app/testeventdev: add perf queue test
authorJerin Jacob <jerin.jacob@caviumnetworks.com>
Tue, 4 Jul 2017 04:53:19 +0000 (10:23 +0530)
committerJerin Jacob <jerin.jacob@caviumnetworks.com>
Fri, 7 Jul 2017 07:35:02 +0000 (09:35 +0200)
This is a performance test case that aims at testing the following:
1. Measure the number of events can be processed in a second.
2. Measure the latency to forward an event.

The perf queue test configures the eventdev with Q queues and P ports,
where Q is nb_producers * nb_stages and P is nb_workers + nb_producers.

The user can choose the number of workers, the number of producers and
number of stages through the --wlcores , --plcores and the --stlist
application command line arguments respectively.

The producer(s) injects the events to eventdev based the
first stage sched type list requested by the user through --stlist
the command line argument.

Based on the number of stages to process(selected through --stlist),
the application forwards the event to next upstream queue and
terminates when it reaches the last stage in the pipeline.
On event termination, application increments the number events
processed and print periodically in one second to get the
number of events processed in one second.

When --fwd_latency command line option selected, the application
inserts the timestamp in the event on the first stage and then
on termination, it updates the number of cycles to forward
a packet. The application uses this value to compute the average
latency to a forward packet.

Example command to run perf queue test:
sudo build/app/dpdk-test-eventdev --vdev=event_sw0 -- --test=perf_queue\
--slcore=1 --plcores=2 --wlcore=3 --stlist=p --nb_pkts=1000000000

Signed-off-by: Jerin Jacob <jerin.jacob@caviumnetworks.com>
Acked-by: Harry van Haaren <harry.van.haaren@intel.com>
app/test-eventdev/Makefile
app/test-eventdev/test_perf_queue.c [new file with mode: 0644]

index 242d3ee..7fed73e 100644 (file)
@@ -48,5 +48,6 @@ SRCS-y += test_order_queue.c
 SRCS-y += test_order_atq.c
 
 SRCS-y += test_perf_common.c
+SRCS-y += test_perf_queue.c
 
 include $(RTE_SDK)/mk/rte.app.mk
diff --git a/app/test-eventdev/test_perf_queue.c b/app/test-eventdev/test_perf_queue.c
new file mode 100644 (file)
index 0000000..1ac8231
--- /dev/null
@@ -0,0 +1,152 @@
+/*
+ *   BSD LICENSE
+ *
+ *   Copyright (C) Cavium 2017.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Cavium nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "test_perf_common.h"
+
+/* See http://dpdk.org/doc/guides/tools/testeventdev.html for test details */
+
+static inline int
+perf_queue_nb_event_queues(struct evt_options *opt)
+{
+       /* nb_queues = number of producers * number of stages */
+       return evt_nr_active_lcores(opt->plcores) * opt->nb_stages;
+}
+
+static int
+perf_queue_eventdev_setup(struct evt_test *test, struct evt_options *opt)
+{
+       uint8_t queue;
+       int nb_stages = opt->nb_stages;
+       int ret;
+
+       const struct rte_event_dev_config config = {
+                       .nb_event_queues = perf_queue_nb_event_queues(opt),
+                       .nb_event_ports = perf_nb_event_ports(opt),
+                       .nb_events_limit  = 4096,
+                       .nb_event_queue_flows = opt->nb_flows,
+                       .nb_event_port_dequeue_depth = 128,
+                       .nb_event_port_enqueue_depth = 128,
+       };
+
+       ret = rte_event_dev_configure(opt->dev_id, &config);
+       if (ret) {
+               evt_err("failed to configure eventdev %d", opt->dev_id);
+               return ret;
+       }
+
+       struct rte_event_queue_conf q_conf = {
+                       .priority = RTE_EVENT_DEV_PRIORITY_NORMAL,
+                       .nb_atomic_flows = opt->nb_flows,
+                       .nb_atomic_order_sequences = opt->nb_flows,
+       };
+       /* queue configurations */
+       for (queue = 0; queue < perf_queue_nb_event_queues(opt); queue++) {
+               q_conf.event_queue_cfg =  evt_sched_type2queue_cfg
+                               (opt->sched_type_list[queue % nb_stages]);
+
+               if (opt->q_priority) {
+                       uint8_t stage_pos = queue % nb_stages;
+                       /* Configure event queues(stage 0 to stage n) with
+                        * RTE_EVENT_DEV_PRIORITY_LOWEST to
+                        * RTE_EVENT_DEV_PRIORITY_HIGHEST.
+                        */
+                       uint8_t step = RTE_EVENT_DEV_PRIORITY_LOWEST /
+                                       (nb_stages - 1);
+                       /* Higher prio for the queues closer to last stage */
+                       q_conf.priority = RTE_EVENT_DEV_PRIORITY_LOWEST -
+                                       (step * stage_pos);
+               }
+               ret = rte_event_queue_setup(opt->dev_id, queue, &q_conf);
+               if (ret) {
+                       evt_err("failed to setup queue=%d", queue);
+                       return ret;
+               }
+       }
+
+       ret = perf_event_dev_port_setup(test, opt, nb_stages /* stride */,
+                                       perf_queue_nb_event_queues(opt));
+       if (ret)
+               return ret;
+
+       ret = rte_event_dev_start(opt->dev_id);
+       if (ret) {
+               evt_err("failed to start eventdev %d", opt->dev_id);
+               return ret;
+       }
+
+       return 0;
+}
+
+static void
+perf_queue_opt_dump(struct evt_options *opt)
+{
+       evt_dump_fwd_latency(opt);
+       perf_opt_dump(opt, perf_queue_nb_event_queues(opt));
+}
+
+static int
+perf_queue_opt_check(struct evt_options *opt)
+{
+       return perf_opt_check(opt, perf_queue_nb_event_queues(opt));
+}
+
+static bool
+perf_queue_capability_check(struct evt_options *opt)
+{
+       struct rte_event_dev_info dev_info;
+
+       rte_event_dev_info_get(opt->dev_id, &dev_info);
+       if (dev_info.max_event_queues < perf_queue_nb_event_queues(opt) ||
+                       dev_info.max_event_ports < perf_nb_event_ports(opt)) {
+               evt_err("not enough eventdev queues=%d/%d or ports=%d/%d",
+                       perf_queue_nb_event_queues(opt),
+                       dev_info.max_event_queues,
+                       perf_nb_event_ports(opt), dev_info.max_event_ports);
+       }
+
+       return true;
+}
+
+static const struct evt_test_ops perf_queue =  {
+       .cap_check          = perf_queue_capability_check,
+       .opt_check          = perf_queue_opt_check,
+       .opt_dump           = perf_queue_opt_dump,
+       .test_setup         = perf_test_setup,
+       .mempool_setup      = perf_mempool_setup,
+       .eventdev_setup     = perf_queue_eventdev_setup,
+       .eventdev_destroy   = perf_eventdev_destroy,
+       .mempool_destroy    = perf_mempool_destroy,
+       .test_result        = perf_test_result,
+       .test_destroy       = perf_test_destroy,
+};
+
+EVT_TEST_REGISTER(perf_queue);