| Message ID | 1469181897-19168-4-git-send-email-matias.elo@nokia.com |
|---|---|
| State | New |
| Headers | show |
On Fri, Jul 22, 2016 at 5:04 AM, Matias Elo <matias.elo@nokia.com> wrote: > Use odp_buffer_alloc_multi() and odp_buffer_free_multi() functions where > appropriate for improved performance. > > Signed-off-by: Matias Elo <matias.elo@nokia.com> > --- > test/common_plat/performance/odp_scheduling.c | 50 > ++++++++++++++++++--------- > 1 file changed, 33 insertions(+), 17 deletions(-) > > diff --git a/test/common_plat/performance/odp_scheduling.c > b/test/common_plat/performance/odp_scheduling.c > index 5a2997f..1d1bf01 100644 > --- a/test/common_plat/performance/odp_scheduling.c > +++ b/test/common_plat/performance/odp_scheduling.c > @@ -29,7 +29,7 @@ > #include <getopt.h> > > #define MSG_POOL_SIZE (4 * 1024 * 1024) /**< Message pool size */ > -#define MAX_ALLOCS 35 /**< Alloc burst size */ > +#define MAX_ALLOCS 32 /**< Alloc burst size */ > Not a biggie, but it's not clear why this change has been made. As far as I can tell, changing this has negligible effect on the performance numbers reported. > #define QUEUES_PER_PRIO 64 /**< Queue per priority */ > #define NUM_PRIOS 2 /**< Number of tested > priorities */ > #define QUEUE_ROUNDS (512 * 1024) /**< Queue test rounds */ > @@ -131,9 +131,10 @@ static void clear_sched_queues(void) > static int enqueue_events(int thr, int prio, int num_queues, int > num_events, > test_globals_t *globals) > { > - odp_buffer_t buf; > + odp_buffer_t buf[num_events]; > + odp_event_t ev[num_events]; > odp_queue_t queue; > - int i, j, k; > + int i, j, k, ret; > > if (prio == ODP_SCHED_PRIO_HIGHEST) > i = 0; > @@ -144,19 +145,28 @@ static int enqueue_events(int thr, int prio, int > num_queues, int num_events, > for (j = 0; j < num_queues; j++) { > queue = globals->queue[i][j]; > > + ret = odp_buffer_alloc_multi(globals->pool, buf, > num_events); > + if (ret != num_events) { > + LOG_ERR(" [%i] buffer alloc failed\n", thr); > + ret = ret < 0 ? 0 : ret; > + odp_buffer_free_multi(buf, ret); > + return -1; > + } > for (k = 0; k < num_events; k++) { > - buf = odp_buffer_alloc(globals->pool); > - > - if (!odp_buffer_is_valid(buf)) { > + if (!odp_buffer_is_valid(buf[k])) { > LOG_ERR(" [%i] buffer alloc failed\n", > thr); > + odp_buffer_free_multi(buf, num_events); > return -1; > } > + ev[k] = odp_buffer_to_event(buf[k]); > + } > > - if (odp_queue_enq(queue, > odp_buffer_to_event(buf))) { > - LOG_ERR(" [%i] Queue enqueue failed.\n", > thr); > - odp_buffer_free(buf); > - return -1; > - } > + ret = odp_queue_enq_multi(queue, ev, num_events); > + if (ret != num_events) { > + LOG_ERR(" [%i] Queue enqueue failed.\n", thr); > + ret = ret < 0 ? 0 : ret; > + odp_buffer_free_multi(&buf[ret], num_events - ret); > + return -1; > } > } > > @@ -210,24 +220,30 @@ static int test_alloc_single(int thr, test_globals_t > *globals) > */ > static int test_alloc_multi(int thr, test_globals_t *globals) > { > - int i, j; > + int i, j, ret; > odp_buffer_t temp_buf[MAX_ALLOCS]; > uint64_t c1, c2, cycles; > > c1 = odp_cpu_cycles(); > > for (i = 0; i < ALLOC_ROUNDS; i++) { > + ret = odp_buffer_alloc_multi(globals->pool, temp_buf, > + MAX_ALLOCS); > + if (ret != MAX_ALLOCS) { > + LOG_ERR(" [%i] buffer alloc failed\n", thr); > + ret = ret < 0 ? 0 : ret; > + odp_buffer_free_multi(temp_buf, ret); > + return -1; > + } > + > for (j = 0; j < MAX_ALLOCS; j++) { > - temp_buf[j] = odp_buffer_alloc(globals->pool); > - > if (!odp_buffer_is_valid(temp_buf[j])) { > LOG_ERR(" [%i] alloc_multi failed\n", > thr); > + odp_buffer_free_multi(temp_buf, > MAX_ALLOCS); > return -1; > } > } > - > - for (; j > 0; j--) > - odp_buffer_free(temp_buf[j - 1]); > + odp_buffer_free_multi(temp_buf, MAX_ALLOCS); > } > > c2 = odp_cpu_cycles(); > -- > 2.7.4 > >
On 07/26/16 03:14, Bill Fischofer wrote: >> > #define MSG_POOL_SIZE (4 * 1024 * 1024) /**< Message pool size */ >> >-#define MAX_ALLOCS 35 /**< Alloc burst size */ >> >+#define MAX_ALLOCS 32 /**< Alloc burst size */ >> > > Not a biggie, but it's not clear why this change has been made. As far as I > can tell, changing this has negligible effect on the performance numbers > reported. > > I guess that is because of round division (4 * 1024 * 1024) / 4194304 = 32 Maxim.
