]>
Commit | Line | Data |
---|---|---|
11fdf7f2 TL |
1 | /* SPDX-License-Identifier: BSD-3-Clause |
2 | * Copyright(c) 2016-2017 Intel Corporation | |
3 | */ | |
4 | ||
5 | #include <stdio.h> | |
6 | #include <inttypes.h> | |
7 | ||
8 | #include <rte_lcore.h> | |
9 | #include <rte_cycles.h> | |
10 | #include <rte_malloc.h> | |
11 | #include <rte_random.h> | |
12 | #include <rte_efd.h> | |
13 | #include <rte_memcpy.h> | |
14 | #include <rte_thash.h> | |
15 | ||
16 | #include "test.h" | |
17 | ||
18 | #define NUM_KEYSIZES 10 | |
19 | #define NUM_SHUFFLES 10 | |
20 | #define MAX_KEYSIZE 64 | |
21 | #define MAX_ENTRIES (1 << 19) | |
22 | #define KEYS_TO_ADD (MAX_ENTRIES * 3 / 4) /* 75% table utilization */ | |
23 | #define NUM_LOOKUPS (KEYS_TO_ADD * 5) /* Loop among keys added, several times */ | |
24 | ||
25 | #if RTE_EFD_VALUE_NUM_BITS == 32 | |
26 | #define VALUE_BITMASK 0xffffffff | |
27 | #else | |
28 | #define VALUE_BITMASK ((1 << RTE_EFD_VALUE_NUM_BITS) - 1) | |
29 | #endif | |
30 | static unsigned int test_socket_id; | |
31 | ||
32 | static inline uint8_t efd_get_all_sockets_bitmask(void) | |
33 | { | |
34 | uint8_t all_cpu_sockets_bitmask = 0; | |
35 | unsigned int i; | |
36 | unsigned int next_lcore = rte_get_master_lcore(); | |
37 | const int val_true = 1, val_false = 0; | |
38 | for (i = 0; i < rte_lcore_count(); i++) { | |
39 | all_cpu_sockets_bitmask |= 1 << rte_lcore_to_socket_id(next_lcore); | |
40 | next_lcore = rte_get_next_lcore(next_lcore, val_false, val_true); | |
41 | } | |
42 | ||
43 | return all_cpu_sockets_bitmask; | |
44 | } | |
45 | ||
46 | enum operations { | |
47 | ADD = 0, | |
48 | LOOKUP, | |
49 | LOOKUP_MULTI, | |
50 | DELETE, | |
51 | NUM_OPERATIONS | |
52 | }; | |
53 | ||
54 | struct efd_perf_params { | |
55 | struct rte_efd_table *efd_table; | |
56 | uint32_t key_size; | |
57 | unsigned int cycle; | |
58 | }; | |
59 | ||
60 | static uint32_t hashtest_key_lens[] = { | |
61 | /* standard key sizes */ | |
62 | 4, 8, 16, 32, 48, 64, | |
63 | /* IPv4 SRC + DST + protocol, unpadded */ | |
64 | 9, | |
65 | /* IPv4 5-tuple, unpadded */ | |
66 | 13, | |
67 | /* IPv6 5-tuple, unpadded */ | |
68 | 37, | |
69 | /* IPv6 5-tuple, padded to 8-byte boundary */ | |
70 | 40 | |
71 | }; | |
72 | ||
73 | /* Array to store number of cycles per operation */ | |
f67539c2 | 74 | static uint64_t cycles[NUM_KEYSIZES][NUM_OPERATIONS]; |
11fdf7f2 TL |
75 | |
76 | /* Array to store the data */ | |
f67539c2 | 77 | static efd_value_t data[KEYS_TO_ADD]; |
11fdf7f2 TL |
78 | |
79 | /* Array to store all input keys */ | |
f67539c2 | 80 | static uint8_t keys[KEYS_TO_ADD][MAX_KEYSIZE]; |
11fdf7f2 TL |
81 | |
82 | /* Shuffle the keys that have been added, so lookups will be totally random */ | |
83 | static void | |
84 | shuffle_input_keys(struct efd_perf_params *params) | |
85 | { | |
86 | efd_value_t temp_data; | |
87 | unsigned int i; | |
88 | uint32_t swap_idx; | |
89 | uint8_t temp_key[MAX_KEYSIZE]; | |
90 | ||
91 | for (i = KEYS_TO_ADD - 1; i > 0; i--) { | |
92 | swap_idx = rte_rand() % i; | |
93 | ||
94 | memcpy(temp_key, keys[i], hashtest_key_lens[params->cycle]); | |
