static pj_highprec_t elapsed_usec( const pj_timestamp *start,
const pj_timestamp *stop )
{
pj_timestamp ts_freq;
pj_highprec_t freq, elapsed;
if (pj_get_timestamp_freq(&ts_freq) != PJ_SUCCESS)
return 0;
/* Convert frequency timestamp */
#if defined(PJ_HAS_INT64) && PJ_HAS_INT64!=0
freq = u64tohighprec(ts_freq.u64);
#else
freq = ts_freq.u32.hi;
pj_highprec_mul(freq, U32MAX);
freq += ts_freq.u32.lo;
#endif
/* Avoid division by zero. */
if (freq == 0) freq = 1;
/* Get elapsed time in cycles. */
elapsed = get_elapsed(start, stop);
/* usec = elapsed * USEC / freq */
pj_highprec_mul(elapsed, USEC);
pj_highprec_div(elapsed, freq);
return elapsed;
}
int echo_srv_common_loop(pj_atomic_t *bytes_counter)
{
pj_highprec_t last_received, avg_bw, highest_bw;
pj_time_val last_print;
unsigned count;
const char *ioqueue_name;
last_received = 0;
pj_gettimeofday(&last_print);
avg_bw = highest_bw = 0;
count = 0;
ioqueue_name = pj_ioqueue_name();
for (;;) {
pj_highprec_t received, cur_received, bw;
unsigned msec;
pj_time_val now, duration;
pj_thread_sleep(1000);
received = cur_received = pj_atomic_get(bytes_counter);
cur_received = cur_received - last_received;
pj_gettimeofday(&now);
duration = now;
PJ_TIME_VAL_SUB(duration, last_print);
msec = PJ_TIME_VAL_MSEC(duration);
bw = cur_received;
pj_highprec_mul(bw, 1000);
pj_highprec_div(bw, msec);
last_print = now;
last_received = received;
avg_bw = avg_bw + bw;
count++;
PJ_LOG(3,("", "%s UDP (%d threads): %u KB/s (avg=%u KB/s) %s",
ioqueue_name,
ECHO_SERVER_MAX_THREADS,
(unsigned)(bw / 1000),
(unsigned)(avg_bw / count / 1000),
(count==20 ? "<ses avg>" : "")));
if (count==20) {
if (avg_bw/count > highest_bw)
highest_bw = avg_bw/count;
count = 0;
avg_bw = 0;
PJ_LOG(3,("", "Highest average bandwidth=%u KB/s",
(unsigned)(highest_bw/1000)));
}
}
PJ_UNREACHED(return 0;)
}
static pj_highprec_t get_elapsed( const pj_timestamp *start,
const pj_timestamp *stop )
{
#if defined(PJ_HAS_INT64) && PJ_HAS_INT64!=0
return u64tohighprec(stop->u64 - start->u64);
#else
pj_highprec_t elapsed_hi, elapsed_lo;
elapsed_hi = stop->u32.hi - start->u32.hi;
elapsed_lo = stop->u32.lo - start->u32.lo;
/* elapsed_hi = elapsed_hi * U32MAX */
pj_highprec_mul(elapsed_hi, U32MAX);
return elapsed_hi + elapsed_lo;
#endif
}
PJ_DEF(pj_status_t) pj_thread_sleep(unsigned msec)
{
pj_highprec_t ticks;
pj_thread_t *thread = pj_thread_this();
PJ_ASSERT_RETURN(thread != NULL, PJ_EBUG);
/* Use high precision calculation to make sure we don't
* crop values:
*
* ticks = HZ * msec / 1000
*/
ticks = HZ;
pj_highprec_mul(ticks, msec);
pj_highprec_div(ticks, 1000);
TRACE_((THIS_FILE, "this thread will sleep for %u ticks", ticks));
interruptible_sleep_on_timeout( &thread->queue, ticks);
return PJ_SUCCESS;
}
/* Calculate the bandwidth for the specific test configuration.
* The test is simple:
* - create sockpair_cnt number of producer-consumer socket pair.
* - create thread_cnt number of worker threads.
* - each producer will send buffer_size bytes data as fast and
* as soon as it can.
* - each consumer will read buffer_size bytes of data as fast
* as it could.
* - measure the total bytes received by all consumers during a
* period of time.
