1 // Copyright (c) 2012-2020 The Bitcoin Core developers
2 // Distributed under the MIT software license, see the accompanying
3 // file COPYING or http://www.opensource.org/licenses/mit-license.php.
4 #include <cuckoocache.h>
5 #include <random.h>
6 #include <script/sigcache.h>
7 #include <test/util/setup_common.h>
8
9 #include <boost/test/unit_test.hpp>
10
11 #include <deque>
12 #include <mutex>
13 #include <shared_mutex>
14 #include <thread>
15 #include <vector>
16
17 /** Test Suite for CuckooCache
18 *
19 * 1. All tests should have a deterministic result (using insecure rand
20 * with deterministic seeds)
21 * 2. Some test methods are templated to allow for easier testing
22 * against new versions / comparing
23 * 3. Results should be treated as a regression test, i.e., did the behavior
24 * change significantly from what was expected. This can be OK, depending on
25 * the nature of the change, but requires updating the tests to reflect the new
26 * expected behavior. For example improving the hit rate may cause some tests
27 * using BOOST_CHECK_CLOSE to fail.
28 *
29 */
30 BOOST_AUTO_TEST_SUITE(cuckoocache_tests);
31
32 /* Test that no values not inserted into the cache are read out of it.
33 *
34 * There are no repeats in the first 200000 insecure_GetRandHash calls
35 */
BOOST_AUTO_TEST_CASE(test_cuckoocache_no_fakes)36 BOOST_AUTO_TEST_CASE(test_cuckoocache_no_fakes)
37 {
38 SeedInsecureRand(SeedRand::ZEROS);
39 CuckooCache::cache<uint256, SignatureCacheHasher> cc{};
40 size_t megabytes = 4;
41 cc.setup_bytes(megabytes << 20);
42 for (int x = 0; x < 100000; ++x) {
43 cc.insert(InsecureRand256());
44 }
45 for (int x = 0; x < 100000; ++x) {
46 BOOST_CHECK(!cc.contains(InsecureRand256(), false));
47 }
48 };
49
50 /** This helper returns the hit rate when megabytes*load worth of entries are
51 * inserted into a megabytes sized cache
52 */
53 template <typename Cache>
test_cache(size_t megabytes,double load)54 static double test_cache(size_t megabytes, double load)
55 {
56 SeedInsecureRand(SeedRand::ZEROS);
57 std::vector<uint256> hashes;
58 Cache set{};
59 size_t bytes = megabytes * (1 << 20);
60 set.setup_bytes(bytes);
61 uint32_t n_insert = static_cast<uint32_t>(load * (bytes / sizeof(uint256)));
62 hashes.resize(n_insert);
63 for (uint32_t i = 0; i < n_insert; ++i) {
64 uint32_t* ptr = (uint32_t*)hashes[i].begin();
65 for (uint8_t j = 0; j < 8; ++j)
66 *(ptr++) = InsecureRand32();
67 }
68 /** We make a copy of the hashes because future optimizations of the
69 * cuckoocache may overwrite the inserted element, so the test is
70 * "future proofed".
71 */
72 std::vector<uint256> hashes_insert_copy = hashes;
73 /** Do the insert */
74 for (const uint256& h : hashes_insert_copy)
75 set.insert(h);
76 /** Count the hits */
77 uint32_t count = 0;
78 for (const uint256& h : hashes)
79 count += set.contains(h, false);
80 double hit_rate = ((double)count) / ((double)n_insert);
81 return hit_rate;
82 }
83
84 /** The normalized hit rate for a given load.
85 *
86 * The semantics are a little confusing, so please see the below
87 * explanation.
88 *
89 * Examples:
90 *
91 * 1. at load 0.5, we expect a perfect hit rate, so we multiply by
92 * 1.0
93 * 2. at load 2.0, we expect to see half the entries, so a perfect hit rate
94 * would be 0.5. Therefore, if we see a hit rate of 0.4, 0.4*2.0 = 0.8 is the
95 * normalized hit rate.
96 *
97 * This is basically the right semantics, but has a bit of a glitch depending on
98 * how you measure around load 1.0 as after load 1.0 your normalized hit rate
99 * becomes effectively perfect, ignoring freshness.
