|
| 1 | +#include "../../hnswlib/hnswlib.h" |
| 2 | + |
| 3 | +#include <assert.h> |
| 4 | + |
| 5 | +#include <vector> |
| 6 | +#include <iostream> |
| 7 | +#include <cstdio> |
| 8 | +#include <thread> |
| 9 | +#include <chrono> |
| 10 | + |
| 11 | +namespace { |
| 12 | + |
| 13 | +const size_t M = 32; |
| 14 | +const size_t ef_construction = 500; |
| 15 | +const size_t random_seed = 100; |
| 16 | +const bool allow_replace_deleted = false; |
| 17 | + |
| 18 | +const size_t dimension = 1024; |
| 19 | +const size_t total_items = 100 * 10000; |
| 20 | +const size_t num_query = 500 * 10000; |
| 21 | +size_t topk = 10; |
| 22 | +const size_t max_thread_num = 48; |
| 23 | +const std::string index_path = "./hnsw.index"; |
| 24 | + |
| 25 | +std::vector<float> data(total_items * dimension); |
| 26 | +std::vector<float> query(num_query * dimension); |
| 27 | + |
| 28 | + |
| 29 | +void check_knn_closer(hnswlib::AlgorithmInterface<float>* alg_hnsw) { |
| 30 | + for (size_t j = 0; j < num_query; ++j) { |
| 31 | + const void* p = query.data() + j * dimension; |
| 32 | + auto gd = alg_hnsw->searchKnn(p, topk); |
| 33 | + auto res = alg_hnsw->searchKnnCloserFirst(p, topk); |
| 34 | + assert(gd.size() == res.size()); |
| 35 | + size_t t = gd.size(); |
| 36 | + while (!gd.empty()) { |
| 37 | + assert(gd.top() == res[--t]); |
| 38 | + gd.pop(); |
| 39 | + } |
| 40 | + } |
| 41 | + std::cout << "test hnsw search knn closer first success..." << std::endl; |
| 42 | +} |
| 43 | + |
| 44 | +void test_compatibility(bool hnsw_first_use_blocks_memory, |
| 45 | + bool hnsw_second_use_blocks_memory) { |
| 46 | + |
| 47 | + std::cout << "================== test compatibility ==================" << std::endl; |
| 48 | + hnswlib::L2Space space(dimension); |
| 49 | + hnswlib::AlgorithmInterface<float>* alg_hnsw_first = new hnswlib::HierarchicalNSW<float>(&space, 2 * total_items, |
| 50 | + M, ef_construction, random_seed, allow_replace_deleted, hnsw_first_use_blocks_memory); |
| 51 | + |
| 52 | + for (size_t i = 0; i < total_items; ++i) { |
| 53 | + alg_hnsw_first->addPoint(data.data() + dimension * i, i); |
| 54 | + } |
| 55 | + check_knn_closer(alg_hnsw_first); |
| 56 | + |
| 57 | + // save hnsw index |
| 58 | + std::remove(index_path.data()); |
| 59 | + alg_hnsw_first->saveIndex(index_path); |
| 60 | + std::cout << "save hnsw(use_small_blocks_memory = " << hnsw_first_use_blocks_memory << ") index success" << std::endl; |
| 61 | + delete alg_hnsw_first; |
| 62 | + |
| 63 | + // load hnsw index |
| 64 | + hnswlib::AlgorithmInterface<float>* alg_hnsw_second = new hnswlib::HierarchicalNSW<float>(&space, false, |
| 65 | + 0, allow_replace_deleted, hnsw_second_use_blocks_memory); |
| 66 | + std::cout << "load hnsw(use_small_blocks_memory = " << hnsw_second_use_blocks_memory << ") index success" << std::endl; |
| 67 | + std::remove(index_path.data()); |
| 68 | + check_knn_closer(alg_hnsw_second); |
| 69 | + |
| 70 | + delete alg_hnsw_second; |
| 71 | +} |
| 72 | + |
| 73 | +void test_performace(bool use_small_blocks_memory) { |
| 74 | + if (total_items == 0) { |
| 75 | + return; |
| 76 | + } |
| 77 | + |
| 78 | + std::cout << "================== test preformace(" |
| 79 | + << "dimension: " << dimension |
| 80 | + << ", M: " << M |
| 81 | + << ", ef_construction: " << ef_construction |
| 82 | + << ", topk: " << topk |
| 83 | + << ", use_small_blocks_memory: " << (use_small_blocks_memory ? "ture" : "false" ) |
| 84 | + << ") ==================" << std::endl; |
| 85 | + hnswlib::L2Space space(dimension); |
| 86 | + hnswlib::HierarchicalNSW<float>* alg_hnsw = new hnswlib::HierarchicalNSW<float>(&space, 2 * total_items, |
| 87 | + M, ef_construction, random_seed, allow_replace_deleted, use_small_blocks_memory); |
| 88 | + |
| 89 | + std::vector<std::thread> threads; |
| 90 | + size_t num_threads = (total_items >= max_thread_num ? max_thread_num : total_items); |
