Multithread test to check udate of deleted elements

.github/workflows/build.yml

@@ -62,6 +62,7 @@ jobs:
           ./searchKnnCloserFirst_test
           ./searchKnnWithFilter_test
           ./multiThreadLoad_test
+          ./multiThread_replace_test
           ./test_updates
           ./test_updates update
         shell: bash

CMakeLists.txt

@@ -9,7 +9,7 @@ if(CMAKE_PROJECT_NAME STREQUAL PROJECT_NAME)
     set(CMAKE_CXX_STANDARD 11)
 
     if (CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
-      SET( CMAKE_CXX_FLAGS  "-Ofast -DNDEBUG -std=c++11 -DHAVE_CXX0X -openmp -march=native -fpic -ftree-vectorize")
+      SET( CMAKE_CXX_FLAGS  "-Ofast -DNDEBUG -std=c++11 -DHAVE_CXX0X -openmp -mcpu=apple-m1 -fpic -ftree-vectorize")
     elseif (CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
       SET( CMAKE_CXX_FLAGS  "-Ofast -lrt -DNDEBUG -std=c++11 -DHAVE_CXX0X -march=native -fpic -w -fopenmp -ftree-vectorize -ftree-vectorizer-verbose=0" )
     elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
@@ -28,6 +28,9 @@ if(CMAKE_PROJECT_NAME STREQUAL PROJECT_NAME)
     add_executable(multiThreadLoad_test examples/multiThreadLoad_test.cpp)
     target_link_libraries(multiThreadLoad_test hnswlib)
 
+    add_executable(multiThread_replace_test examples/multiThread_replace_test.cpp)
+    target_link_libraries(multiThread_replace_test hnswlib)
+
     add_executable(main main.cpp sift_1b.cpp)
     target_link_libraries(main hnswlib)
 endif()

examples/multiThread_replace_test.cpp

@@ -0,0 +1,118 @@
+#include "../hnswlib/hnswlib.h"
+#include <thread>
+#include <chrono>
+
+
+template<class Function>
+inline void ParallelFor(size_t start, size_t end, size_t numThreads, Function fn) {
+    if (numThreads <= 0) {
+        numThreads = std::thread::hardware_concurrency();
+    }
+
+    if (numThreads == 1) {
+        for (size_t id = start; id < end; id++) {
+            fn(id, 0);
+        }
+    } else {
+        std::vector<std::thread> threads;
+        std::atomic<size_t> current(start);
+
+        // keep track of exceptions in threads
+        // https://stackoverflow.com/a/32428427/1713196
+        std::exception_ptr lastException = nullptr;
+        std::mutex lastExceptMutex;
+
+        for (size_t threadId = 0; threadId < numThreads; ++threadId) {
+            threads.push_back(std::thread([&, threadId] {
+                while (true) {
+                    size_t id = current.fetch_add(1);
+
+                    if (id >= end) {
+                        break;
+                    }
+
+                    try {
+                        fn(id, threadId);
+                    } catch (...) {
+                        std::unique_lock<std::mutex> lastExcepLock(lastExceptMutex);
+                        lastException = std::current_exception();
+                        /*
+                         * This will work even when current is the largest value that
+                         * size_t can fit, because fetch_add returns the previous value
+                         * before the increment (what will result in overflow
+                         * and produce 0 instead of current + 1).
+                         */
+                        current = end;
+                        break;
+                    }
+                }
+            }));
+        }
+        for (auto &thread : threads) {
+            thread.join();
+        }
+        if (lastException) {
+            std::rethrow_exception(lastException);
+        }
+    }
+}
+
+
+int main() {
+    std::cout << "Running multithread load test" << std::endl;
+    int d = 16;
+    int num_elements = 1000;
+    int max_elements = 2 * num_elements;
+    int num_threads = 50;
+
+    std::mt19937 rng;
+    rng.seed(47);
+    std::uniform_real_distribution<> distrib_real;
+
+    hnswlib::L2Space space(d);
+
+    float batch1[d * max_elements];
+    for (int i = 0; i < d * max_elements; i++) {
+        batch1[i] = distrib_real(rng);
+    }
+    float batch2[d * num_elements];
+    for (int i = 0; i < d * num_elements; i++) {
+        batch2[i] = distrib_real(rng);
+    }
+
+    std::vector<int> rand_labels(max_elements);
+    for (int i = 0; i < max_elements; i++) {
+        rand_labels[i] = i;
+    }
+    std::shuffle(rand_labels.begin(), rand_labels.end(), rng);
+
+    int iter = 0;
+    while (iter < 1000) {
+        std::cout << iter << std::endl;
+
+        hnswlib::HierarchicalNSW<float>* alg_hnsw = new hnswlib::HierarchicalNSW<float>(&space, max_elements, 16, 200, 123, true);
+
+        std::cout << "Building index\n";
+        ParallelFor(0, max_elements, num_threads, [&](size_t row, size_t threadId) {
+            alg_hnsw->addPoint((void*)(batch1 + d * row), row);
+        });
+
+        std::cout << "Deleting\n";
+        for (int i = 0; i < num_elements; i++) {
+            alg_hnsw->markDelete(rand_labels[i]);
+        }
+
+        std::cout << "Updating elements\n";
+        ParallelFor(0, num_elements, num_threads, [&](size_t row, size_t threadId) {
+            int label = rand_labels[row] + 10000;
+            alg_hnsw->addPoint((void*)(batch2 + d * row), label, true);
+        });
+
+        iter += 1;
+
+        delete alg_hnsw;
+    }
+
+    std::cout << "Finish" << std::endl;
+    return 0;
+}