100 */
normalize_hit_rate(double hits,double load)101 static double normalize_hit_rate(double hits, double load)
102 {
103 return hits * std::max(load, 1.0);
104 }
105
106 /** Check the hit rate on loads ranging from 0.1 to 1.6 */
BOOST_AUTO_TEST_CASE(cuckoocache_hit_rate_ok)107 BOOST_AUTO_TEST_CASE(cuckoocache_hit_rate_ok)
108 {
109 /** Arbitrarily selected Hit Rate threshold that happens to work for this test
110 * as a lower bound on performance.
111 */
112 double HitRateThresh = 0.98;
113 size_t megabytes = 4;
114 for (double load = 0.1; load < 2; load *= 2) {
115 double hits = test_cache<CuckooCache::cache<uint256, SignatureCacheHasher>>(megabytes, load);
116 BOOST_CHECK(normalize_hit_rate(hits, load) > HitRateThresh);
117 }
118 }
119
120
121 /** This helper checks that erased elements are preferentially inserted onto and
122 * that the hit rate of "fresher" keys is reasonable*/
123 template <typename Cache>
test_cache_erase(size_t megabytes)124 static void test_cache_erase(size_t megabytes)
125 {
126 double load = 1;
127 SeedInsecureRand(SeedRand::ZEROS);
128 std::vector<uint256> hashes;
129 Cache set{};
130 size_t bytes = megabytes * (1 << 20);
131 set.setup_bytes(bytes);
132 uint32_t n_insert = static_cast<uint32_t>(load * (bytes / sizeof(uint256)));
133 hashes.resize(n_insert);
134 for (uint32_t i = 0; i < n_insert; ++i) {
135 uint32_t* ptr = (uint32_t*)hashes[i].begin();
136 for (uint8_t j = 0; j < 8; ++j)
137 *(ptr++) = InsecureRand32();
138 }
139 /** We make a copy of the hashes because future optimizations of the
140 * cuckoocache may overwrite the inserted element, so the test is
141 * "future proofed".
142 */
143 std::vector<uint256> hashes_insert_copy = hashes;
144
145 /** Insert the first half */
146 for (uint32_t i = 0; i < (n_insert / 2); ++i)
147 set.insert(hashes_insert_copy[i]);
148 /** Erase the first quarter */
149 for (uint32_t i = 0; i < (n_insert / 4); ++i)
150 BOOST_CHECK(set.contains(hashes[i], true));
151 /** Insert the second half */
152 for (uint32_t i = (n_insert / 2); i < n_insert; ++i)
153 set.insert(hashes_insert_copy[i]);
154
155 /** elements that we marked as erased but are still there */
156 size_t count_erased_but_contained = 0;
157 /** elements that we did not erase but are older */
158 size_t count_stale = 0;
159 /** elements that were most recently inserted */
160 size_t count_fresh = 0;
161
162 for (uint32_t i = 0; i < (n_insert / 4); ++i)
163 count_erased_but_contained += set.contains(hashes[i], false);
164 for (uint32_t i = (n_insert / 4); i < (n_insert / 2); ++i)
165 count_stale += set.contains(hashes[i], false);
166 for (uint32_t i = (n_insert / 2); i < n_insert; ++i)
167 count_fresh += set.contains(hashes[i], false);
168
169 double hit_rate_erased_but_contained = double(count_erased_but_contained) / (double(n_insert) / 4.0);
170 double hit_rate_stale = double(count_stale) / (double(n_insert) / 4.0);
171 double hit_rate_fresh = double(count_fresh) / (double(n_insert) / 2.0);
172
173 // Check that our hit_rate_fresh is perfect
174 BOOST_CHECK_EQUAL(hit_rate_fresh, 1.0);
175 // Check that we have a more than 2x better hit rate on stale elements than
176 // erased elements.
177 BOOST_CHECK(hit_rate_stale > 2 * hit_rate_erased_but_contained);
178 }
179
BOOST_AUTO_TEST_CASE(cuckoocache_erase_ok)180 BOOST_AUTO_TEST_CASE(cuckoocache_erase_ok)
181 {
182 size_t megabytes = 4;
183 test_cache_erase<CuckooCache::cache<uint256, SignatureCacheHasher>>(megabytes);
184 }
185
186 template <typename Cache>
test_cache_erase_parallel(size_t megabytes)187 static void test_cache_erase_parallel(size_t megabytes)
188 {
189 double load = 1;
190 SeedInsecureRand(SeedRand::ZEROS);
191 std::vector<uint256> hashes;
192 Cache set{};
193 size_t bytes = megabytes * (1 << 20);
194 set.setup_bytes(bytes);
195 uint32_t n_insert = static_cast<uint32_t>(load * (bytes / sizeof(uint256)));
196 hashes.resize(n_insert);
197 for (uint32_t i = 0; i < n_insert; ++i) {
198 uint32_t* ptr = (uint32_t*)hashes[i].begin();
199 for (uint8_t j = 0; j < 8; ++j)
200 *(ptr++) = InsecureRand32();
201 }
202 /** We make a copy of the hashes because future optimizations of the
203 * cuckoocache may overwrite the inserted element, so the test is
204 * "future proofed".
