int
main(int argc,char *argv[])
{
gaspi_rank_t rank,tnc;
gaspi_float vers;
gaspi_config_t gconf;
char mtype[16];
int i;
mcycles_t t0, t1;
mcycles_t delta[1024];
gaspi_float cpu_freq;
gaspi_config_get(&gconf);
gconf.mtu = 4096;
gconf.queue_num = 1;
gaspi_config_set(gconf);
GPI2_ASSERT(gaspi_proc_init(GASPI_BLOCK));
GPI2_ASSERT( gaspi_version(&vers) );
GPI2_ASSERT( gaspi_proc_rank(&rank) );
GPI2_ASSERT( gaspi_proc_num(&tnc) );
GPI2_ASSERT( gaspi_machine_type(mtype) );
GPI2_ASSERT( gaspi_cpu_frequency(&cpu_freq));
if ( 0 == rank)
printf("my rank: %d tnc: %d (vers: %.2f) machine:%s\n",rank,tnc,vers,mtype);
GPI2_ASSERT(gaspi_barrier(GASPI_GROUP_ALL,GASPI_BLOCK));
if(0 == rank)
printf("cpu freq: %.2f\n",cpu_freq);
//benchmark
for(i = 0; i < 1000; i++)
{
t0 = get_mcycles();
GPI2_ASSERT(gaspi_barrier(GASPI_GROUP_ALL,GASPI_BLOCK));
t1 = get_mcycles();
delta[i] = (t1 - t0);
}
if(0 == rank)
{
qsort(delta,1000,sizeof *delta,mcycles_compare);
const double div = 1.0 / cpu_freq;
const double ts = (double)delta[500] * div;
printf("time: %f usec\n",ts);
}
fflush(stdout);
GPI2_ASSERT(gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
GPI2_ASSERT(gaspi_proc_term(GASPI_BLOCK));
return 0;
}
int
main(int argc,char *argv[])
{
gaspi_rank_t rank,tnc;
gaspi_return_t ret;
gaspi_float vers;
gaspi_config_t gconf;
char mtype[16];
int commSize=1,cnt=0;
gaspi_number_t queueSize;
void *pret;
int i,l;
mcycles_t t0,t1,dt;
mcycles_t stamp[1024], stamp2[1024], delta[1024];
int amount_work = 1000;
gaspi_float cpu_freq;
gaspi_config_get(&gconf);
gconf.mtu = 4096;
gconf.queue_num = 1;
gaspi_config_set(gconf);
GPI2_ASSERT(gaspi_proc_init(GASPI_BLOCK));
GPI2_ASSERT( gaspi_version(&vers) );
GPI2_ASSERT( gaspi_proc_rank(&rank) );
GPI2_ASSERT( gaspi_proc_num(&tnc) );
GPI2_ASSERT( gaspi_machine_type(mtype) );
GPI2_ASSERT( gaspi_cpu_frequency(&cpu_freq));
if(0 == rank)
{
printf("cpu freq: %.2f\n", cpu_freq);
printf("my rank: %d tnc: %d (vers: %.2f) machine:%s\n",rank, tnc, vers, mtype);
}
GPI2_ASSERT(gaspi_barrier(GASPI_GROUP_ALL,GASPI_BLOCK));
//benchmark
for(i = 0; i < 1000; i++)
{
t0=t1=dt=0;
do
{
t0 = get_mcycles();
ret = gaspi_barrier(GASPI_GROUP_ALL,GASPI_TEST);
t1 = get_mcycles();
dt += (t1-t0);
usleep(amount_work); //useful work here..
}while(ret!=GASPI_SUCCESS);
delta[i]=dt;
}
if(0 == rank)
{
qsort(delta,1000,sizeof *delta,mcycles_compare);
const double div = 1.0 / cpu_freq;
const double ts = (double)delta[500] * div;
printf("time: %f usec\n",ts);
}
GPI2_ASSERT(gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
GPI2_ASSERT(gaspi_proc_term(GASPI_BLOCK));
return 0;
}
int
main (int argc, char *argv[])
{
int i, j, t;
gaspi_rank_t myrank;
char *ptr0;
//on numa architectures you have to map this process to the numa node where nic is installed
if (start_bench (2) != 0)
{
printf ("Initialization failed\n");
exit (-1);
}
// BENCH //
gaspi_proc_rank (&myrank);
if (gaspi_segment_ptr (0, (void **) &ptr0) != GASPI_SUCCESS)
{
printf ("gaspi_segment_ptr failed !\n");
exit (-1);
}
gaspi_float cpu_freq;
gaspi_cpu_frequency(&cpu_freq);
if (myrank < 2)
{
if(myrank == 0)
{
printf("-----------------------------------\n");
printf ("%12s\t%5s\n", "Bytes", "Lat(usecs)");
printf("-----------------------------------\n");
}
int bytes = 2;
volatile char *postBuf = (volatile char *) ptr0;
for (i = 1; i < 24; i++)
{
volatile char *pollBuf = (volatile char *) (ptr0 + ( 2 * bytes -1 ));
int rcnt = 0;
int cnt = 0;
gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK);
for (j = 0; j < ITERATIONS; j++)
{
if (rcnt < ITERATIONS && !(cnt < 1 && myrank == 1))
{
rcnt++;
while (*pollBuf != (char) rcnt)
{
#ifdef MIC
_mm_delay_32(32);
#else
_mm_pause();
#endif
}
}
stamp[j] = get_mcycles ();
postBuf[bytes - 1] = (char) ++cnt;
gaspi_write (0, 0, myrank ^ 0x1,
0, bytes, bytes,
0, GASPI_BLOCK);
gaspi_wait (0, GASPI_BLOCK);
}
for (t = 0; t < (ITERATIONS - 1); t++)
delta[t] = stamp[t + 1] - stamp[t];
qsort (delta, (ITERATIONS - 1), sizeof *delta, mcycles_compare);
const double div = 1.0 / cpu_freq;
const double ts = (double) delta[ITERATIONS / 2] * div * 0.5;
if(myrank == 0)
printf ("%12d\t%4.2f\n", bytes, ts);
bytes <<= 1;
}
}
end_bench ();
return 0;
}
