死鎖：創建（所有進程終止）

Question

問題：

我正在寫使用MPI C++結合（MPICH 2）MPI PROGRAMM。在調用MPI::COMM::Create方法（所有正在運行的進程終止）後，在下面列出的特定數量的調用函數之後，我立即發生死鎖。

LOCAL_COMM = LOCAL_COMM.Create(localGroup)在代碼中，需要更具體。

我的猜測是出現了某種溢出，但我是MPI編程的新手，無法找到問題的根源。預先感謝您的幫助。

代碼：

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

我發現問題在於我創建了新的Group和Communicator。也許COMM_WORLD.Dup（）從一開始就是錯誤的選擇，因爲在離開函數後，我真的不知道編譯器如何處理它的值。

無論如何糾正的代碼如下所列。

固定碼：

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 

Answer 2

明顯，但無益，答案是，你已經寫了一個程序死鎖。可以說這比編寫一個無死鎖的容易得多。

你有沒有調查過任何可能幫助你找到問題的工具？工具，如：

Vampir

英特爾Trace Analyzer（其實際上可基於VAMPIR）

或並行調試器

RogueWave的TotalView

Allinea的DDT

至少其中之一應該在你的段落中llel編程工具箱。