205 */
206 std::vector<uint256> hashes_insert_copy = hashes;
207 std::shared_mutex mtx;
208
209 {
210 /** Grab lock to make sure we release inserts */
211 std::unique_lock<std::shared_mutex> l(mtx);
212 /** Insert the first half */
213 for (uint32_t i = 0; i < (n_insert / 2); ++i)
214 set.insert(hashes_insert_copy[i]);
215 }
216
217 /** Spin up 3 threads to run contains with erase.
218 */
219 std::vector<std::thread> threads;
220 /** Erase the first quarter */
221 for (uint32_t x = 0; x < 3; ++x)
222 /** Each thread is emplaced with x copy-by-value
223 */
224 threads.emplace_back([&, x] {
225 std::shared_lock<std::shared_mutex> l(mtx);
226 size_t ntodo = (n_insert/4)/3;
227 size_t start = ntodo*x;
228 size_t end = ntodo*(x+1);
229 for (uint32_t i = start; i < end; ++i) {
230 bool contains = set.contains(hashes[i], true);
231 assert(contains);
232 }
233 });
234
235 /** Wait for all threads to finish
236 */
237 for (std::thread& t : threads)
238 t.join();
239 /** Grab lock to make sure we observe erases */
240 std::unique_lock<std::shared_mutex> l(mtx);
241 /** Insert the second half */
242 for (uint32_t i = (n_insert / 2); i < n_insert; ++i)
243 set.insert(hashes_insert_copy[i]);
244
245 /** elements that we marked erased but that are still there */
246 size_t count_erased_but_contained = 0;
247 /** elements that we did not erase but are older */
248 size_t count_stale = 0;
249 /** elements that were most recently inserted */
250 size_t count_fresh = 0;
251
252 for (uint32_t i = 0; i < (n_insert / 4); ++i)
253 count_erased_but_contained += set.contains(hashes[i], false);
254 for (uint32_t i = (n_insert / 4); i < (n_insert / 2); ++i)
255 count_stale += set.contains(hashes[i], false);
256 for (uint32_t i = (n_insert / 2); i < n_insert; ++i)
257 count_fresh += set.contains(hashes[i], false);
258
259 double hit_rate_erased_but_contained = double(count_erased_but_contained) / (double(n_insert) / 4.0);
260 double hit_rate_stale = double(count_stale) / (double(n_insert) / 4.0);
261 double hit_rate_fresh = double(count_fresh) / (double(n_insert) / 2.0);
262
263 // Check that our hit_rate_fresh is perfect
264 BOOST_CHECK_EQUAL(hit_rate_fresh, 1.0);
265 // Check that we have a more than 2x better hit rate on stale elements than
266 // erased elements.
267 BOOST_CHECK(hit_rate_stale > 2 * hit_rate_erased_but_contained);
268 }
BOOST_AUTO_TEST_CASE(cuckoocache_erase_parallel_ok)269 BOOST_AUTO_TEST_CASE(cuckoocache_erase_parallel_ok)
270 {
271 size_t megabytes = 4;
272 test_cache_erase_parallel<CuckooCache::cache<uint256, SignatureCacheHasher>>(megabytes);
273 }
274
275
276 template <typename Cache>
test_cache_generations()277 static void test_cache_generations()
278 {
279 // This test checks that for a simulation of network activity, the fresh hit
280 // rate is never below 99%, and the number of times that it is worse than
281 // 99.9% are less than 1% of the time.
282 double min_hit_rate = 0.99;
283 double tight_hit_rate = 0.999;
284 double max_rate_less_than_tight_hit_rate = 0.01;
285 // A cache that meets this specification is therefore shown to have a hit
286 // rate of at least tight_hit_rate * (1 - max_rate_less_than_tight_hit_rate) +
287 // min_hit_rate*max_rate_less_than_tight_hit_rate = 0.999*99%+0.99*1% == 99.89%
288 // hit rate with low variance.