95 | temp_data = data[i]; | |
96 | ||
97 | memcpy(keys[i], keys[swap_idx], hashtest_key_lens[params->cycle]); | |
98 | data[i] = data[swap_idx]; | |
99 | ||
100 | memcpy(keys[swap_idx], temp_key, hashtest_key_lens[params->cycle]); | |
101 | data[swap_idx] = temp_data; | |
102 | } | |
103 | } | |
104 | ||
105 | static int key_compare(const void *key1, const void *key2) | |
106 | { | |
107 | return memcmp(key1, key2, MAX_KEYSIZE); | |
108 | } | |
109 | ||
110 | /* | |
111 | * TODO: we could "error proof" these as done in test_hash_perf.c ln 165: | |
112 | * | |
113 | * The current setup may give errors if too full in some cases which we check | |
114 | * for. However, since EFD allows for ~99% capacity, these errors are rare for | |
115 | * #"KEYS_TO_ADD" which is 75% capacity. | |
116 | */ | |
117 | static int | |
118 | setup_keys_and_data(struct efd_perf_params *params, unsigned int cycle) | |
119 | { | |
120 | unsigned int i, j; | |
121 | int num_duplicates; | |
122 | ||
123 | params->key_size = hashtest_key_lens[cycle]; | |
124 | params->cycle = cycle; | |
125 | ||
126 | /* Reset all arrays */ | |
127 | for (i = 0; i < params->key_size; i++) | |
128 | keys[0][i] = 0; | |
129 | ||
130 | /* Generate a list of keys, some of which may be duplicates */ | |
131 | for (i = 0; i < KEYS_TO_ADD; i++) { | |
132 | for (j = 0; j < params->key_size; j++) | |
133 | keys[i][j] = rte_rand() & 0xFF; | |
134 | ||
135 | data[i] = rte_rand() & VALUE_BITMASK; | |
136 | } | |
137 | ||
138 | /* Remove duplicates from the keys array */ | |
139 | do { | |
140 | num_duplicates = 0; | |
141 | ||
142 | /* Sort the list of keys to make it easier to find duplicates */ | |
143 | qsort(keys, KEYS_TO_ADD, MAX_KEYSIZE, key_compare); | |
144 | ||
145 | /* Sift through the list of keys and look for duplicates */ | |
146 | int num_duplicates = 0; | |
147 | for (i = 0; i < KEYS_TO_ADD - 1; i++) { | |
148 | if (memcmp(keys[i], keys[i + 1], params->key_size) == 0) { | |
149 | /* This key already exists, try again */ | |
150 | num_duplicates++; | |
151 | for (j = 0; j < params->key_size; j++) | |
152 | keys[i][j] = rte_rand() & 0xFF; | |
153 | } | |
154 | } | |
155 | } while (num_duplicates != 0); | |
156 | ||
157 | /* Shuffle the random values again */ | |
158 | shuffle_input_keys(params); | |
159 | ||
160 | params->efd_table = rte_efd_create("test_efd_perf", | |
161 | MAX_ENTRIES, params->key_size, | |
162 | efd_get_all_sockets_bitmask(), test_socket_id); | |
163 | TEST_ASSERT_NOT_NULL(params->efd_table, "Error creating the efd table\n"); | |
164 | ||
165 | return 0; | |
166 | } | |
167 | ||
168 | static int | |
169 | timed_adds(struct efd_perf_params *params) | |
170 | { | |
171 | const uint64_t start_tsc = rte_rdtsc(); | |
172 | unsigned int i, a; | |
173 | int32_t ret; | |
174 | ||
175 | for (i = 0; i < KEYS_TO_ADD; i++) { | |
176 | ret = rte_efd_update(params->efd_table, test_socket_id, keys[i], | |
177 | data[i]); | |
178 | if (ret != 0) { | |
179 | printf("Error %d in rte_efd_update - key=0x", ret); | |
180 | for (a = 0; a < params->key_size; a++) | |
181 | printf("%02x", keys[i][a]); | |
