MPI_Scatterv: segmentation fault 11 on process 0 only

Question

I'm trying to scatter values among processes belonging to an hypercube group (quicksort project). Depending on the amount of processes I either create a new communicator excluding excessive processes, or I duplicate MPI_COMM_WORLD if it fits exactly any hypercube (power of 2).

In both cases, processes other than 0 receive their data, but: - On first scenario, process 0 throws a segmentation fault 11 - On second scenario, nothing faults, but process 0 received values are gibberish.

NOTE: If I try a regular MPI_Scatter everything works well.

//Input
vector<int> LoadFromFile();

int d;                      //dimension of hypercube
int p;                      //active processes
int idle;                   //idle processes 
vector<int> values;         //values loaded
int arraySize;              //number of total values to distribute

int main(int argc, char* argv[])
{       
int mpiWorldRank;
int mpiWorldSize;

int mpiRank; 
int mpiSize;

MPI_Init(&argc, &argv);

MPI_Comm_rank(MPI_COMM_WORLD, &mpiWorldRank);
MPI_Comm_size(MPI_COMM_WORLD, &mpiWorldSize);
MPI_Comm MPI_COMM_HYPERCUBE;

d = log2(mpiWorldSize);     
p = pow(2, d);                  //Number of processes belonging to the hypercube
idle = mpiWorldSize - p;        //number of processes in excess
int toExclude[idle];            //array of idle processes to exclude from communicator
int sendCounts[p];              //array of values sizes to be sent to processes

//
int i = 0;
while (i < idle)
{
    toExclude[i] = mpiWorldSize - 1 - i;
    ++i;
}

//CREATING HYPERCUBE GROUP: Group of size of power of 2 -----------------
MPI_Group world_group;
MPI_Comm_group(MPI_COMM_WORLD, &world_group);

// Remove excessive processors if any from communicator
if (idle > 0)
{
    MPI_Group newGroup;     
    MPI_Group_excl(world_group, 1, toExclude, &newGroup);
    MPI_Comm_create(MPI_COMM_WORLD, newGroup, &MPI_COMM_HYPERCUBE);
    //Abort any processor not part of the hypercube.    
    if (mpiWorldRank > p)
    {
        cout << "aborting: " << mpiWorldRank <<endl;
        MPI_Finalize();
        return 0;
    }   
}   
else 
{
    MPI_Comm_dup(MPI_COMM_WORLD, &MPI_COMM_HYPERCUBE);
}

MPI_Comm_rank(MPI_COMM_HYPERCUBE, &mpiRank);
MPI_Comm_size(MPI_COMM_HYPERCUBE, &mpiSize);
//END OF: CREATING HYPERCUBE GROUP --------------------------

if (mpiRank == 0)
{
    //STEP1: Read input
    values = LoadFromFile();
    arraySize = values.size();
}

//Transforming input vector into an array
int valuesArray[values.size()];
if(mpiRank == 0)
{
    copy(values.begin(), values.end(), valuesArray);
}

//Broadcast input size to all processes
MPI_Bcast(&arraySize, 1, MPI_INT, 0, MPI_COMM_HYPERCUBE);

//MPI_Scatterv: determining size of arrays to be received and displacement
int nmin = arraySize / p;
int remainingData = arraySize % p;
int displs[p];
int recvCount;

int k = 0;
for (i=0; i<p; i++)
{
    sendCounts[i] = i < remainingData
        ? nmin+1
        : nmin;
    displs[i] = k;
    k += sendCounts[i];
}

recvCount = sendCounts[mpiRank];
int recvValues[recvCount];

//Following MPI_Scatter works well:     
// MPI_Scatter(&valuesArray, 13, MPI_INT, recvValues , 13, MPI_INT, 0, MPI_COMM_HYPERCUBE);

MPI_Scatterv(&valuesArray, sendCounts, displs, MPI_INT, recvValues , recvCount, MPI_INT, 0, MPI_COMM_HYPERCUBE);

int j = 0;
while (j < recvCount)
{
    cout << "rank " << mpiRank << " received: " << recvValues[j] << endl;
    ++j;
}   

MPI_Finalize();
return 0;
}

I would add error checking to all operations that can fail and a bunch of debug outputs (printf or similar). There are a bunch of arrays that could be overflowed if the values aren't what you think they should be or if an operation fails. Having multi-processes makes this much harder or impossible to debug by manually reading the code. — uesp, Oct 30 '14 at 13:17

Hristo Iliev · Accepted Answer · 2014-10-30T13:57:44.343

First of all, you are supplying wrong arguments to MPI_Group_excl:

MPI_Group_excl(world_group, 1, toExclude, &newGroup);
//                          ^

The second argument specifies the number of entries in the exclusion list and should therefore be equal to idle. Since you are excluding a single rank only, the resulting group has mpiWorldSize-1 ranks and hence MPI_Scatterv expects that both sendCounts[] and displs[] have that many elements. Of those only p elements are properly initialised and and the rest are random, therefore MPI_Scatterv crashes in the root.

Another error is the code that aborts the idle processes: it should read if (mpiWorldRank >= p).

I would recommend that the entire exclusion code is replaced by a single call to MPI_Comm_split instead:

MPI_Comm comm_hypercube;
int colour = mpiWorldRank >= p ? MPI_UNDEFINED : 0;

MPI_Comm_split(MPI_COMM_WORLD, colour, mpiWorldRank, &comm_hypercube);
if (comm_hypercube == MPI_COMM_NULL)
{
   MPI_Finalize();
   return 0;
}

When no process supplies MPI_UNDEFINED as its colour, the call is equivalent to MPI_Comm_dup.

Note that you should avoid using in your code names starting with MPI_ as those could clash with symbols from the MPI implementation.

Additional note: std::vector<T> uses contiguous storage, therefore you could do without copying the elements into a regular array and simply provide the address of the first element in the call to MPI_Scatter(v):

MPI_Scatterv(&values[0], ...);

MPI_Scatterv: segmentation fault 11 on process 0 only

1 Answers1