std::set fast and slow, what is going on?

I've come up across a strange behaviour of std::set.

Here is the code:

#include <cstdio>
#include <windows.h>
#include <stdlib.h>

#include <vector>
#include <set>

using namespace std;

int main(int argc, char *argv[])
{
    set<int> b[100];

    for (int o=0; o<10; o++)
    {
        int tt = GetTickCount();

        for (int i=0; i<5000000; i++)
        {
            b[o].insert(i);
        }

        tt = GetTickCount() - tt;

        b[o].clear();

        printf("%d\n", tt);
    }

    return 0;
}

I'm running on Windows XP.

Here is the interesting part: this first printed time is about 3500 ms, while all next are over 9000 ms! Why is that happening?

Oh, and this only happens on release version (-O2 optimization).

It doesn't happen on Linux (after changing code to compile there).

One more thing: when I run it while profiling with Intel VTune it always takes about 3000 ms, so it's the way it should be.

UPDATE: Here is some new code:

#include <cstdio>
#include <windows.h>
#include <stdlib.h>

int main(int argc, char *argv[])
{
const int count = 10000000;
int **a = new int*[count];

for (int o=0; o<10; o++)
{
    int ttt = GetTickCount();

    for (int i=0; i<count; i++)
    {
        a[i] = new int;

        *a[i] = i;
    }

    int ttt2 = GetTickCount();

    for (int i=0; i<count; i++)
    {
        int r1 = rand() * 10000 + rand();
        int r2 = rand() * 10000 + rand();
        r1 = r1%count;
        r2 = r2%count;

        int *e = a[r1];
        a[r1] = a[r2];
        a[r2] = e;
    }

    int ttt3 = GetTickCount();

    for (int i=0; i<count; i++)
    {
        delete a[i];
    }

    int ttt4 = GetTickCount();

    printf("%d %d\n", ttt2-ttt, ttt4-ttt3);
}

return 0;
}

This is the same problem. What happens is I allocate many many small objects and then delete the