*/
static int perform_test(pj_bool_t allow_concur,
int sock_type, const char *type_name,
unsigned thread_cnt, unsigned sockpair_cnt,
pj_size_t buffer_size,
pj_size_t *p_bandwidth)
{
enum { MSEC_DURATION = 5000 };
pj_pool_t *pool;
test_item *items;
pj_thread_t **thread;
pj_ioqueue_t *ioqueue;
pj_status_t rc;
pj_ioqueue_callback ioqueue_callback;
pj_uint32_t total_elapsed_usec, total_received;
pj_highprec_t bandwidth;
pj_timestamp start, stop;
unsigned i;
TRACE_((THIS_FILE, " starting test.."));
ioqueue_callback.on_read_complete = &on_read_complete;
ioqueue_callback.on_write_complete = &on_write_complete;
thread_quit_flag = 0;
pool = pj_pool_create(mem, NULL, 4096, 4096, NULL);
if (!pool)
return -10;
items = (test_item*) pj_pool_alloc(pool, sockpair_cnt*sizeof(test_item));
thread = (pj_thread_t**)
pj_pool_alloc(pool, thread_cnt*sizeof(pj_thread_t*));
TRACE_((THIS_FILE, " creating ioqueue.."));
rc = pj_ioqueue_create(pool, sockpair_cnt*2, &ioqueue);
if (rc != PJ_SUCCESS) {
app_perror("...error: unable to create ioqueue", rc);
return -15;
}
rc = pj_ioqueue_set_default_concurrency(ioqueue, allow_concur);
if (rc != PJ_SUCCESS) {
app_perror("...error: pj_ioqueue_set_default_concurrency()", rc);
return -16;
}
/* Initialize each producer-consumer pair. */
for (i=0; i<sockpair_cnt; ++i) {
pj_ssize_t bytes;
items[i].ioqueue = ioqueue;
items[i].buffer_size = buffer_size;
items[i].outgoing_buffer = (char*) pj_pool_alloc(pool, buffer_size);
items[i].incoming_buffer = (char*) pj_pool_alloc(pool, buffer_size);
items[i].bytes_recv = items[i].bytes_sent = 0;
/* randomize outgoing buffer. */
pj_create_random_string(items[i].outgoing_buffer, buffer_size);
/* Create socket pair. */
TRACE_((THIS_FILE, " calling socketpair.."));
rc = app_socketpair(pj_AF_INET(), sock_type, 0,
&items[i].server_fd, &items[i].client_fd);
if (rc != PJ_SUCCESS) {
app_perror("...error: unable to create socket pair", rc);
return -20;
}
/* Register server socket to ioqueue. */
TRACE_((THIS_FILE, " register(1).."));
rc = pj_ioqueue_register_sock(pool, ioqueue,
items[i].server_fd,
&items[i], &ioqueue_callback,
&items[i].server_key);
if (rc != PJ_SUCCESS) {
app_perror("...error: registering server socket to ioqueue", rc);
return -60;
}
/* Register client socket to ioqueue. */
TRACE_((THIS_FILE, " register(2).."));
rc = pj_ioqueue_register_sock(pool, ioqueue,
items[i].client_fd,
&items[i], &ioqueue_callback,
&items[i].client_key);
if (rc != PJ_SUCCESS) {
app_perror("...error: registering server socket to ioqueue", rc);
return -70;
}
/* Start reading. */
TRACE_((THIS_FILE, " pj_ioqueue_recv.."));
bytes = items[i].buffer_size;
rc = pj_ioqueue_recv(items[i].server_key, &items[i].recv_op,
items[i].incoming_buffer, &bytes,
0);
if (rc != PJ_EPENDING) {
app_perror("...error: pj_ioqueue_recv", rc);
return -73;
}
/* Start writing. */
TRACE_((THIS_FILE, " pj_ioqueue_write.."));
bytes = items[i].buffer_size;
rc = pj_ioqueue_send(items[i].client_key, &items[i].send_op,
items[i].outgoing_buffer, &bytes, 0);
if (rc != PJ_SUCCESS && rc != PJ_EPENDING) {
app_perror("...error: pj_ioqueue_write", rc);
return -76;
}
items[i].has_pending_send = (rc==PJ_EPENDING);
}
/* Create the threads. */
for (i=0; i<thread_cnt; ++i) {
struct thread_arg *arg;
arg = (struct thread_arg*) pj_pool_zalloc(pool, sizeof(*arg));
arg->id = i;
arg->ioqueue = ioqueue;
arg->counter = 0;
rc = pj_thread_create( pool, NULL,
&worker_thread,
arg,
PJ_THREAD_DEFAULT_STACK_SIZE,
PJ_THREAD_SUSPENDED, &thread[i] );
if (rc != PJ_SUCCESS) {
app_perror("...error: unable to create thread", rc);
return -80;
}
}
/* Mark start time. */
rc = pj_get_timestamp(&start);
if (rc != PJ_SUCCESS)
return -90;
/* Start the thread. */
TRACE_((THIS_FILE, " resuming all threads.."));
for (i=0; i<thread_cnt; ++i) {
rc = pj_thread_resume(thread[i]);
if (rc != 0)
return -100;
}
/* Wait for MSEC_DURATION seconds.
