Question

问题：

我正在使用MPI C ++绑定（MPICH 2）编写MPI程序。在下面列出的特定数量的调用函数之后，我在调用MPI::COMM::Create方法后立即出现死锁（所有正在运行的进程都终止）。

代码中的

LOCAL_COMM = LOCAL_COMM.Create(localGroup)需要更加具体。

我的猜测是出现了某种溢出，但我是MPI编程的新手，无法找到问题的根源。提前感谢您的帮助。

代码：

 Matrix Matrix::operator* (Matrix &M)
{
    //Synchronize all threads
    //COMM_WORLD.Barrier();

    int hisCol = M.myCol;
    int commSize = COMM_WORLD.Get_size();

    //Create local Group and Communicator equal to COMM_WORLD
    Intracomm LOCAL_COMM = COMM_WORLD.Dup();
    Group localGroup(LOCAL_COMM.Get_group());

    //Change locals if required
    if(commSize > myRow) {
        //Make array of ranks to include
        vector<int> masRanks(myRow);
        for(int i = 0; i < myRow; i++)
            masRanks[i] = i;

        //Create new localGroup and LOCAL_COMM
        localGroup = localGroup.Incl(myRow,&masRanks[0]);
        LOCAL_COMM = LOCAL_COMM.Create(localGroup); 
        if(LOCAL_COMM == COMM_NULL) {
            localGroup.Free();
            return Matrix(myRow,hisCol);
        }
        commSize = LOCAL_COMM.Get_size();
    }

    //Calculate number of Rows displacement and amount to Send per Thread
    vector<int> masRows(commSize,0);
    vector<int> amntToSend(commSize,0), amntToRecv(commSize,0);
    int tmpRows = myRow;

    for(int i = commSize; i > 0; i--) {
        masRows[i-1] = ((double)tmpRows / i + 0.5);
        tmpRows -= masRows[i-1];

        amntToSend[i-1] = masRows[i-1] * myCol;
        amntToRecv[i-1] = masRows[i-1] * hisCol;
    }

    //Form array specifying displacement to send to each thread
    vector<int> elemDisp(1,0), elemRecvDisp(1,0);
    if(myRank == 0)         
        for(int i = 1; i < commSize; i++) {
                elemDisp.push_back(amntToSend[i-1] + elemDisp[i-1]);
                elemRecvDisp.push_back(amntToRecv[i-1] + elemRecvDisp[i-1]);
        }

    //Prepare variables and send / receive left and right matrix
    int maxRow = *max_element(masRows.begin(),masRows.end());

    Matr tmpData(maxRow*myCol,0);
    LOCAL_COMM.Scatterv(&data[0],&amntToSend[0],&elemDisp[0],DOUBLE,&tmpData[0],masRows[myRank]*myCol,DOUBLE,0);
    LOCAL_COMM.Bcast(&M.data[0],M.myRow*hisCol,DOUBLE,0);

    //Start evaluation of each element
    Matr resData(masRows[myRank]*hisCol,0);
    for(int row = 0; row < masRows[myRank]; row++)
        for(int i = 0; i < hisCol; i++)
            for(int j = 0; j < myCol; j++)
                resData[row*hisCol + i] += tmpData[row*myCol + j] * M.data[j*hisCol + i];

    //Gather all data at 0-thread
    Matr outData(myRow*hisCol);
    LOCAL_COMM.Gatherv(&resData[0],masRows[myRank]*hisCol,DOUBLE,&outData[0],&amntToRecv[0],&elemRecvDisp[0],DOUBLE,0);

    //Return result Matrix-object
    Matrix resMatr(myRow, hisCol);
    resMatr.data = outData;

    //Free group and communicator
    localGroup.Free();
    LOCAL_COMM.Free();

    return resMatr;
}

Answer 1

明显但无益的答案是你编写了一个程序死锁。可以说这比写一个无死锁的要容易得多。

您是否调查了可能有助于您找到问题的任何工具？工具如：

Vampir

英特尔的Trace Analyzer（实际上可能基于吸血鬼）

或并行调试器

RogueWave的TotalView

Allinea的DDT

其中至少有一个应该在您的并行编程工具箱中。

Answer 2

我发现问题出现在我创建新Group和Communicator的方式中。也许COMM_WORLD.Dup（）从一开始就是错误的选择，因为在离开函数之后，我真的不知道编译器对它的值做了什么。

无论如何纠正的代码如下所列。

固定代码：

Matrix Matrix::operator* (Matrix &M)
{
    int hisCol = M.myCol;
    int commSize = COMM_WORLD.Get_size();

    //Make array of ranks to include
    vector<int> vRanks;
    for(int i = 0; i < myRow && i < commSize; i++)
        vRanks.push_back(i);

    //Create new LocalGroup and LOCAL_COMM
    Group LocalGroup = COMM_WORLD.Get_group().Incl(vRanks.size(),&vRanks[0]);
    Intracomm LOCAL_COMM = COMM_WORLD.Create(LocalGroup);

    COMM_WORLD.Barrier();

    //Make unused processes quit
    if(LOCAL_COMM == COMM_NULL) {
        LocalGroup.Free();
        return Matrix(myRow,hisCol);
    }
    commSize = LOCAL_COMM.Get_size();

    //The rest is pretty much the same

MPI :: COMM :: Create之后的死锁（所有进程终止）

2 个答案: