xref: /dpdk/app/test/test_distributor_perf.c (revision 089e5ed727a15da2729cfee9b63533dd120bd04c)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2017 Intel Corporation
3  */
4 
5 #include "test.h"
6 
7 #include <unistd.h>
8 #include <string.h>
9 #include <rte_mempool.h>
10 #include <rte_cycles.h>
11 #include <rte_common.h>
12 #include <rte_mbuf.h>
13 #include <rte_distributor.h>
14 #include <rte_pause.h>
15 
16 #define ITER_POWER_CL 25 /* log 2 of how many iterations  for Cache Line test */
17 #define ITER_POWER 21 /* log 2 of how many iterations we do when timing. */
18 #define BURST 64
19 #define BIG_BATCH 1024
20 
21 /* static vars - zero initialized by default */
22 static volatile int quit;
23 static volatile unsigned worker_idx;
24 
25 struct worker_stats {
26 	volatile unsigned handled_packets;
27 } __rte_cache_aligned;
28 struct worker_stats worker_stats[RTE_MAX_LCORE];
29 
30 /*
31  * worker thread used for testing the time to do a round-trip of a cache
32  * line between two cores and back again
33  */
34 static int
35 flip_bit(volatile uint64_t *arg)
36 {
37 	uint64_t old_val = 0;
38 	while (old_val != 2) {
39 		while (!*arg)
40 			rte_pause();
41 		old_val = *arg;
42 		*arg = 0;
43 	}
44 	return 0;
45 }
46 
47 /*
48  * test case to time the number of cycles to round-trip a cache line between
49  * two cores and back again.
50  */
51 static void
52 time_cache_line_switch(void)
53 {
54 	/* allocate a full cache line for data, we use only first byte of it */
55 	uint64_t data[RTE_CACHE_LINE_SIZE*3 / sizeof(uint64_t)];
56 
57 	unsigned i, slaveid = rte_get_next_lcore(rte_lcore_id(), 0, 0);
58 	volatile uint64_t *pdata = &data[0];
59 	*pdata = 1;
60 	rte_eal_remote_launch((lcore_function_t *)flip_bit, &data[0], slaveid);
61 	while (*pdata)
62 		rte_pause();
63 
64 	const uint64_t start_time = rte_rdtsc();
65 	for (i = 0; i < (1 << ITER_POWER_CL); i++) {
66 		while (*pdata)
67 			rte_pause();
68 		*pdata = 1;
69 	}
70 	const uint64_t end_time = rte_rdtsc();
71 
72 	while (*pdata)
73 		rte_pause();
74 	*pdata = 2;
75 	rte_eal_wait_lcore(slaveid);
76 	printf("==== Cache line switch test ===\n");
77 	printf("Time for %u iterations = %"PRIu64" ticks\n", (1<<ITER_POWER_CL),
78 			end_time-start_time);
79 	printf("Ticks per iteration = %"PRIu64"\n\n",
80 			(end_time-start_time) >> ITER_POWER_CL);
81 }
82 
83 /*
84  * returns the total count of the number of packets handled by the worker
85  * functions given below.
86  */
87 static unsigned
88 total_packet_count(void)
89 {
90 	unsigned i, count = 0;
91 	for (i = 0; i < worker_idx; i++)
92 		count += worker_stats[i].handled_packets;
93 	return count;
94 }
95 
96 /* resets the packet counts for a new test */
97 static void
98 clear_packet_count(void)
99 {
100 	memset(&worker_stats, 0, sizeof(worker_stats));
101 }
102 
103 /*
104  * This is the basic worker function for performance tests.
105  * it does nothing but return packets and count them.
106  */
107 static int
108 handle_work(void *arg)
109 {
110 	struct rte_distributor *d = arg;
111 	unsigned int count = 0;
112 	unsigned int num = 0;
113 	int i;
114 	unsigned int id = __atomic_fetch_add(&worker_idx, 1, __ATOMIC_RELAXED);
115 	struct rte_mbuf *buf[8] __rte_cache_aligned;
116 
117 	for (i = 0; i < 8; i++)
118 		buf[i] = NULL;
119 
120 	num = rte_distributor_get_pkt(d, id, buf, buf, num);
121 	while (!quit) {
122 		worker_stats[id].handled_packets += num;
123 		count += num;
124 		num = rte_distributor_get_pkt(d, id, buf, buf, num);
125 	}
126 	worker_stats[id].handled_packets += num;
127 	count += num;
128 	rte_distributor_return_pkt(d, id, buf, num);
129 	return 0;
130 }
131 
132 /*
133  * This basic performance test just repeatedly sends in 32 packets at a time
134  * to the distributor and verifies at the end that we got them all in the worker
135  * threads and finally how long per packet the processing took.