* This should be as simple as pj_thread_sleep(MSEC_DURATION) actually,
* but unfortunately it doesn't work when system doesn't employ
* timeslicing for threads.
*/
TRACE_((THIS_FILE, " wait for few seconds.."));
do {
pj_thread_sleep(1);
/* Mark end time. */
rc = pj_get_timestamp(&stop);
if (thread_quit_flag) {
TRACE_((THIS_FILE, " transfer limit reached.."));
break;
}
if (pj_elapsed_usec(&start,&stop)<MSEC_DURATION * 1000) {
TRACE_((THIS_FILE, " time limit reached.."));
break;
}
} while (1);
/* Terminate all threads. */
TRACE_((THIS_FILE, " terminating all threads.."));
thread_quit_flag = 1;
for (i=0; i<thread_cnt; ++i) {
TRACE_((THIS_FILE, " join thread %d..", i));
pj_thread_join(thread[i]);
}
/* Close all sockets. */
TRACE_((THIS_FILE, " closing all sockets.."));
for (i=0; i<sockpair_cnt; ++i) {
pj_ioqueue_unregister(items[i].server_key);
pj_ioqueue_unregister(items[i].client_key);
}
/* Destroy threads */
for (i=0; i<thread_cnt; ++i) {
pj_thread_destroy(thread[i]);
}
/* Destroy ioqueue. */
TRACE_((THIS_FILE, " destroying ioqueue.."));
pj_ioqueue_destroy(ioqueue);
/* Calculate actual time in usec. */
total_elapsed_usec = pj_elapsed_usec(&start, &stop);
/* Calculate total bytes received. */
total_received = 0;
for (i=0; i<sockpair_cnt; ++i) {
total_received = (pj_uint32_t)items[i].bytes_recv;
}
/* bandwidth = total_received*1000/total_elapsed_usec */
bandwidth = total_received;
pj_highprec_mul(bandwidth, 1000);
pj_highprec_div(bandwidth, total_elapsed_usec);
*p_bandwidth = (pj_uint32_t)bandwidth;
PJ_LOG(3,(THIS_FILE, " %.4s %2d %2d %8d KB/s",
type_name, thread_cnt, sockpair_cnt,
*p_bandwidth));
/* Done. */
pj_pool_release(pool);
TRACE_((THIS_FILE, " done.."));
return 0;
}
static int uri_benchmark(unsigned *p_parse, unsigned *p_print, unsigned *p_cmp)
{
unsigned i, loop;
pj_status_t status = PJ_SUCCESS;
pj_timestamp zero;
pj_time_val elapsed;
pj_highprec_t avg_parse, avg_print, avg_cmp, kbytes;
pj_bzero(&var, sizeof(var));
zero.u32.hi = zero.u32.lo = 0;
var.parse_len = var.print_len = var.cmp_len = 0;
var.parse_time.u32.hi = var.parse_time.u32.lo = 0;
var.print_time.u32.hi = var.print_time.u32.lo = 0;
var.cmp_time.u32.hi = var.cmp_time.u32.lo = 0;
for (loop=0; loop<LOOP_COUNT; ++loop) {
for (i=0; i<PJ_ARRAY_SIZE(uri_test_array); ++i) {
pj_pool_t *pool;
pool = pjsip_endpt_create_pool(endpt, "", POOL_SIZE, POOL_SIZE);
status = do_uri_test(pool, &uri_test_array[i]);
pjsip_endpt_release_pool(endpt, pool);
if (status != PJ_SUCCESS) {
PJ_LOG(3,(THIS_FILE, " error %d when testing entry %d",
status, i));
pjsip_endpt_release_pool(endpt, pool);
goto on_return;
}
}
}
kbytes = var.parse_len;
pj_highprec_mod(kbytes, 1000000);
pj_highprec_div(kbytes, 100000);
elapsed = pj_elapsed_time(&zero, &var.parse_time);
avg_parse = pj_elapsed_usec(&zero, &var.parse_time);
pj_highprec_mul(avg_parse, AVERAGE_URL_LEN);
pj_highprec_div(avg_parse, var.parse_len);
if (avg_parse == 0)
avg_parse = 1;
avg_parse = 1000000 / avg_parse;
PJ_LOG(3,(THIS_FILE,
" %u.%u MB of urls parsed in %d.%03ds (avg=%d urls/sec)",
(unsigned)(var.parse_len/1000000), (unsigned)kbytes,
elapsed.sec, elapsed.msec,
(unsigned)avg_parse));
*p_parse = (unsigned)avg_parse;
kbytes = var.print_len;
pj_highprec_mod(kbytes, 1000000);
pj_highprec_div(kbytes, 100000);
elapsed = pj_elapsed_time(&zero, &var.print_time);
avg_print = pj_elapsed_usec(&zero, &var.print_time);
pj_highprec_mul(avg_print, AVERAGE_URL_LEN);
