Re: [dpdk-dev] [PATCH v4 3/4] Add autotests for RIB library

[Bruce Richardson <bruce.richardson@intel.com>]

On Fri, Apr 27, 2018 at 01:03:33AM +0300, Medvedkin Vladimir wrote:
> Signed-off-by: Medvedkin Vladimir <medvedkinv@gmail.com>
> ---
>  test/test/Makefile               |   5 +
>  test/test/meson.build            |   8 +
>  test/test/test_rib.c             | 308 +++++++++++++++++++++++++++++++++++++++
>  test/test/test_rib_generate_rt.c | 297 +++++++++++++++++++++++++++++++++++++
>  test/test/test_rib_generate_rt.h |  38 +++++
>  test/test/test_rib_lpm_comp.c    | 189 ++++++++++++++++++++++++
>  test/test/test_rib_perf.c        | 145 ++++++++++++++++++
>  7 files changed, 990 insertions(+)
>  create mode 100644 test/test/test_rib.c
>  create mode 100644 test/test/test_rib_generate_rt.c
>  create mode 100644 test/test/test_rib_generate_rt.h
>  create mode 100644 test/test/test_rib_lpm_comp.c
>  create mode 100644 test/test/test_rib_perf.c
> 

<snip>

> diff --git a/test/test/test_rib_perf.c b/test/test/test_rib_perf.c
> new file mode 100644
> index 0000000..42fbd1e
> --- /dev/null
> +++ b/test/test/test_rib_perf.c
> @@ -0,0 +1,145 @@
> +/* SPDX-License-Identifier: BSD-3-Clause
> + * Copyright(c) 2018 Vladimir Medvedkin <medvedkinv@gmail.com>
> + */
> +
> +#include <stdio.h>
> +#include <stdint.h>
> +#include <stdlib.h>
> +
> +#include <rte_cycles.h>
> +#include <rte_random.h>
> +#include <rte_branch_prediction.h>
> +#include <rte_ip.h>
> +#include <rte_malloc.h>
> +#include <rte_rib.h>
> +#include <rte_dir24_8.h>
> +
> +#include "test.h"
> +#include "test_xmmt_ops.h"
> +#include "test_rib_generate_rt.h"
> +
> +#define TEST_RIB_ASSERT(cond) do {				\
> +	if (!(cond)) {						\
> +		printf("Error at line %d:\n", __LINE__);	\
> +		return -1;					\
> +	}							\
> +} while (0)
> +
> +#define ITERATIONS (1 << 15)
> +#define BATCH_SIZE (1 << 12)
> +#define BULK_SIZE 32
> +
> +#define NH_MSK(nh_sz)	((1ULL << ((1 << (3 + nh_sz)) - 1)) - 1)
> +
> +static int
> +test_rib_perf(void)
> +{
> +	struct rte_rib *rib = NULL;
> +	struct rte_rib_conf conf;
> +	struct route_rule *rt;
> +	uint64_t begin, total_time;
> +	uint64_t next_hop_add;
> +	uint64_t default_nh = 0;
> +	int64_t count = 0;
> +	unsigned int i, j;
> +	int status = 0;
> +	int ret, nh_bits, nr_tbl8;
> +	uint32_t num_routes;
> +
> +	conf.max_nodes = 3000000;
> +	conf.node_sz = sizeof(struct rte_rib_node);
> +	conf.type = RTE_RIB_DIR24_8;
> +	conf.fib_conf.dir24_8.def_nh = default_nh;
> +	conf.fib_conf.dir24_8.nh_sz = RTE_DIR24_8_8B;

It's good that you are taking the worst-case to show the perf, but it means
that the library comes out a bit slower than LPM in the autotest.
How about running the same test cases for multiple data sizes, 8, 4, 2?

> +
> +	rte_srand(rte_rdtsc());
> +
> +	nh_bits = RTE_MIN(((1 << (3 + conf.fib_conf.dir24_8.nh_sz)) - 1), 24);
> +	nr_tbl8 = RTE_MIN(((1 << nh_bits) - 1), 131071);
> +	conf.fib_conf.dir24_8.num_tbl8 = nr_tbl8;
> +	num_routes = 1200000;
> +
> +	rt = rte_zmalloc("struct route_rule *", sizeof(struct route_rule) *
> +		num_routes, 0);
> +	TEST_RIB_ASSERT(rt != NULL);
> +
> +	num_routes = generate_large_route_rule_table(num_routes, rt);
> +	TEST_RIB_ASSERT(num_routes != 0);
> +
> +	printf("No. routes = %u\n", (unsigned int) num_routes);
> +
> +	shuffle_rt(rt, num_routes);
> +
> +	print_route_distribution(rt, (uint32_t) num_routes);
> +
> +	rib = rte_rib_create(__func__, SOCKET_ID_ANY, &conf);
> +	TEST_RIB_ASSERT(rib != NULL);
> +
> +	/* Measue add. */
> +	begin = rte_rdtsc();
> +
> +	for (i = 0; i < num_routes; i++) {
> +		do {
> +			next_hop_add = rte_rand() & NH_MSK(conf.fib_conf.dir24_8.nh_sz);
> +		} while (next_hop_add == default_nh);
> +
> +		ret = rte_rib_add(rib, rt[i].ip, rt[i].depth, next_hop_add);
> +		if ((ret == 0))
> +			status++;
> +	}
> +
> +	total_time = rte_rdtsc() - begin;
> +
> +	printf("Unique added entries = %d\n", status);
> +	printf("Average RIB Add: %g cycles\n",
> +			(double)total_time / num_routes);
> +
> +	/* Measure bulk Lookup */
> +	total_time = 0;
> +	count = 0;
> +	for (i = 0; i < ITERATIONS; i++) {
> +		static uint32_t ip_batch[BATCH_SIZE];
> +		uint64_t next_hops[BULK_SIZE];
> +
> +		/* Create array of random IP addresses */
> +		for (j = 0; j < BATCH_SIZE; j++)
> +			ip_batch[j] = rte_rand();
> +
> +		/* Lookup per batch */
> +		begin = rte_rdtsc();
> +		for (j = 0; j < BATCH_SIZE; j += BULK_SIZE)
> +			rte_rib_fib_lookup_bulk(rib, &ip_batch[j], next_hops,
> +				BULK_SIZE);
> +
> +		total_time += rte_rdtsc() - begin;
> +		for (j = 0; j < BULK_SIZE; j++) {
> +			if (next_hops[j] == default_nh)
> +				count++;
> +		}
> +	}
> +	printf("BULK RIB Lookup: %.1f cycles (fails = %.1f%%)\n",
> +			(double)total_time / ((double)ITERATIONS * BATCH_SIZE),
> +			(count * 100.0) / (double)(ITERATIONS * BATCH_SIZE));
> +
> +	/* Delete */
> +	status = 0;
> +	begin = rte_rdtsc();
> +
> +	for (i = 0; i < num_routes; i++) {
> +		ret = rte_rib_delete(rib, rt[i].ip, rt[i].depth);
> +		if (ret == 0)
> +			status++;
> +	}
> +
> +	total_time = rte_rdtsc() - begin;
> +
> +	printf("Average RIB Delete: %g cycles\n",
> +			(double)total_time / num_routes);
> +
> +	rte_rib_free(rib);
> +	rte_free(rt);
> +
> +	return 0;
> +}
> +
> +REGISTER_TEST_COMMAND(rib_perf_autotest, test_rib_perf);
> -- 
> 1.8.3.1
>