c++ - 如何将 C++ 写入速度加速到 CrystalDiskMark 测试的速度？

Question

现在我在内存中每秒获得大约 3.6GB 的数据，我需要将它们连续写入我的 SSD。我用 CrystalDiskMark 测试了我的 SSD 的写入速度，几乎是每秒 6GB，所以我认为这项工作应该没有那么难。

![我的SSD测试结果][1]:

[1] https://plus.google.com/u/0/photos/photo/106876803948041178149/6649598887699308850?authkey=CNbb5KjF8-jxJQ “测试结果”：

我的电脑是 Windows 10，使用 Visual Studio 2017 社区。

我找到了这个问题并尝试了投票最高的答案。不幸的是，他的option_2的写入速度只有1s/GB左右，远低于CrystalDiskMark测试的速度。然后我尝试了内存映射，这次写入速度变快了，大约 630ms/GB，但还是慢了很多。然后我尝试了多线程内存映射，好像线程数为4的时候，速度大概是350ms/GB，加上线程数，写入速度就没有再上去了。

内存映射代码：

#include <fstream>
#include <chrono>
#include <vector>
#include <cstdint>
#include <numeric>
#include <random>
#include <algorithm>
#include <iostream>
#include <cassert>
#include <thread>
#include <windows.h>
#include <sstream>


// Generate random data
std::vector<int> GenerateData(std::size_t bytes) {
    assert(bytes % sizeof(int) == 0);
    std::vector<int> data(bytes / sizeof(int));
    std::iota(data.begin(), data.end(), 0);
    std::shuffle(data.begin(), data.end(), std::mt19937{ std::random_device{}() });
    return data;
}

// Memory mapping
int map_write(int* data, int size, int id){
    char* name = (char*)malloc(100);
    sprintf_s(name, 100, "D:\\data_%d.bin",id);
    HANDLE hFile = CreateFile(name, GENERIC_READ | GENERIC_WRITE, 0, NULL, OPEN_ALWAYS, FILE_ATTRIBUTE_NORMAL, NULL);//
    if (hFile == INVALID_HANDLE_VALUE){
        return -1;
    }

    Sleep(0);

    DWORD dwFileSize = size;

    char* rname = (char*)malloc(100);
    sprintf_s(rname, 100, "data_%d.bin", id);

    HANDLE hFileMap = CreateFileMapping(hFile, NULL, PAGE_READWRITE, 0, dwFileSize, rname);//create file  
    if (hFileMap == NULL) {
        CloseHandle(hFile);
        return -2;
    }

    PVOID pvFile = MapViewOfFile(hFileMap, FILE_MAP_WRITE, 0, 0, 0);//Acquire the address of file on disk
    if (pvFile == NULL) {
        CloseHandle(hFileMap);
        CloseHandle(hFile);
        return -3;
}

    PSTR pchAnsi = (PSTR)pvFile;
    memcpy(pchAnsi, data, dwFileSize);//memery copy 

    UnmapViewOfFile(pvFile);

    CloseHandle(hFileMap);
    CloseHandle(hFile);

    return 0;
}

// Multi-thread memory mapping
void Mem2SSD_write(int* data, int size){
    int part = size / sizeof(int) / 4;

    int index[4];

    index[0] = 0;
    index[1] = part;
    index[2] = part * 2;
    index[3] = part * 3;

    std::thread ta(map_write, data + index[0], size / 4, 10);
    std::thread tb(map_write, data + index[1], size / 4, 11);
    std::thread tc(map_write, data + index[2], size / 4, 12);
    std::thread td(map_write, data + index[3], size / 4, 13);

    ta.join();
    tb.join();
    tc.join();
    td.join();
 }

//Test:
int main() {
    const std::size_t kB = 1024;
    const std::size_t MB = 1024 * kB;
    const std::size_t GB = 1024 * MB;

    for (int i = 0; i < 10; ++i) {
        std::vector<int> data = GenerateData(1 * GB);
        auto startTime = std::chrono::high_resolution_clock::now();
        Mem2SSD_write(&data[0], 1 * GB);
        auto endTime = std::chrono::high_resolution_clock::now();
        auto duration = std::chrono::duration_cast<std::chrono::milliseconds>(endTime - startTime).count();
        std::cout << "1G writing cost: " << duration << " ms" << std::endl;
    }

    system("pause");
    return 0;
}

所以我想问一下，C++写大文件有没有更快的写法？或者，为什么我不能像 CrystalDiskMark 测试的那样快速写入？CrystalDiskMark 是怎么写的？

任何帮助将不胜感激。谢谢！

标签： c++file-writing

Answer 1

首先，这不是c++问题，而是 os 相关问题。为了获得最大性能，需要使用操作系统特定的低级 api 调用，这在一般的c++库中不存在。从您的代码中可以清楚地看到您使用的是 windows api，因此搜索 windows 的解决方案有多少。