289
290 // We use deterministic values, but this test has also passed on many
291 // iterations with non-deterministic values, so it isn't "overfit" to the
292 // specific entropy in FastRandomContext(true) and implementation of the
293 // cache.
294 SeedInsecureRand(SeedRand::ZEROS);
295
296 // block_activity models a chunk of network activity. n_insert elements are
297 // added to the cache. The first and last n/4 are stored for removal later
298 // and the middle n/2 are not stored. This models a network which uses half
299 // the signatures of recently (since the last block) added transactions
300 // immediately and never uses the other half.
301 struct block_activity {
302 std::vector<uint256> reads;
303 block_activity(uint32_t n_insert, Cache& c) : reads()
304 {
305 std::vector<uint256> inserts;
306 inserts.resize(n_insert);
307 reads.reserve(n_insert / 2);
308 for (uint32_t i = 0; i < n_insert; ++i) {
309 uint32_t* ptr = (uint32_t*)inserts[i].begin();
310 for (uint8_t j = 0; j < 8; ++j)
311 *(ptr++) = InsecureRand32();
312 }
313 for (uint32_t i = 0; i < n_insert / 4; ++i)
314 reads.push_back(inserts[i]);
315 for (uint32_t i = n_insert - (n_insert / 4); i < n_insert; ++i)
316 reads.push_back(inserts[i]);
317 for (const auto& h : inserts)
318 c.insert(h);
319 }
320 };
321
322 const uint32_t BLOCK_SIZE = 1000;
323 // We expect window size 60 to perform reasonably given that each epoch
324 // stores 45% of the cache size (~472k).
325 const uint32_t WINDOW_SIZE = 60;
326 const uint32_t POP_AMOUNT = (BLOCK_SIZE / WINDOW_SIZE) / 2;
327 const double load = 10;
328 const size_t megabytes = 4;
329 const size_t bytes = megabytes * (1 << 20);
330 const uint32_t n_insert = static_cast<uint32_t>(load * (bytes / sizeof(uint256)));
331
332 std::vector<block_activity> hashes;
333 Cache set{};
334 set.setup_bytes(bytes);
335 hashes.reserve(n_insert / BLOCK_SIZE);
336 std::deque<block_activity> last_few;
337 uint32_t out_of_tight_tolerance = 0;
338 uint32_t total = n_insert / BLOCK_SIZE;
339 // we use the deque last_few to model a sliding window of blocks. at each
340 // step, each of the last WINDOW_SIZE block_activities checks the cache for
341 // POP_AMOUNT of the hashes that they inserted, and marks these erased.
342 for (uint32_t i = 0; i < total; ++i) {
343 if (last_few.size() == WINDOW_SIZE)
344 last_few.pop_front();
345 last_few.emplace_back(BLOCK_SIZE, set);
346 uint32_t count = 0;
347 for (auto& act : last_few)
348 for (uint32_t k = 0; k < POP_AMOUNT; ++k) {
349 count += set.contains(act.reads.back(), true);
350 act.reads.pop_back();
351 }
352 // We use last_few.size() rather than WINDOW_SIZE for the correct
353 // behavior on the first WINDOW_SIZE iterations where the deque is not
354 // full yet.
355 double hit = (double(count)) / (last_few.size() * POP_AMOUNT);
356 // Loose Check that hit rate is above min_hit_rate
357 BOOST_CHECK(hit > min_hit_rate);
358 // Tighter check, count number of times we are less than tight_hit_rate
359 // (and implicitly, greater than min_hit_rate)
360 out_of_tight_tolerance += hit < tight_hit_rate;
361 }
362 // Check that being out of tolerance happens less than
363 // max_rate_less_than_tight_hit_rate of the time
364 BOOST_CHECK(double(out_of_tight_tolerance) / double(total) < max_rate_less_than_tight_hit_rate);
365 }
BOOST_AUTO_TEST_CASE(cuckoocache_generations)366 BOOST_AUTO_TEST_CASE(cuckoocache_generations)
367 {
368 test_cache_generations<CuckooCache::cache<uint256, SignatureCacheHasher>>();
369 }
370
371 BOOST_AUTO_TEST_SUITE_END();
372