加载/清除大量数据时，std::vector 变得越来越慢答案

【问题标题】：std::vector get slower and slower when load/clear huge amount of data加载/清除大量数据时，std::vector 变得越来越慢
【发布时间】：2016-11-24 10:25:47
【问题描述】：

问题

我有一个相当复杂的图像处理应用程序，其中一个子模块需要将巨大的二进制位图加载到内存中。实际上高达 96 GB（意味着 888 888 x 888 888 像素图像）。磁盘是 2xSSD raid0，读/写速度约为 1 GB/s。它将图像加载到一个向量（每个元素代表位图中的一行）中，该向量由智能指针指向带有字节的向量（每个元素代表 8 个像素）。这里奇怪的问题是，在重复加载和清除向量之后（我看到内存实际上是被填充和清除而没有内存泄漏），每次迭代似乎需要越来越长的时间。专门清理内存需要很长时间。

测试

我做了一些简单的测试应用程序来测试这个孤立的和不同的角度。用原始指针替换智能指针会产生同样奇怪的行为。然后我尝试使用本机数组而不是向量，这成功了。在加载/清除 100 次迭代后，使用向量时 24 GB 时间急剧增加，而数组实现在时间上是稳定的。下面是使用 24 GB 垃圾填充内存的测试应用程序，而不是加载实际图像，结果相同。在具有 128 GB RAM 并使用 Visual Studio 2013 Update 5 构建的 Windows 10 Pro 上完成的测试。

此函数使用向量进行加载/清除：

void SimpleLoadAndClear_Vector(int width, int height) {
    time_t start_time, end_time;

    // Load memory
    time(&start_time);
    cout << "Loading image into memory...";
    auto width_bytes = width / 8;
    auto image = new vector<vector<unsigned char>*>(height);
    for (auto y = 0; y < height; y++) {
        (*image)[y] = new vector<unsigned char>(width_bytes);
        auto row_ptr = (*image)[y];
        for (auto b = 0; b < width_bytes; b++) {
            (*row_ptr)[b] = 0xFF;
        }
    }
    cout << "DONE: ";
    time(&end_time);
    auto mem_load = (int)difftime(end_time, start_time);
    cout << to_string(mem_load) << " sec" << endl;

    // Clear memory
    time(&start_time);
    cout << "Clearing memory...";
    for (auto y = 0; y < height; y++) {
        delete (*image)[y];
    }
    delete image;
    cout << "DONE: ";
    time(&end_time);
    auto mem_clear = (int)difftime(end_time, start_time);
    cout << to_string(mem_clear) + " sec" << endl;
}

此函数使用数组进行加载清除：

void SimpleLoadAndClear_Array(int width, int height) {
    time_t start_time, end_time;

    // Load memory
    time(&start_time);
    cout << "Loading image into memory...";

    auto width_bytes = width / 8;
    auto image = new unsigned char*[height];
    for (auto y = 0; y < height; y++) {
        image[y] = new unsigned char[width_bytes];
        auto row_ptr = image[y];
        for (auto b = 0; b < width_bytes; b++) {
            row_ptr[b] = 0xFF;
        }
    }
    cout << "DONE: ";
    time(&end_time);
    auto mem_load = (int)difftime(end_time, start_time);
    cout << to_string(mem_load) << " sec" << endl;

    // Clear memory
    time(&start_time);
    cout << "Clearing memory...";

    for (auto y = 0; y < height; y++) {
        delete[] image[y];
    }
    delete[] image;
    cout << "DONE: ";
    time(&end_time);
    auto mem_clear = (int)difftime(end_time, start_time);
    cout << to_string(mem_clear) + " sec" << endl;
}

这是调用上述加载/清除函数的主函数：

void main()
{
    auto width = 455960;
    auto height = 453994;
    auto i_max = 50;
    for (auto i = 0; i < i_max; i++){
        SimpleLoadAndClear_Vector(width, height);
    }
}