从CreateFileW功能：

当FILE_FLAG_NO_BUFFERING与 结合使用时FILE_FLAG_OVERLAPPED，这些标志可提供最大的异步性能，因为 I/O 不依赖于内存管理器的同步操作。

CreateFileW所以我们需要在通话或FILE_NO_INTERMEDIATE_BUFFERING通话中使用这两个标志的组合NtCreateFile

还扩展文件大小和有效数据长度需要一些时间，所以如果知道开始时的最终文件会更好 - 只需通过NtSetInformationFilewithFileEndOfFileInformation 或 via SetFileInformationByHandlewith设置文件最终大小FileEndOfFileInfo。然后使用SetFileValidData或通过NtSetInformationFile使用FileValidDataLengthInformation设置有效数据长度。设置有效数据长度需要SE_MANAGE_VOLUME_NAME在最初打开文件时启用特权（但不是在调用时SetFileValidData）

还要寻找文件压缩 - 如果文件压缩（如果在压缩文件夹中创建，默认情况下将被压缩）这是非常慢的写入。所以需要通过禁用文件压缩FSCTL_SET_COMPRESSION

那么当我们使用异步 I/O（最快的方式）时，我们不需要创建多个专用线程。相反，我们需要确定并发运行的 I/O 请求数。如果您使用CrystalDiskMark，它实际上会运行CdmResource\diskspd\diskspd64.exe进行测试，这是与其-o<count>参数对应的（运行diskspd64.exe /? > h.txt查看参数列表）。

使用非缓冲 I/O使任务更加困难，因为存在 3 个附加要求：

1. 传递给 WriteFile 的任何 ByteOffset 必须是扇区大小的倍数。
2. 传递给 WriteFile 的长度必须是扇区大小的整数
3. 缓冲区必须按照底层设备的对齐要求进行对齐。要获取此信息，请 NtQueryInformationFile使用FileAlignmentInformation 或GetFileInformationByHandleEx使用FileAlignmentInfo调用

在大多数情况下，页面对齐的内存也会是扇区对齐的，因为扇区大小大于页面大小的情况很少见。

所以几乎总是使用 VirtualAlloc 函数和多页大小（4,096 字节）分配的缓冲区是可以的。在针对较小代码大小的具体测试中，我使用了这个假设

struct WriteTest 
{
    enum { opCompression, opWrite };

    struct REQUEST : IO_STATUS_BLOCK 
    {
        WriteTest* pTest;
        ULONG opcode;
        ULONG offset;
    };

    LONGLONG _TotalSize, _BytesLeft;
    HANDLE _hFile;
    ULONG64 _StartTime;
    void* _pData;
    REQUEST* _pRequests;
    ULONG _BlockSize;
    ULONG _ConcurrentRequestCount;
    ULONG _dwThreadId;
    LONG _dwRefCount;

    WriteTest(ULONG BlockSize, ULONG ConcurrentRequestCount) 
    {
        if (BlockSize & (BlockSize - 1))
        {
            __debugbreak();
        }
        _BlockSize = BlockSize, _ConcurrentRequestCount = ConcurrentRequestCount;
        _dwRefCount = 1, _hFile = 0, _pRequests = 0, _pData = 0;
        _dwThreadId = GetCurrentThreadId();
    }

    ~WriteTest()
    {
        if (_pData)
        {
            VirtualFree(_pData, 0, MEM_RELEASE);
        }

        if (_pRequests)
        {
            delete [] _pRequests;
        }

        if (_hFile)
        {
            NtClose(_hFile);
        }

        PostThreadMessageW(_dwThreadId, WM_QUIT, 0, 0);
    }

    void Release()
    {
        if (!InterlockedDecrement(&_dwRefCount))
        {
            delete this;
        }
    }

    void AddRef()
    {
        InterlockedIncrementNoFence(&_dwRefCount);
    }

    void StartWrite()
    {
        IO_STATUS_BLOCK iosb;

        FILE_VALID_DATA_LENGTH_INFORMATION fvdl;
        fvdl.ValidDataLength.QuadPart = _TotalSize;
        NTSTATUS status;

        if (0 > (status = NtSetInformationFile(_hFile, &iosb, &_TotalSize, sizeof(_TotalSize), FileEndOfFileInformation)) ||
            0 > (status = NtSetInformationFile(_hFile, &iosb, &fvdl, sizeof(fvdl), FileValidDataLengthInformation)))
        {
            DbgPrint("FileValidDataLength=%x\n", status);
        }

        ULONG offset = 0;
        ULONG dwNumberOfBytesTransfered = _BlockSize;

        _BytesLeft = _TotalSize + dwNumberOfBytesTransfered;

        ULONG ConcurrentRequestCount = _ConcurrentRequestCount;

        REQUEST* irp = _pRequests;