136  */
137 static inline int
138 perf_test(struct rte_distributor *d, struct rte_mempool *p)
139 {
140 	unsigned int i;
141 	uint64_t start, end;
142 	struct rte_mbuf *bufs[BURST];
143 
144 	clear_packet_count();
145 	if (rte_mempool_get_bulk(p, (void *)bufs, BURST) != 0) {
146 		printf("Error getting mbufs from pool\n");
147 		return -1;
148 	}
149 	/* ensure we have different hash value for each pkt */
150 	for (i = 0; i < BURST; i++)
151 		bufs[i]->hash.usr = i;
152 
153 	start = rte_rdtsc();
154 	for (i = 0; i < (1<<ITER_POWER); i++)
155 		rte_distributor_process(d, bufs, BURST);
156 	end = rte_rdtsc();
157 
158 	do {
159 		usleep(100);
160 		rte_distributor_process(d, NULL, 0);
161 	} while (total_packet_count() < (BURST << ITER_POWER));
162 
163 	rte_distributor_clear_returns(d);
164 
165 	printf("Time per burst:  %"PRIu64"\n", (end - start) >> ITER_POWER);
166 	printf("Time per packet: %"PRIu64"\n\n",
167 			((end - start) >> ITER_POWER)/BURST);
168 	rte_mempool_put_bulk(p, (void *)bufs, BURST);
169 
170 	for (i = 0; i < rte_lcore_count() - 1; i++)
171 		printf("Worker %u handled %u packets\n", i,
172 				worker_stats[i].handled_packets);
173 	printf("Total packets: %u (%x)\n", total_packet_count(),
174 			total_packet_count());
175 	printf("=== Perf test done ===\n\n");
176 
177 	return 0;
178 }
179 
180 /* Useful function which ensures that all worker functions terminate */
181 static void
182 quit_workers(struct rte_distributor *d, struct rte_mempool *p)
183 {
184 	const unsigned int num_workers = rte_lcore_count() - 1;
185 	unsigned int i;
186 	struct rte_mbuf *bufs[RTE_MAX_LCORE];
187 
188 	rte_mempool_get_bulk(p, (void *)bufs, num_workers);
189 
190 	quit = 1;
191 	for (i = 0; i < num_workers; i++)
192 		bufs[i]->hash.usr = i << 1;
193 	rte_distributor_process(d, bufs, num_workers);
194 
195 	rte_mempool_put_bulk(p, (void *)bufs, num_workers);
196 
197 	rte_distributor_process(d, NULL, 0);
198 	rte_eal_mp_wait_lcore();
199 	quit = 0;
200 	worker_idx = 0;
201 }
202 
203 static int
204 test_distributor_perf(void)
205 {
206 	static struct rte_distributor *ds;
207 	static struct rte_distributor *db;
208 	static struct rte_mempool *p;
209 
210 	if (rte_lcore_count() < 2) {
211 		printf("Not enough cores for distributor_perf_autotest, expecting at least 2\n");
212 		return TEST_SKIPPED;
213 	}
214 
215 	/* first time how long it takes to round-trip a cache line */
216 	time_cache_line_switch();
217 
218 	if (ds == NULL) {
219 		ds = rte_distributor_create("Test_perf", rte_socket_id(),
220 				rte_lcore_count() - 1,
221 				RTE_DIST_ALG_SINGLE);
222 		if (ds == NULL) {
223 			printf("Error creating distributor\n");
224 			return -1;
225 		}
226 	} else {
227 		rte_distributor_clear_returns(ds);
228 	}
229 
230 	if (db == NULL) {
231 		db = rte_distributor_create("Test_burst", rte_socket_id(),
232 				rte_lcore_count() - 1,
233 				RTE_DIST_ALG_BURST);
234 		if (db == NULL) {
235 			printf("Error creating burst distributor\n");
236 			return -1;
237 		}
238 	} else {
239 		rte_distributor_clear_returns(db);
240 	}
241 
242 	const unsigned nb_bufs = (511 * rte_lcore_count()) < BIG_BATCH ?
243 			(BIG_BATCH * 2) - 1 : (511 * rte_lcore_count());
244 	if (p == NULL) {
245 		p = rte_pktmbuf_pool_create("DPT_MBUF_POOL", nb_bufs, BURST,
246 			0, RTE_MBUF_DEFAULT_BUF_SIZE, rte_socket_id());
247 		if (p == NULL) {
248 			printf("Error creating mempool\n");
249 			return -1;
250 		}
251 	}
252 
253 	printf("=== Performance test of distributor (single mode) ===\n");
254 	rte_eal_mp_remote_launch(handle_work, ds, SKIP_MASTER);
255 	if (perf_test(ds, p) < 0)
256 		return -1;
257 	quit_workers(ds, p);
258 
259 	printf("=== Performance test of distributor (burst mode) ===\n");
260 	rte_eal_mp_remote_launch(handle_work, db, SKIP_MASTER);
261 	if (perf_test(db, p) < 0)
262 		return -1;
263 	quit_workers(db, p);
264 
265 	return 0;
266 }
267 
268 REGISTER_TEST_COMMAND(distributor_perf_autotest, test_distributor_perf);
269