MPI_Recv未收到所有MPI_Send请求

Question

I have a bug in my code. I have multiple processes all processing data from a binary tree. At the end, they should send the results to the master node (node 0) where the results will be processed. However, for some reason, some of the MPI_Sends are not being received.

int *output=(int*) malloc(sizeof(int)*(varNum+2)); //contains all variable values and maxSAT and assignNum

if(proc_id!=0 && proc_id<nodeNums){
    output[0]=maxSAT;
    output[1]=assignNum;
    for(i=2;i<varNum+2;i++){
        output[i]=varValues[i-2];
    }
    MPI_Send(output,varNum+2,MPI_INT,0,TAG,MPI_COMM_WORLD);
    printf("proc %d sent data\n",proc_id);
}
else if(proc_id==0){
    for(i=1;i<nodeNums;i++){
        printf("receiving data from %d\n",i);
        MPI_Recv(output,varNum+2,MPI_INT,i,TAG,MPI_COMM_WORLD,MPI_STATUS_IGNORE);
        if(output[0]>maxSAT){
            maxSAT=output[0];
            assignNum=output[1];
            for(i=0;i<varNum;i++){
                varValues[i]=output[i+2];
            }   
        }
        else if(output[0]==maxSAT){
            assignNum+=output[1];
        }
    }
}

When I run it with 8 processes (nodeNums=8), this is the output.

proc 2 sent data
receiving data from 1
proc 5 sent data
proc 6 sent data
proc 3 sent data
proc 7 sent data
proc 1 sent data
proc 4 sent data

For some reason, all processes are sending data, but it is only receiving from 1. However, if I run it with 4 processes, everything is sent/received. Anyone has any idea why this happens?

Answer 1

The problem has nothing to do with MPI. Your mistake is the use of the same variable in two different but nested loops:

else if(proc_id==0){
    for(i=1;i<nodeNums;i++){ <----------------- (1)
        ...
            for(i=0;i<varNum;i++){ <----------- (2)
                varValues[i]=output[i+2];
            }
        ...
    }
}

After the inner loop completes, the value of i is equal to varNum and if it happens that varNum is greater or equal to nodeNums , the outer loop terminates too. Change the name of the loop variable of the inner loop.

Answer 2

This is not really the way to use MPI. What you want here is MPI_Gather() , which is where all the processes (including the root) send a chunk of data and the gathering process receives them all. Like this:

rbuf = (int *)malloc(nodeNums*(varNum+2)*sizeof(int));
MPI_Gather(output, varNum+2, MPI_INT, rbuf, varNum+2, MPI_INT, 0, MPI_COMM_WORLD);

All your processes should execute the above in the same part of their execution. All the data will end up in rbuf .

In your case, if the root doesn't want to send anything, just have it send empty data which it can simply ignore (after all, it doesn't need to physically "send" to itself, so this is not very inefficient).