经过 50 次迭代后，矢量版本的测试输出如下所示（显然加载/清除时间越来越长）：

Loading image into memory...DONE: 19 sec
Clearing memory...DONE: 24 sec
Loading image into memory...DONE: 40 sec
Clearing memory...DONE: 20 sec
Loading image into memory...DONE: 27 sec
Clearing memory...DONE: 39 sec
Loading image into memory...DONE: 35 sec
Clearing memory...DONE: 24 sec
Loading image into memory...DONE: 27 sec
Clearing memory...DONE: 34 sec
Loading image into memory...DONE: 33 sec
Clearing memory...DONE: 29 sec
Loading image into memory...DONE: 27 sec
Clearing memory...DONE: 35 sec
Loading image into memory...DONE: 32 sec
Clearing memory...DONE: 33 sec
Loading image into memory...DONE: 28 sec
Clearing memory...DONE: 37 sec
Loading image into memory...DONE: 31 sec
Clearing memory...DONE: 35 sec
Loading image into memory...DONE: 30 sec
Clearing memory...DONE: 38 sec
Loading image into memory...DONE: 31 sec
Clearing memory...DONE: 38 sec
Loading image into memory...DONE: 31 sec
Clearing memory...DONE: 41 sec
Loading image into memory...DONE: 32 sec
Clearing memory...DONE: 40 sec
Loading image into memory...DONE: 33 sec
Clearing memory...DONE: 42 sec
Loading image into memory...DONE: 35 sec
Clearing memory...DONE: 43 sec
Loading image into memory...DONE: 34 sec
Clearing memory...DONE: 46 sec
Loading image into memory...DONE: 36 sec
Clearing memory...DONE: 47 sec
Loading image into memory...DONE: 35 sec
Clearing memory...DONE: 49 sec
Loading image into memory...DONE: 37 sec
Clearing memory...DONE: 50 sec
Loading image into memory...DONE: 37 sec
Clearing memory...DONE: 51 sec
Loading image into memory...DONE: 39 sec
Clearing memory...DONE: 51 sec
Loading image into memory...DONE: 39 sec
Clearing memory...DONE: 53 sec
Loading image into memory...DONE: 40 sec
Clearing memory...DONE: 52 sec
Loading image into memory...DONE: 40 sec
Clearing memory...DONE: 55 sec
Loading image into memory...DONE: 41 sec
Clearing memory...DONE: 56 sec
Loading image into memory...DONE: 41 sec
Clearing memory...DONE: 59 sec
Loading image into memory...DONE: 42 sec
Clearing memory...DONE: 59 sec
Loading image into memory...DONE: 42 sec
Clearing memory...DONE: 60 sec
Loading image into memory...DONE: 44 sec
Clearing memory...DONE: 60 sec
Loading image into memory...DONE: 44 sec
Clearing memory...DONE: 63 sec
Loading image into memory...DONE: 44 sec
Clearing memory...DONE: 63 sec
Loading image into memory...DONE: 45 sec
Clearing memory...DONE: 64 sec
Loading image into memory...DONE: 46 sec
Clearing memory...DONE: 65 sec
Loading image into memory...DONE: 45 sec
Clearing memory...DONE: 67 sec
Loading image into memory...DONE: 47 sec
Clearing memory...DONE: 69 sec
Loading image into memory...DONE: 47 sec
Clearing memory...DONE: 70 sec
Loading image into memory...DONE: 48 sec
Clearing memory...DONE: 72 sec
Loading image into memory...DONE: 48 sec
Clearing memory...DONE: 74 sec
Loading image into memory...DONE: 49 sec
Clearing memory...DONE: 74 sec
Loading image into memory...DONE: 50 sec
Clearing memory...DONE: 74 sec
Loading image into memory...DONE: 50 sec
Clearing memory...DONE: 76 sec
Loading image into memory...DONE: 51 sec
Clearing memory...DONE: 78 sec
Loading image into memory...DONE: 53 sec
Clearing memory...DONE: 78 sec
Loading image into memory...DONE: 53 sec
Clearing memory...DONE: 80 sec
Loading image into memory...DONE: 54 sec
Clearing memory...DONE: 80 sec
Loading image into memory...DONE: 54 sec
Clearing memory...DONE: 82 sec
Loading image into memory...DONE: 55 sec
Clearing memory...DONE: 91 sec
Loading image into memory...DONE: 56 sec
Clearing memory...DONE: 84 sec
Loading image into memory...DONE: 56 sec
Clearing memory...DONE: 88 sec

50次迭代后，数组版本的测试输出如下（显然加载/清除时间是稳定的，不会越来越多）：

Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 26 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 17 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 26 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 26 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 17 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 26 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 26 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 17 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 26 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 26 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 25 sec
Loading image into memory...DONE: 27 sec
Clearing memory...DONE: 17 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 26 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 17 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 26 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 17 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 26 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 26 sec
Clearing memory...DONE: 17 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 25 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 25 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 25 sec
Clearing memory...DONE: 19 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 25 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 25 sec
Loading image into memory...DONE: 26 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 25 sec
Loading image into memory...DONE: 26 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 25 sec
Loading image into memory...DONE: 25 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 25 sec
Loading image into memory...DONE: 26 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 25 sec
Clearing memory...DONE: 17 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 26 sec
Loading image into memory...DONE: 25 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 25 sec
Loading image into memory...DONE: 26 sec
Clearing memory...DONE: 18 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 25 sec
Loading image into memory...DONE: 25 sec
Clearing memory...DONE: 19 sec
Loading image into memory...DONE: 18 sec
Clearing memory...DONE: 25 sec
Loading image into memory...DONE: 26 sec
Clearing memory...DONE: 18 sec

问题

这个 Windows 是否在处理巨大的 std::vectors？
是 std::vectors 只是执行蹩脚大量数据，是否有意设计？
我是不是完全错过了什么？
是否有任何其他明显的 std 容器我应该使用（我需要通过不同线程的 x 和 y 中的索引访问图像数据）？
还有其他好的解释和建议的解决方案吗？

【问题讨论】：

你可以只保留一次内存并重复使用它吗？
您是在测试发布还是调试版本？
由于缺少 RAM，我无法测试，但可能的解释是 std::vector 在后台使用了不同的内存分配器，甚至可能是内存池，当你扔咳咳 100GB 分配给它。
@Arvid 是 std::vectors 在设计上只是对大量数据执行糟糕的操作吗？ -- 如果你看看你编写代码的方式，你正在创建一个向量对象并调用分配器数千次。创建对象需要构造，仅使用指针和new 不会发生这种情况。所以不是vector 本身是“糟糕的”，而是你需要重新排列你的代码，这样你就不会创建太多的对象，也不会调用太多的分配器。
或者使用带有column/row major索引方案的单个（w x h）向量？

标签： c++ arrays windows memory vector

【解决方案1】：

我做错的是我为图像中的每一行调用向量分配器（数千次）。首先将整个事物分配为一个向量，然后将不同的行映射到大向量中的正确位置，问题就解决了。

感谢@PaulMcKenzie 为我指明正确方向的答案。

【讨论】：