void
save_minimum()
{
if (results->N == 0) {
save_n(1);
settime(0);
} else {
save_n(results->v[results->N - 1].n);
settime(results->v[results->N - 1].u);
}
}
void
save_minimum()
{
if (results.N == 0) {
save_n(1);
settime(0);
} else {
save_n(results.n[results.N - 1]);
settime(results.u[results.N - 1]);
}
}
void
save_median()
{
int i = results.N / 2;
uint64 u, n;
if (results.N == 0) {
n = 1;
u = 0;
} else if (results.N % 2) {
n = results.n[i];
u = results.u[i];
} else {
n = (results.n[i] + results.n[i-1]) / 2;
u = (results.u[i] + results.u[i-1]) / 2;
}
save_n(n);
settime(u);
}
void
save_median()
{
int i = results->N / 2;
uint64 u, n;
if (results->N == 0) {
n = 1;
u = 0;
} else if (results->N % 2) {
n = results->v[i].n;
u = results->v[i].u;
} else {
n = (results->v[i].n + results->v[i-1].n) / 2;
u = (results->v[i].u + results->v[i-1].u) / 2;
}
#ifdef _DEBUG
fprintf(stderr, "save_median: N=%d, n=%lu, u=%lu\n", results->N, (unsigned long)n, (unsigned long)u);
#endif /* _DEBUG */
save_n(n); settime(u);
}
void
loads(size_t len, size_t range, size_t stride,
int parallel, int warmup, int repetitions)
{
double result;
size_t count;
struct mem_state state;
if (range < stride) return;
state.width = 1;
state.len = range;
state.maxlen = len;
state.line = stride;
state.pagesize = getpagesize();
count = 100 * (state.len / (state.line * 100) + 1);
#if 0
(*fpInit)(0, &state);
fprintf(stderr, "loads: after init\n");
(*benchmark_loads)(2, &state);
fprintf(stderr, "loads: after benchmark\n");
mem_cleanup(0, &state);
fprintf(stderr, "loads: after cleanup\n");
settime(1);
save_n(1);
#else
/*
* Now walk them and time it.
*/
benchmp(fpInit, benchmark_loads, mem_cleanup,
100000, parallel, warmup, repetitions, &state);
#endif
/* We want to get to nanoseconds / load. */
save_minimum();
result = (1000. * (double)gettime()) / (double)(count * get_n());
fprintf(stderr, "%.5f %.3f\n", range / (1024. * 1024.), result);
}
iter_t
benchmp_interval(void* _state)
{
char c;
iter_t iterations;
double result;
fd_set fds;
struct timeval timeout;
benchmp_child_state* state = (benchmp_child_state*)_state;
iterations = (state->state == timing_interval ? state->iterations : state->iterations_batch);
if (state->need_warmup) {
/* remove spurious compilation warning */
result = state->enough;
} else {
result = stop(0,0);
if (state->cleanup) {
if (benchmp_sigchld_handler == SIG_DFL)
signal(SIGCHLD, SIG_DFL);
(*state->cleanup)(iterations, state->cookie);
}
save_n(state->iterations);
result -= t_overhead() + get_n() * l_overhead();
settime(result >= 0. ? (uint64)result : 0.);
}
/* if the parent died, then give up */
if (getppid() == 1 && state->cleanup) {
if (benchmp_sigchld_handler == SIG_DFL)
signal(SIGCHLD, SIG_DFL);
(*state->cleanup)(0, state->cookie);
exit(0);
}
timeout.tv_sec = 0;
timeout.tv_usec = 0;
FD_ZERO(&fds);
switch (state->state) {
case warmup:
iterations = state->iterations_batch;
FD_SET(state->start_signal, &fds);
select(state->start_signal+1, &fds, NULL,
NULL, &timeout);
if (FD_ISSET(state->start_signal, &fds)) {
state->state = timing_interval;
read(state->start_signal, &c, sizeof(char));
iterations = state->iterations;
}
if (state->need_warmup) {
state->need_warmup = 0;
/* send 'ready' */
write(state->response, &c, sizeof(char));
}
break;
case timing_interval:
iterations = state->iterations;
if (state->parallel > 1 || result > 0.95 * state->enough) {
insertsort(gettime(), get_n(), get_results());
state->i++;
/* we completed all the experiments, return results */
if (state->i >= state->repetitions) {
state->state = cooldown;
}
}
if (state->parallel == 1
&& (result < 0.99 * state->enough || result > 1.2 * state->enough)) {
if (result > 150.) {
double tmp = iterations / result;
tmp *= 1.1 * state->enough;
iterations = (iter_t)(tmp + 1);
} else {
iterations <<= 3;
if (iterations > 1<<27
|| (result < 0. && iterations > 1<<20)) {
state->state = cooldown;
}
}
}
state->iterations = iterations;
if (state->state == cooldown) {
/* send 'done' */
write(state->response, (void*)&c, sizeof(char));
iterations = state->iterations_batch;
}
break;
case cooldown:
iterations = state->iterations_batch;
FD_SET(state->result_signal, &fds);
select(state->result_signal+1, &fds, NULL, NULL, &timeout);
if (FD_ISSET(state->result_signal, &fds)) {
/*
* At this point all children have stopped their
* measurement loops, so we can block waiting for
* the parent to tell us to send our results back.
* From this point on, we will do no more "work".
