Using Vcluster for Dynamic Sparse Data Exchange

Dynamic Sparse Data Exchange

Dynamic Sparse Data Exchange or DSDE, is a typical point to point communication in which senders know to which processor to receive, but receivers has not knowledge about from where they are receiving. OpenFPM use the NBX method or Non blocking consensus exchange. (Said without bombastic word each processor wait for incoming messages. Pretty basic achievement and technique in standard server programming, presented as bombastic and incredible discovery in MPI)

Initialization

Before using any functionality the library must be initialized. After initialization we can create the Vcluster object

 
    openfpm_init(&argc,&argv);
    

Dynamic Sparse Data Exchange

To do dynamic sparse data exchange, each processor fill a send processor list and create a message for each processor. In this case the message will be a complex object. OpenFPM use the capability to serialize complex object into sequence of byte to send over the network and de-serializa or re-assemble the object into another processors. In this case the complex object is a list of double numbers. At the end of the example each processor print what it received

 
    // id of the processor calling this function
    long int proc_id = v_cl.getProcessUnitID();
 
    // number of processors executing this program
    long int n_proc = v_cl.getProcessingUnits();
 
    // List of processors we communicate with
    openfpm::vector<size_t> prc_send;
 
    // For each processor we want to send a vector of doubles
    // in this case each processor send N_NEXT vectors.
    // In general we can think to openfpm::vector<T> as a set of objects T.
    // where we want to send one object T to each processor in out sending list.
    // In this case T is a list of double or (openfpm::vector<double>)
    openfpm::vector<openfpm::vector<double>> messages_send(N_NEXT);
 
    // Here we prepare the senbding buffer
    for (size_t i = 0, m = 0 ; i < N_NEXT ; i++, m++)
    {
        // create the sending processor list
        prc_send.add(openfpm::math::positive_modulo(proc_id + i + 1,n_proc));
 
        // Fill with soma data the vectors
        for (size_t j = 0 ; j < N_NUMBERS ; j++)
            messages_send.get(m).add(j+N_NUMBERS*proc_id);
 
    }
 
    // Buffer that receive messages
    openfpm::vector<double> messages_recv2;
 
    // List of processor from which we receove
    openfpm::vector<size_t> prc_recv2;
 
    // number of elements we receive from each processors
    openfpm::vector<size_t> sz_recv2;
 
    v_cl.SSendRecv(messages_send,messages_recv2,prc_send,prc_recv2,sz_recv2);
 
    // here each processor print the received message
 
    std::cout << "Processor " << proc_id << " received ";
    for (size_t i = 0 ; i < messages_recv2.size() ; i++)
        std::cout << messages_recv2.get(i) << "   ";
 
    std::cout << std::endl;
 

Finalize

At the very end of the program we have always to de-initialize the library

 
    openfpm_finalize();