        _StartTime = GetTickCount64();

        do 
        {
            irp->opcode = opWrite;
            irp->pTest = this;
            irp->offset = offset;
            offset += dwNumberOfBytesTransfered;
            DoWrite(irp++);
        } while (--ConcurrentRequestCount);
    }

    void FillBuffer(PULONGLONG pu, LONGLONG ByteOffset)
    {
        ULONG n = _BlockSize / sizeof(ULONGLONG);
        do 
        {
            *pu++ = ByteOffset, ByteOffset += sizeof(ULONGLONG);
        } while (--n);
    }

    void DoWrite(REQUEST* irp)
    {
        LONG BlockSize = _BlockSize;

        LONGLONG BytesLeft = InterlockedExchangeAddNoFence64(&_BytesLeft, -BlockSize) - BlockSize;

        if (0 < BytesLeft)
        {
            LARGE_INTEGER ByteOffset;
            ByteOffset.QuadPart = _TotalSize - BytesLeft;

            PVOID Buffer = RtlOffsetToPointer(_pData, irp->offset);

            FillBuffer((PULONGLONG)Buffer, ByteOffset.QuadPart);

            AddRef();

            NTSTATUS status = NtWriteFile(_hFile, 0, 0, irp, irp, Buffer, BlockSize, &ByteOffset, 0);

            if (0 > status)
            {
                OnComplete(status, 0, irp);
            }
        }
        else if (!BytesLeft)
        {
            // write end
            ULONG64 time = GetTickCount64() - _StartTime;

            WCHAR sz[64];
            StrFormatByteSizeW((_TotalSize * 1000) / time, sz, RTL_NUMBER_OF(sz));
            DbgPrint("end:%S\n", sz);
        }
    }

    static VOID NTAPI _OnComplete(
        _In_    NTSTATUS status,
        _In_    ULONG_PTR dwNumberOfBytesTransfered,
        _Inout_ PVOID Ctx
        )
    {
        reinterpret_cast<REQUEST*>(Ctx)->pTest->OnComplete(status, dwNumberOfBytesTransfered, reinterpret_cast<REQUEST*>(Ctx));
    }

    VOID OnComplete(NTSTATUS status, ULONG_PTR dwNumberOfBytesTransfered, REQUEST* irp)
    {
        if (0 > status)
        {
            DbgPrint("OnComplete[%x]: %x\n", irp->opcode, status);
        }
        else 
        switch (irp->opcode)
        {
        default:
            __debugbreak();

        case opCompression:
            StartWrite();
            break;

        case opWrite:
            if (dwNumberOfBytesTransfered == _BlockSize)
            {
                DoWrite(irp);
            }
            else
            {
                DbgPrint(":%I64x != %x\n", dwNumberOfBytesTransfered, _BlockSize);
            }
        }

        Release();
    }

    NTSTATUS Create(POBJECT_ATTRIBUTES poa, ULONGLONG size)
    {
        if (!(_pRequests = new REQUEST[_ConcurrentRequestCount]) ||
            !(_pData = VirtualAlloc(0, _BlockSize * _ConcurrentRequestCount, MEM_COMMIT, PAGE_READWRITE)))
        {
            return STATUS_INSUFFICIENT_RESOURCES;
        }

        ULONGLONG sws = _BlockSize - 1;
        LARGE_INTEGER as;

        _TotalSize = as.QuadPart = (size + sws) & ~sws;

        HANDLE hFile;
        IO_STATUS_BLOCK iosb;

        NTSTATUS status = NtCreateFile(&hFile,
            DELETE|FILE_GENERIC_READ|FILE_GENERIC_WRITE&~FILE_APPEND_DATA,
            poa, &iosb, &as, 0, 0, FILE_OVERWRITE_IF, 
            FILE_NON_DIRECTORY_FILE|FILE_NO_INTERMEDIATE_BUFFERING, 0, 0);

        if (0 > status)
        {
            return status;
        }

        _hFile = hFile;

        if (0 > (status = RtlSetIoCompletionCallback(hFile, _OnComplete, 0)))
        {
            return status;
        }

        static USHORT cmp = COMPRESSION_FORMAT_NONE;

        REQUEST* irp = _pRequests;

        irp->pTest = this;
        irp->opcode = opCompression;

        AddRef();
        status = NtFsControlFile(hFile, 0, 0, irp, irp, FSCTL_SET_COMPRESSION, &cmp, sizeof(cmp), 0, 0);

        if (0 > status)
        {
            OnComplete(status, 0, irp);
        }

        return status;
    }
};

void WriteSpeed(POBJECT_ATTRIBUTES poa, ULONGLONG size, ULONG BlockSize, ULONG ConcurrentRequestCount)
{
    BOOLEAN b;
    NTSTATUS status = RtlAdjustPrivilege(SE_MANAGE_VOLUME_PRIVILEGE, TRUE, FALSE, &b);

    if (0 <= status)
    {
        status = STATUS_INSUFFICIENT_RESOURCES;

        if (WriteTest * pTest = new WriteTest(BlockSize, ConcurrentRequestCount))
        {
            status = pTest->Create(poa, size);

            pTest->Release();

            if (0 <= status)
            {
                MessageBoxW(0, 0, L"Test...", MB_OK|MB_ICONINFORMATION);
            }
        }
    }
}