repeatedly till you get to a total run-time of about 1sec.

I checked; on my machine, if I run the C code below with 100k entries vs 
50k entries the performance is 4.7s vs. 0.7s ~ 6.5x slowdown vs 2x 
change in size [I'm using time under cygwin on a Xeon 3.06GHz 
dual-processor system]. This indicates that I have a 512KB L2-cache.

Now, I added the bounds check. At 50k, there is a 27% slowdown, at 100k 
there is about a 1% slowdown.

You have to be *very* careful that what you're measuring isn't hidden 
behind some other effect.

----------------------------------------------------------

#define N 50000

void
setup_perm(
         int     perm[]
         )
{
   int   i;
   for( i = 0; i < N; i++ ) {
     perm[i] = i;
   }
}

void
do_perm(
         int     perm[],
         int     val[]
         )
{
   int   i;
   for( i = 0; i < N; i++ ) {
     /*
     if( perm[i] >= N ) {
       abort();
     }
     */
     val[perm[i]] = i;
   }
}

int     perm[N];
int     val[N];

int
main(void)
{
   int   i;

   setup_perm(perm);
   for( i = 0; i < 10000; i++ ) {
     do_perm(perm, val);
   }
   return 0;
}