为何tbb::concurrent_hash_map读写慢于std::unordered_map？求高效替代方案

核心问题 #

为什么使用tbb::parallel_for后，tbb::concurrent_hash_map的读写性能仍低于std::unordered_map？尤其是读取耗时表现差距明显。

耗时统计结果 #

std::unordered_map time：267ms
-------------------find time：55ms

tbb::concurrent_hash_map tbb parrallel time：295ms
-------------------find time：235ms

测试代码 #

std::unordered_map 测试代码 #

const int kNumElements = 10'000'000;
// TODO: 1. test unordered_map
std::unordered_map<int, int> map_int_unordered(kNumElements);
auto start_time = std::chrono::steady_clock::now();
for (int i = 0; i < kNumElements; i++)
{
    map_int_unordered[i] = i;
}
auto end_time = std::chrono::steady_clock::now();
auto duration = std::chrono::duration_cast<std::chrono::milliseconds>(end_time - start_time);
std::cout << DEBUG << "1. std::unordered_map time：" << duration.count() << "ms" << std::endl;
// print_map(map_int_unordered);

start_time = std::chrono::steady_clock::now();
for (int i = 0; i < kNumElements; i++)
{
    map_int_unordered.find(i);
}
end_time = std::chrono::steady_clock::now();
duration = std::chrono::duration_cast<std::chrono::milliseconds>(end_time - start_time);
std::cout << GREEN << " -------------------find time：" << duration.count() << "ms" << std::endl;

tbb::concurrent_hash_map 测试代码 #

tbb::concurrent_hash_map<int, int> map_tbb2(kNumElements);
start_time = std::chrono::steady_clock::now();
tbb::parallel_for(0, kNumElements, [&](int i)
                  { map_tbb2.insert(std::make_pair(i, i)); });
end_time = std::chrono::steady_clock::now();
duration = std::chrono::duration_cast<std::chrono::milliseconds>(end_time - start_time);
std::cout << DEBUG << "3. tbb::concurrent_hash_map tbb parrallel time：" << duration.count() << "ms" << std::endl;

start_time = std::chrono::steady_clock::now();
tbb::parallel_for(0, kNumElements, [&](int i)
                  {  
                    tbb::concurrent_hash_map<int, int>::accessor acc;
                    map.find(acc,i); });
end_time = std::chrono::steady_clock::now();
duration = std::chrono::duration_cast<std::chrono::milliseconds>(end_time - start_time);
std::cout << GREEN << " -------------------find time：" << duration.count() << "ms" << std::endl;

额外问题与测试结果 #

是否存在比std::unordered_map更快的哈希表？已尝试ankerl/unordered_dense、tsl/robin_map，测试结果如下：

tsl::robin_map time：224ms
-------------------find time：21ms
tsl::robin_map multi-threads time：133ms
-------------------find time：17ms
ankerl::unordered_dense::map time：338ms
-------------------find time：270ms

尽管tsl::robin_map性能更优，但在自定义类中使用时速度大幅下降。

内容的提问来源于stack exchange，提问作者YZH