pj_highprec_div(avg_print, var.parse_len);
if (avg_print == 0)
avg_print = 1;
avg_print = 1000000 / avg_print;
PJ_LOG(3,(THIS_FILE,
" %u.%u MB of urls printed in %d.%03ds (avg=%d urls/sec)",
(unsigned)(var.print_len/1000000), (unsigned)kbytes,
elapsed.sec, elapsed.msec,
(unsigned)avg_print));
*p_print = (unsigned)avg_print;
kbytes = var.cmp_len;
pj_highprec_mod(kbytes, 1000000);
pj_highprec_div(kbytes, 100000);
elapsed = pj_elapsed_time(&zero, &var.cmp_time);
avg_cmp = pj_elapsed_usec(&zero, &var.cmp_time);
pj_highprec_mul(avg_cmp, AVERAGE_URL_LEN);
pj_highprec_div(avg_cmp, var.cmp_len);
if (avg_cmp == 0)
avg_cmp = 1;
avg_cmp = 1000000 / avg_cmp;
PJ_LOG(3,(THIS_FILE,
" %u.%u MB of urls compared in %d.%03ds (avg=%d urls/sec)",
(unsigned)(var.cmp_len/1000000), (unsigned)kbytes,
elapsed.sec, elapsed.msec,
(unsigned)avg_cmp));
*p_cmp = (unsigned)avg_cmp;
on_return:
return status;
}
/*
* sock_producer_consumer()
*
* Simple producer-consumer benchmarking. Send loop number of
* buf_size size packets as fast as possible.
*/
static int sock_producer_consumer(int sock_type,
unsigned buf_size,
unsigned loop,
unsigned *p_bandwidth)
{
pj_sock_t consumer, producer;
pj_pool_t *pool;
char *outgoing_buffer, *incoming_buffer;
pj_timestamp start, stop;
unsigned i;
pj_highprec_t elapsed, bandwidth;
pj_size_t total_received;
pj_status_t rc;
/* Create pool. */
pool = pj_pool_create(mem, NULL, 4096, 4096, NULL);
if (!pool)
return -10;
/* Create producer-consumer pair. */
rc = app_socketpair(PJ_AF_INET, sock_type, 0, &consumer, &producer);
if (rc != PJ_SUCCESS) {
app_perror("...error: create socket pair", rc);
return -20;
}
/* Create buffers. */
outgoing_buffer = pj_pool_alloc(pool, buf_size);
incoming_buffer = pj_pool_alloc(pool, buf_size);
/* Start loop. */
pj_get_timestamp(&start);
total_received = 0;
for (i=0; i<loop; ++i) {
pj_ssize_t sent, part_received, received;
pj_time_val delay;
sent = buf_size;
rc = pj_sock_send(producer, outgoing_buffer, &sent, 0);
if (rc != PJ_SUCCESS || sent != (pj_ssize_t)buf_size) {
app_perror("...error: send()", rc);
return -61;
}
/* Repeat recv() until all data is part_received.
* This applies only for non-UDP of course, since for UDP
* we would expect all data to be part_received in one packet.
*/
received = 0;
do {
part_received = buf_size-received;
rc = pj_sock_recv(consumer, incoming_buffer+received,
&part_received, 0);
if (rc != PJ_SUCCESS) {
app_perror("...recv error", rc);
return -70;
}
if (part_received <= 0) {
PJ_LOG(3,("", "...error: socket has closed (part_received=%d)!",
part_received));
return -73;
}
if ((pj_size_t)part_received != buf_size-received) {
if (sock_type != PJ_SOCK_STREAM) {
PJ_LOG(3,("", "...error: expecting %u bytes, got %u bytes",
buf_size-received, part_received));
return -76;
}
}
received += part_received;
} while ((pj_size_t)received < buf_size);
total_received += received;
/* Stop test if it's been runnign for more than 10 secs. */
pj_get_timestamp(&stop);
delay = pj_elapsed_time(&start, &stop);
if (delay.sec > 10)
break;
}
/* Stop timer. */
pj_get_timestamp(&stop);
elapsed = pj_elapsed_usec(&start, &stop);
/* bandwidth = total_received * 1000 / elapsed */
bandwidth = total_received;
pj_highprec_mul(bandwidth, 1000);
pj_highprec_div(bandwidth, elapsed);
*p_bandwidth = (pj_uint32_t)bandwidth;
/* Close sockets. */
pj_sock_close(consumer);
pj_sock_close(producer);
/* Done */
pj_pool_release(pool);
return 0;
}
