app/eventdev: optimize producer routine
authorPavan Nikhilesh <pbhagavatula@marvell.com>
Wed, 3 Jul 2019 05:51:35 +0000 (11:21 +0530)
committerJerin Jacob <jerinj@marvell.com>
Sun, 7 Jul 2019 12:48:42 +0000 (14:48 +0200)
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>
app/test-eventdev/test_perf_common.c

index 2e9e83e..aa925a7 100644 (file)
@@ -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);