| 91 | + size_t batch_num = (total_items / (num_threads <= 1 ? 1 : (num_threads - 1))) + 1; |
| 92 | + auto start_time = std::chrono::system_clock::now(); |
| 93 | + for (size_t idx = 0; idx < total_items; idx += batch_num) { |
| 94 | + size_t start = idx; |
| 95 | + size_t end = std::min(idx + batch_num, total_items); |
| 96 | + threads.push_back( |
| 97 | + std::thread( |
| 98 | + [alg_hnsw, start, end] { |
| 99 | + for (size_t i = start; i < end; i++) { |
| 100 | + alg_hnsw->addPoint(data.data() + i * dimension, i); |
| 101 | + } |
| 102 | + } |
| 103 | + ) |
| 104 | + ); |
| 105 | + } |
| 106 | + for (auto &thread : threads) { |
| 107 | + thread.join(); |
| 108 | + } |
| 109 | + threads.clear(); |
| 110 | + auto end_time = std::chrono::system_clock::now(); |
| 111 | + double duration_in_ms = std::chrono::duration_cast<std::chrono::milliseconds>(end_time - start_time).count(); |
| 112 | + double duration_in_seconds = static_cast<double>((std::chrono::duration_cast<std::chrono::milliseconds>(end_time - start_time)).count()) / 1000.0; |
| 113 | + size_t qps = (duration_in_seconds == 0 ? total_items : total_items / duration_in_seconds); |
| 114 | + double latency = (total_items == 0 ? 0 : duration_in_ms / total_items); |
| 115 | + std::cout << "Start " << num_threads << " thread to add " << total_items << " items to hnsw index, cost " |
| 116 | + << duration_in_seconds << " seconds, qps: " << qps << ", latency: " << latency << "ms" << std::endl; |
| 117 | + |
| 118 | + |
| 119 | + num_threads = (num_query >= max_thread_num ? max_thread_num : num_query); |
| 120 | + batch_num = (num_query / (num_threads <= 1 ? 1 : (num_threads - 1))) + 1; |
| 121 | + start_time = std::chrono::system_clock::now(); |
| 122 | + for (size_t idx = 0; idx < num_query; idx += batch_num) { |
| 123 | + size_t start = idx; |
| 124 | + size_t end = std::min(idx + batch_num, num_query); |
| 125 | + threads.push_back( |
| 126 | + std::thread( |
| 127 | + [alg_hnsw, start, end] { |
| 128 | + for (size_t i = start; i < end; i++) { |
| 129 | + const void* p = query.data() + i * dimension; |
| 130 | + auto gd = alg_hnsw->searchKnn(p, topk); |
| 131 | + } |
| 132 | + } |
| 133 | + ) |
| 134 | + ); |
| 135 | + } |
| 136 | + for (auto &thread : threads) { |
| 137 | + thread.join(); |
| 138 | + } |
| 139 | + threads.clear(); |
| 140 | + end_time = std::chrono::system_clock::now(); |
| 141 | + duration_in_ms = std::chrono::duration_cast<std::chrono::milliseconds>(end_time - start_time).count(); |
| 142 | + duration_in_seconds = static_cast<double>((std::chrono::duration_cast<std::chrono::milliseconds>(end_time - start_time)).count()) / 1000.0; |
| 143 | + qps = (duration_in_seconds == 0 ? num_query : num_query / duration_in_seconds); |
| 144 | + latency = (num_query == 0 ? 0 : duration_in_ms / num_query); |
| 145 | + std::cout << "Start " << num_threads << " thread to exec " << num_query << " searchKnn, cost " |
| 146 | + << duration_in_seconds << " seconds, qps: " << qps << ", latency: " << latency << "ms" << std::endl; |
| 147 | + |
| 148 | + delete alg_hnsw; |
| 149 | +} |
| 150 | + |
| 151 | +} // namespace |
| 152 | + |
| 153 | +int main() { |
| 154 | + |
| 155 | + std::mt19937 rng; |
| 156 | + rng.seed(47); |
| 157 | + std::uniform_real_distribution<> distrib; |
| 158 | + |
| 159 | + for (size_t i = 0; i < total_items * dimension; ++i) { |
| 160 | + data[i] = distrib(rng); |
| 161 | + } |
| 162 | + for (size_t i = 0; i < num_query * dimension; ++i) { |
| 163 | + query[i] = distrib(rng); |
| 164 | + } |
| 165 | + |
| 166 | + test_compatibility(true, true); |
| 167 | + test_compatibility(false, false); |
| 168 | + test_compatibility(true, false); |
| 169 | + test_compatibility(false, true); |
| 170 | + |
| 171 | + test_performace(true); |
| 172 | + test_performace(false); |
| 173 | + |
| 174 | + return 0; |
| 175 | +} |
0 commit comments