2 * Copyright 2017 Facebook, Inc.
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
17 #include <folly/ThreadCachedArena.h>
18 #include <folly/Memory.h>
26 #include <unordered_map>
28 #include <glog/logging.h>
30 #include <folly/Range.h>
31 #include <folly/Benchmark.h>
32 #include <folly/portability/GTest.h>
34 using namespace folly;
40 explicit ArenaTester(ThreadCachedArena& arena) : arena_(&arena) { }
42 void allocate(size_t count, size_t maxSize);
44 void merge(ArenaTester&& other);
47 std::mutex mergeMutex_;
48 std::vector<std::pair<uint8_t, Range<uint8_t*>>> areas_;
49 ThreadCachedArena* arena_;
52 void ArenaTester::allocate(size_t count, size_t maxSize) {
53 // Allocate chunks of memory of random sizes
55 std::uniform_int_distribution<uint32_t> sizeDist(1, maxSize - 1);
57 areas_.reserve(count);
58 for (size_t i = 0; i < count; i++) {
59 size_t size = sizeDist(rnd);
60 uint8_t* p = static_cast<uint8_t*>(arena_->allocate(size));
61 areas_.emplace_back(rnd() & 0xff, Range<uint8_t*>(p, size));
64 // Fill each area with a different value, to prove that they don't overlap
65 // Fill in random order.
67 areas_.begin(), areas_.end(),
68 [&rnd] (int n) -> int {
69 return std::uniform_int_distribution<uint32_t>(0, n-1)(rnd);
72 for (auto& p : areas_) {
73 std::fill(p.second.begin(), p.second.end(), p.first);
77 void ArenaTester::verify() {
78 for (auto& p : areas_) {
79 for (auto v : p.second) {
80 EXPECT_EQ(p.first, v);
85 void ArenaTester::merge(ArenaTester&& other) {
87 std::lock_guard<std::mutex> lock(mergeMutex_);
88 std::move(other.areas_.begin(), other.areas_.end(),
89 std::back_inserter(areas_));
96 TEST(ThreadCachedArena, BlockSize) {
97 static const size_t alignment = alignof(std::max_align_t);
98 static const size_t requestedBlockSize = 64;
100 ThreadCachedArena arena(requestedBlockSize);
101 size_t blockSize = alignment;
102 uint8_t* prev = static_cast<uint8_t*>(arena.allocate(1));
104 // Keep allocating until we're no longer one single alignment away from the
105 // previous allocation -- that's when we've gotten to the next block.
107 while ((p = static_cast<uint8_t*>(arena.allocate(1))) ==
110 blockSize += alignment;
113 VLOG(1) << "Requested block size: " << requestedBlockSize << ", actual: "
115 EXPECT_LE(requestedBlockSize, blockSize);
118 TEST(ThreadCachedArena, SingleThreaded) {
119 static const size_t requestedBlockSize = 64;
120 ThreadCachedArena arena(requestedBlockSize);
121 EXPECT_EQ(arena.totalSize(), sizeof(ThreadCachedArena));
123 ArenaTester tester(arena);
124 tester.allocate(100, 100 << 10);
127 EXPECT_GT(arena.totalSize(), sizeof(ThreadCachedArena));
130 TEST(ThreadCachedArena, MultiThreaded) {
131 static const size_t requestedBlockSize = 64;
132 ThreadCachedArena arena(requestedBlockSize);
133 ArenaTester mainTester(arena);
135 // Do this twice, to catch the possibility that memory from the first
137 static const size_t numThreads = 20;
138 for (size_t i = 0; i < 2; i++) {
139 std::vector<std::thread> threads;
140 threads.reserve(numThreads);
141 for (size_t j = 0; j < numThreads; j++) {
142 threads.emplace_back(
143 [&arena, &mainTester] () {
144 ArenaTester tester(arena);
145 tester.allocate(500, 1 << 10);
147 mainTester.merge(std::move(tester));
150 for (auto& t : threads) {
158 TEST(ThreadCachedArena, StlAllocator) {
159 typedef std::unordered_map<
160 int, int, std::hash<int>, std::equal_to<int>,
161 StlAllocator<ThreadCachedArena, std::pair<const int, int>>> Map;
163 static const size_t requestedBlockSize = 64;
164 ThreadCachedArena arena(requestedBlockSize);
166 Map map {0, std::hash<int>(), std::equal_to<int>(),
167 StlAllocator<ThreadCachedArena, std::pair<const int, int>>(&arena)};
169 for (int i = 0; i < 1000; i++) {
173 for (int i = 0; i < 1000; i++) {
174 EXPECT_EQ(i, map[i]);
180 static const int kNumValues = 10000;
182 BENCHMARK(bmUMStandard, iters) {
183 typedef std::unordered_map<int, int> Map;
187 for (int i = 0; i < kNumValues; i++) {
193 BENCHMARK(bmUMArena, iters) {
194 typedef std::unordered_map<
195 int, int, std::hash<int>, std::equal_to<int>,
196 StlAllocator<ThreadCachedArena, std::pair<const int, int>>> Map;
199 ThreadCachedArena arena;
201 Map map {0, std::hash<int>(), std::equal_to<int>(),
202 StlAllocator<ThreadCachedArena, std::pair<const int, int>>(
205 for (int i = 0; i < kNumValues; i++) {
211 BENCHMARK_DRAW_LINE()
213 BENCHMARK(bmMStandard, iters) {
214 typedef std::map<int, int> Map;
218 for (int i = 0; i < kNumValues; i++) {
224 BENCHMARK_DRAW_LINE()
226 BENCHMARK(bmMArena, iters) {
228 int, int, std::less<int>,
229 StlAllocator<ThreadCachedArena, std::pair<const int, int>>> Map;
232 ThreadCachedArena arena;
234 Map map {std::less<int>(),
235 StlAllocator<ThreadCachedArena, std::pair<const int, int>>(
238 for (int i = 0; i < kNumValues; i++) {
244 BENCHMARK_DRAW_LINE()
249 // Benchmark Iters Total t t/iter iter/sec
250 // ----------------------------------------------------------------------------
251 // Comparing benchmarks: bmUMStandard,bmUMArena
252 // + 143% bmUMStandard 1570 2.005 s 1.277 ms 782.9
253 // * bmUMArena 3817 2.003 s 524.7 us 1.861 k
254 // ----------------------------------------------------------------------------
255 // Comparing benchmarks: bmMStandard,bmMArena
256 // +79.0% bmMStandard 1197 2.009 s 1.678 ms 595.8
257 // * bmMArena 2135 2.002 s 937.6 us 1.042 k
258 // ----------------------------------------------------------------------------
260 int main(int argc, char *argv[]) {
261 testing::InitGoogleTest(&argc, argv);
262 gflags::ParseCommandLineFlags(&argc, &argv, true);
263 auto ret = RUN_ALL_TESTS();
264 if (!ret && FLAGS_benchmark) {
265 folly::runBenchmarks();