*/
read(state->result_signal, (void*)&c, sizeof(char));
write(state->response, (void*)get_results(), state->r_size);
if (state->cleanup) {
if (benchmp_sigchld_handler == SIG_DFL)
signal(SIGCHLD, SIG_DFL);
(*state->cleanup)(0, state->cookie);
}
/* Now wait for signal to exit */
read(state->exit_signal, (void*)&c, sizeof(char));
exit(0);
}
};
if (state->initialize) {
(*state->initialize)(iterations, state->cookie);
}
start(0);
return (iterations);
}
void
benchmp(benchmp_f initialize,
benchmp_f benchmark,
benchmp_f cleanup,
int enough,
int parallel,
int warmup,
int repetitions,
void* cookie)
{
iter_t iterations = 1;
long i;
pid_t *pids = NULL;
int response[2];
int start_signal[2];
int result_signal[2];
int exit_signal[2];
#ifdef _DEBUG
fprintf(stderr, "benchmp(%p, %p, %p, %d, %d, %d, %d, %p): entering\n", initialize, benchmark, cleanup, enough, parallel, warmup, repetitions, cookie);
#endif
enough = get_enough(enough);
#ifdef _DEBUG
fprintf(stderr, "\tenough=%d\n", enough);
#endif
if (repetitions < 0)
repetitions = (1 < parallel || 1000000 <= enough ? 1 : TRIES);
/* initialize results */
settime(0);
save_n(1);
if (parallel > 1) {
/* Compute the baseline performance */
benchmp(initialize, benchmark, cleanup,
enough, 1, warmup, repetitions, cookie);
/* if we can't even do a single job, then give up */
if (gettime() == 0)
return;
/* calculate iterations for 1sec runtime */
iterations = get_n();
if (enough < SHORT) {
double tmp = (double)SHORT * (double)get_n();
tmp /= (double)gettime();
iterations = (iter_t)tmp + 1;
}
settime(0);
save_n(1);
}
/* Create the necessary pipes for control */
if (pipe(response) < 0
|| pipe(start_signal) < 0
|| pipe(result_signal) < 0
|| pipe(exit_signal) < 0) {
#ifdef _DEBUG
fprintf(stderr, "BENCHMP: Could not create control pipes\n");
#endif /* _DEBUG */
return;
}
/* fork the necessary children */
benchmp_sigchld_received = 0;
benchmp_sigterm_received = 0;
benchmp_sigterm_handler = signal(SIGTERM, benchmp_sigterm);
benchmp_sigchld_handler = signal(SIGCHLD, benchmp_sigchld);
pids = (pid_t*)malloc(parallel * sizeof(pid_t));
if (!pids) return;
bzero((void*)pids, parallel * sizeof(pid_t));
for (i = 0; i < parallel; ++i) {
if (benchmp_sigterm_received)
goto error_exit;
#ifdef _DEBUG
fprintf(stderr, "benchmp(%p, %p, %p, %d, %d, %d, %d, %p): creating child %d\n", initialize, benchmark, cleanup, enough, parallel, warmup, repetitions, cookie, i);
#endif
switch(pids[i] = fork()) {
case -1:
/* could not open enough children! */
#ifdef _DEBUG
fprintf(stderr, "BENCHMP: fork() failed!\n");
#endif /* _DEBUG */
goto error_exit;
case 0:
/* If child */
close(response[0]);
close(start_signal[1]);
close(result_signal[1]);
close(exit_signal[1]);
handle_scheduler(i, 0, 0);
benchmp_child(initialize,
benchmark,
cleanup,
i,
response[1],
start_signal[0],
result_signal[0],
exit_signal[0],
enough,
iterations,
parallel,
repetitions,
cookie
);
exit(0);
default:
break;
}
}
close(response[1]);
close(start_signal[0]);
close(result_signal[0]);
close(exit_signal[0]);
benchmp_parent(response[0],
start_signal[1],
result_signal[1],
exit_signal[1],
pids,
parallel,
iterations,
warmup,
repetitions,
enough
);
goto cleanup_exit;
error_exit:
/* give the children a chance to clean up gracefully */
signal(SIGCHLD, SIG_DFL);
while (--i >= 0) {
kill(pids[i], SIGTERM);
waitpid(pids[i], NULL, 0);
}
cleanup_exit:
/*
* Clean up and kill all children
*
* NOTE: the children themselves SHOULD exit, and
* Killing them could prevent them from
* cleanup up subprocesses, etc... So, we only
* want to kill child processes when it appears
* that they will not die of their own accord.
* We wait twice the timing interval plus two seconds
* for children to die. If they haven't died by
* that time, then we start killing them.
*/
benchmp_sigalrm_timeout = (int)((2 * enough)/1000000) + 2;
if (benchmp_sigalrm_timeout < 5)
benchmp_sigalrm_timeout = 5;
signal(SIGCHLD, SIG_DFL);
while (i-- > 0) {
/* wait timeout seconds for child to die, then kill it */
benchmp_sigalrm_pid = pids[i];
benchmp_sigalrm_handler = signal(SIGALRM, benchmp_sigalrm);
alarm(benchmp_sigalrm_timeout);
waitpid(pids[i], NULL, 0);
alarm(0);
signal(SIGALRM, benchmp_sigalrm_handler);
}
if (pids) free(pids);
#ifdef _DEBUG
fprintf(stderr, "benchmp(0x%x, 0x%x, 0x%x, %d, %d, 0x%x): exiting\n", (unsigned int)initialize, (unsigned int)benchmark, (unsigned int)cleanup, enough, parallel, (unsigned int)cookie);
#endif
}