diff --git a/test/common_plat/performance/odp_scheduling.c b/test/common_plat/performance/odp_scheduling.c index 5a2997f..1d1bf01 100644 --- a/test/common_plat/performance/odp_scheduling.c +++ b/test/common_plat/performance/odp_scheduling.c @@ -29,7 +29,7 @@ #include <getopt.h> #define MSG_POOL_SIZE (4 * 1024 * 1024) /**< Message pool size */ -#define MAX_ALLOCS 35 /**< Alloc burst size */ +#define MAX_ALLOCS 32 /**< Alloc burst size */ #define QUEUES_PER_PRIO 64 /**< Queue per priority */ #define NUM_PRIOS 2 /**< Number of tested priorities */ #define QUEUE_ROUNDS (512 * 1024) /**< Queue test rounds */ @@ -131,9 +131,10 @@ static void clear_sched_queues(void) static int enqueue_events(int thr, int prio, int num_queues, int num_events, test_globals_t *globals) { - odp_buffer_t buf; + odp_buffer_t buf[num_events]; + odp_event_t ev[num_events]; odp_queue_t queue; - int i, j, k; + int i, j, k, ret; if (prio == ODP_SCHED_PRIO_HIGHEST) i = 0; @@ -144,19 +145,28 @@ static int enqueue_events(int thr, int prio, int num_queues, int num_events, for (j = 0; j < num_queues; j++) { queue = globals->queue[i][j]; + ret = odp_buffer_alloc_multi(globals->pool, buf, num_events); + if (ret != num_events) { + LOG_ERR(" [%i] buffer alloc failed\n", thr); + ret = ret < 0 ? 0 : ret; + odp_buffer_free_multi(buf, ret); + return -1; + } for (k = 0; k < num_events; k++) { - buf = odp_buffer_alloc(globals->pool); - - if (!odp_buffer_is_valid(buf)) { + if (!odp_buffer_is_valid(buf[k])) { LOG_ERR(" [%i] buffer alloc failed\n", thr); + odp_buffer_free_multi(buf, num_events); return -1; } + ev[k] = odp_buffer_to_event(buf[k]); + } - if (odp_queue_enq(queue, odp_buffer_to_event(buf))) { - LOG_ERR(" [%i] Queue enqueue failed.\n", thr); - odp_buffer_free(buf); - return -1; - } + ret = odp_queue_enq_multi(queue, ev, num_events); + if (ret != num_events) { + LOG_ERR(" [%i] Queue enqueue failed.\n", thr); + ret = ret < 0 ? 0 : ret; + odp_buffer_free_multi(&buf[ret], num_events - ret); + return -1; } } @@ -210,24 +220,30 @@ static int test_alloc_single(int thr, test_globals_t *globals) */ static int test_alloc_multi(int thr, test_globals_t *globals) { - int i, j; + int i, j, ret; odp_buffer_t temp_buf[MAX_ALLOCS]; uint64_t c1, c2, cycles; c1 = odp_cpu_cycles(); for (i = 0; i < ALLOC_ROUNDS; i++) { + ret = odp_buffer_alloc_multi(globals->pool, temp_buf, + MAX_ALLOCS); + if (ret != MAX_ALLOCS) { + LOG_ERR(" [%i] buffer alloc failed\n", thr); + ret = ret < 0 ? 0 : ret; + odp_buffer_free_multi(temp_buf, ret); + return -1; + } + for (j = 0; j < MAX_ALLOCS; j++) { - temp_buf[j] = odp_buffer_alloc(globals->pool); - if (!odp_buffer_is_valid(temp_buf[j])) { LOG_ERR(" [%i] alloc_multi failed\n", thr); + odp_buffer_free_multi(temp_buf, MAX_ALLOCS); return -1; } } - - for (; j > 0; j--) - odp_buffer_free(temp_buf[j - 1]); + odp_buffer_free_multi(temp_buf, MAX_ALLOCS); } c2 = odp_cpu_cycles();
Use odp_buffer_alloc_multi() and odp_buffer_free_multi() functions where appropriate for improved performance. Signed-off-by: Matias Elo <matias.elo@nokia.com> --- test/common_plat/performance/odp_scheduling.c | 50 ++++++++++++++++++--------- 1 file changed, 33 insertions(+), 17 deletions(-) -- 2.7.4