182 | printf(" value=%d\n", data[i]); | |
183 | ||
184 | return -1; | |
185 | } | |
186 | } | |
187 | ||
188 | const uint64_t end_tsc = rte_rdtsc(); | |
189 | const uint64_t time_taken = end_tsc - start_tsc; | |
190 | ||
191 | cycles[params->cycle][ADD] = time_taken / KEYS_TO_ADD; | |
192 | return 0; | |
193 | } | |
194 | ||
195 | static int | |
196 | timed_lookups(struct efd_perf_params *params) | |
197 | { | |
198 | unsigned int i, j, a; | |
199 | const uint64_t start_tsc = rte_rdtsc(); | |
200 | efd_value_t ret_data; | |
201 | ||
202 | for (i = 0; i < NUM_LOOKUPS / KEYS_TO_ADD; i++) { | |
203 | for (j = 0; j < KEYS_TO_ADD; j++) { | |
204 | ret_data = rte_efd_lookup(params->efd_table, | |
205 | test_socket_id, keys[j]); | |
206 | if (ret_data != data[j]) { | |
207 | printf("Value mismatch using rte_efd_lookup: " | |
208 | "key #%d (0x", i); | |
209 | for (a = 0; a < params->key_size; a++) | |
210 | printf("%02x", keys[i][a]); | |
211 | printf(")\n"); | |
212 | printf(" Expected %d, got %d\n", data[i], | |
213 | ret_data); | |
214 | ||
215 | return -1; | |
216 | } | |
217 | ||
218 | } | |
219 | } | |
220 | ||
221 | const uint64_t end_tsc = rte_rdtsc(); | |
222 | const uint64_t time_taken = end_tsc - start_tsc; | |
223 | ||
224 | cycles[params->cycle][LOOKUP] = time_taken / NUM_LOOKUPS; | |
225 | ||
226 | return 0; | |
227 | } | |
228 | ||
229 | static int | |
230 | timed_lookups_multi(struct efd_perf_params *params) | |
231 | { | |
232 | unsigned int i, j, k, a; | |
233 | efd_value_t result[RTE_EFD_BURST_MAX] = {0}; | |
234 | const void *keys_burst[RTE_EFD_BURST_MAX]; | |
235 | const uint64_t start_tsc = rte_rdtsc(); | |
236 | ||
237 | for (i = 0; i < NUM_LOOKUPS / KEYS_TO_ADD; i++) { | |
238 | for (j = 0; j < KEYS_TO_ADD / RTE_EFD_BURST_MAX; j++) { | |
239 | for (k = 0; k < RTE_EFD_BURST_MAX; k++) | |
240 | keys_burst[k] = keys[j * RTE_EFD_BURST_MAX + k]; | |
241 | ||
242 | rte_efd_lookup_bulk(params->efd_table, test_socket_id, | |
243 | RTE_EFD_BURST_MAX, | |
244 | keys_burst, result); | |
245 | ||
246 | for (k = 0; k < RTE_EFD_BURST_MAX; k++) { | |
247 | uint32_t data_idx = j * RTE_EFD_BURST_MAX + k; | |
248 | if (result[k] != data[data_idx]) { | |
249 | printf("Value mismatch using " | |
250 | "rte_efd_lookup_bulk: key #%d " | |
251 | "(0x", i); | |
252 | for (a = 0; a < params->key_size; a++) | |
253 | printf("%02x", | |
254 | keys[data_idx][a]); | |
255 | printf(")\n"); | |
256 | printf(" Expected %d, got %d\n", | |
257 | data[data_idx], result[k]); | |
258 | ||
259 | return -1; | |
260 | } | |
261 | } | |
262 | } | |
263 | } | |
264 | ||
265 | const uint64_t end_tsc = rte_rdtsc(); | |
266 | const uint64_t time_taken = end_tsc - start_tsc; | |
267 | ||
268 | cycles[params->cycle][LOOKUP_MULTI] = time_taken / NUM_LOOKUPS; | |
269 | ||
270 | return 0; | |
271 | } | |
272 | ||
273 | static int | |
274 | timed_deletes(struct efd_perf_params *params) | |
275 | { | |
276 | unsigned int i, a; | |
277 | const uint64_t start_tsc = rte_rdtsc(); | |
278 | int32_t ret; | |
279 | ||
280 | for (i = 0; i < KEYS_TO_ADD; i++) { | |
