app/test-eventdev: optimize producer routine
Checks
Commit Message
From: Pavan Nikhilesh <pbhagavatula@marvell.com>
When using synthetic and timer event producer reduce the calls made to
mempool library by using get_bulk() instead of get().
Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
---
app/test-eventdev/test_perf_common.c | 61 +++++++++++++++-------------
1 file changed, 33 insertions(+), 28 deletions(-)
Comments
> -----Original Message-----
> From: pbhagavatula@marvell.com <pbhagavatula@marvell.com>
> Sent: Wednesday, July 3, 2019 11:22 AM
> To: Jerin Jacob Kollanukkaran <jerinj@marvell.com>
> Cc: dev@dpdk.org; Pavan Nikhilesh Bhagavatula <pbhagavatula@marvell.com>
> Subject: [dpdk-dev][PATCH] app/test-eventdev: optimize producer routine
>
> From: Pavan Nikhilesh <pbhagavatula@marvell.com>
>
> When using synthetic and timer event producer reduce the calls made to
> mempool library by using get_bulk() instead of get().
>
> Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
Acked-by: Jerin Jacob <jerinj@marvell.com>
Applied to dpdk-next-eventdev/master. Thanks.
@@ -28,6 +28,7 @@ perf_test_result(struct evt_test *test, struct evt_options *opt)
static inline int
perf_producer(void *arg)
{
+ int i;
struct prod_data *p = arg;
struct test_perf *t = p->t;
struct evt_options *opt = t->opt;
@@ -38,7 +39,7 @@ perf_producer(void *arg)
const uint32_t nb_flows = t->nb_flows;
uint32_t flow_counter = 0;
uint64_t count = 0;
- struct perf_elt *m;
+ struct perf_elt *m[BURST_SIZE + 1] = {NULL};
struct rte_event ev;
if (opt->verbose_level > 1)
@@ -54,19 +55,21 @@ perf_producer(void *arg)
ev.sub_event_type = 0; /* stage 0 */
while (count < nb_pkts && t->done == false) {
- if (rte_mempool_get(pool, (void **)&m) < 0)
+ if (rte_mempool_get_bulk(pool, (void **)m, BURST_SIZE) < 0)
continue;
-
- ev.flow_id = flow_counter++ % nb_flows;
- ev.event_ptr = m;
- m->timestamp = rte_get_timer_cycles();
- while (rte_event_enqueue_burst(dev_id, port, &ev, 1) != 1) {
- if (t->done)
- break;
- rte_pause();
- m->timestamp = rte_get_timer_cycles();
+ for (i = 0; i < BURST_SIZE; i++) {
+ ev.flow_id = flow_counter++ % nb_flows;
+ ev.event_ptr = m[i];
+ m[i]->timestamp = rte_get_timer_cycles();
+ while (rte_event_enqueue_burst(dev_id,
+ port, &ev, 1) != 1) {
+ if (t->done)
+ break;
+ rte_pause();
+ m[i]->timestamp = rte_get_timer_cycles();
+ }
}
- count++;
+ count += BURST_SIZE;
}
return 0;
@@ -75,6 +78,7 @@ perf_producer(void *arg)
static inline int
perf_event_timer_producer(void *arg)
{
+ int i;
struct prod_data *p = arg;
struct test_perf *t = p->t;
struct evt_options *opt = t->opt;
@@ -85,7 +89,7 @@ perf_event_timer_producer(void *arg)
const uint32_t nb_flows = t->nb_flows;
const uint64_t nb_timers = opt->nb_timers;
struct rte_mempool *pool = t->pool;
- struct perf_elt *m;
+ struct perf_elt *m[BURST_SIZE + 1] = {NULL};
struct rte_event_timer_adapter **adptr = t->timer_adptr;
struct rte_event_timer tim;
uint64_t timeout_ticks = opt->expiry_nsec / opt->timer_tick_nsec;
@@ -107,23 +111,24 @@ perf_event_timer_producer(void *arg)
printf("%s(): lcore %d\n", __func__, rte_lcore_id());
while (count < nb_timers && t->done == false) {
- if (rte_mempool_get(pool, (void **)&m) < 0)
+ if (rte_mempool_get_bulk(pool, (void **)m, BURST_SIZE) < 0)
continue;
-
- m->tim = tim;
- m->tim.ev.flow_id = flow_counter++ % nb_flows;
- m->tim.ev.event_ptr = m;
- m->timestamp = rte_get_timer_cycles();
- while (rte_event_timer_arm_burst(
- adptr[flow_counter % nb_timer_adptrs],
- (struct rte_event_timer **)&m, 1) != 1) {
- if (t->done)
- break;
- rte_pause();
- m->timestamp = rte_get_timer_cycles();
+ for (i = 0; i < BURST_SIZE; i++) {
+ rte_prefetch0(m[i + 1]);
+ m[i]->tim = tim;
+ m[i]->tim.ev.flow_id = flow_counter++ % nb_flows;
+ m[i]->tim.ev.event_ptr = m[i];
+ m[i]->timestamp = rte_get_timer_cycles();
+ while (rte_event_timer_arm_burst(
+ adptr[flow_counter % nb_timer_adptrs],
+ (struct rte_event_timer **)&m[i], 1) != 1) {
+ if (t->done)
+ break;
+ m[i]->timestamp = rte_get_timer_cycles();
+ }
+ arm_latency += rte_get_timer_cycles() - m[i]->timestamp;
}
- arm_latency += rte_get_timer_cycles() - m->timestamp;
- count++;
+ count += BURST_SIZE;
}
fflush(stdout);
rte_delay_ms(1000);