281 | ret = rte_efd_delete(params->efd_table, test_socket_id, keys[i], | |
282 | NULL); | |
283 | ||
284 | if (ret != 0) { | |
285 | printf("Error %d in rte_efd_delete - key=0x", ret); | |
286 | for (a = 0; a < params->key_size; a++) | |
287 | printf("%02x", keys[i][a]); | |
288 | printf("\n"); | |
289 | ||
290 | return -1; | |
291 | } | |
292 | } | |
293 | ||
294 | const uint64_t end_tsc = rte_rdtsc(); | |
295 | const uint64_t time_taken = end_tsc - start_tsc; | |
296 | ||
297 | cycles[params->cycle][DELETE] = time_taken / KEYS_TO_ADD; | |
298 | ||
299 | return 0; | |
300 | } | |
301 | ||
302 | static void | |
303 | perform_frees(struct efd_perf_params *params) | |
304 | { | |
305 | if (params->efd_table != NULL) { | |
306 | rte_efd_free(params->efd_table); | |
307 | params->efd_table = NULL; | |
308 | } | |
309 | } | |
310 | ||
311 | static int | |
312 | exit_with_fail(const char *testname, struct efd_perf_params *params, | |
313 | unsigned int i) | |
314 | { | |
315 | ||
316 | printf("<<<<<Test %s failed at keysize %d iteration %d >>>>>\n", | |
317 | testname, hashtest_key_lens[params->cycle], i); | |
318 | perform_frees(params); | |
319 | return -1; | |
320 | } | |
321 | ||
322 | static int | |
323 | run_all_tbl_perf_tests(void) | |
324 | { | |
325 | unsigned int i, j; | |
326 | struct efd_perf_params params; | |
327 | ||
328 | printf("Measuring performance, please wait\n"); | |
329 | fflush(stdout); | |
330 | ||
331 | test_socket_id = rte_socket_id(); | |
332 | ||
333 | for (i = 0; i < NUM_KEYSIZES; i++) { | |
334 | ||
335 | if (setup_keys_and_data(¶ms, i) < 0) { | |
336 | printf("Could not create keys/data/table\n"); | |
337 | return -1; | |
338 | } | |
339 | ||
340 | if (timed_adds(¶ms) < 0) | |
341 | return exit_with_fail("timed_adds", ¶ms, i); | |
342 | ||
343 | for (j = 0; j < NUM_SHUFFLES; j++) | |
344 | shuffle_input_keys(¶ms); | |
345 | ||
346 | if (timed_lookups(¶ms) < 0) | |
347 | return exit_with_fail("timed_lookups", ¶ms, i); | |
348 | ||
349 | if (timed_lookups_multi(¶ms) < 0) | |
350 | return exit_with_fail("timed_lookups_multi", ¶ms, i); | |
351 | ||
352 | if (timed_deletes(¶ms) < 0) | |
353 | return exit_with_fail("timed_deletes", ¶ms, i); | |
354 | ||
355 | /* Print a dot to show progress on operations */ | |
356 | printf("."); | |
357 | fflush(stdout); | |
358 | ||
359 | perform_frees(¶ms); | |
360 | } | |
361 | ||
362 | printf("\nResults (in CPU cycles/operation)\n"); | |
363 | printf("-----------------------------------\n"); | |
364 | printf("\n%-18s%-18s%-18s%-18s%-18s\n", | |
365 | "Keysize", "Add", "Lookup", "Lookup_bulk", "Delete"); | |
366 | for (i = 0; i < NUM_KEYSIZES; i++) { | |
367 | printf("%-18d", hashtest_key_lens[i]); | |
368 | for (j = 0; j < NUM_OPERATIONS; j++) | |
369 | printf("%-18"PRIu64, cycles[i][j]); | |
370 | printf("\n"); | |
371 | } | |
372 | return 0; | |
373 | } | |
374 | ||
375 | static int | |
376 | test_efd_perf(void) | |
377 | { | |
378 | ||
379 | if (run_all_tbl_perf_tests() < 0) | |
380 | return -1; | |
381 | ||
382 | return 0; | |
383 | } | |
384 | ||
385 | REGISTER_TEST_COMMAND(efd_perf_autotest, test_efd_perf); |