Global Reduction Operations

void Intracomm.reduce(Object sendbuf, int sendoffset,
                      Object recvbuf, int recvoffset,
                      int count, Datatype datatype,
                      Op op, int root)

sendbuf	send buffer array
sendoffset	initial offset in send buffer
recvbuf	receive buffer array
recvoffset	initial offset in receive buffer
count	number of items in send buffer
datatype	data type of each item in send buffer
op	reduce operation
dest	rank of root process

Combine elements in input buffer of each process using the reduce operation, and return the combined value in the output buffer of the root process. Java binding of the MPI operation MPI_REDUCE.

The predefined operations are available in Java as MPI.MAX, MPI.MIN, MPI.SUM, MPI.PROD, MPI.LAND, MPI.BAND, MPI.LOR, MPI.BOR, MPI.LXOR, MPI.BXOR, MPI.MINLOC and MPI.MAXLOC.

The handling of MINLOC and MAXLOC is modelled on the Fortran binding. The extra predefined types MPI.SHORT2, MPI.INT2, MPI.LONG2, MPI.FLOAT2, MPI.DOUBLE2 describe pairs of Java numeric primitive types.

Op.Op(UserFunction function, boolean commute)

function	user defined function
commute	true if commutative, false otherwise

Bind a user-defined global reduction operation to an Op object. Java binding of the MPI operation MPI_OP_CREATE. The abstract base class UserFunction is defined by

  class UserFunction {
    public abstract void call(Object invec, Object inoutvec,
                              Datatype datatype) ;
  }

To define a new operation, the programmer should define a concrete subclass of UserFunction, implementing the call method, then pass an object from this class to the Op constructor. The UserFunction.call method plays exactly the same role as the function argument in the standard bindings of MPI. The actual arguments invec and outvec passed to call will be equal-length arrays with elements of the type specified in the datatype argument. The user-defined call method combines these in the way specified in the MPI standard.

• Advice to implementors. There is a problem for JNI-based wrapper implementations here: how does the C code inside the wrapper persuade the MPI reduction operation to invoke the Java method, call? Some C MPI_User_function must dispatch the call method of a Java object of type UserFunction. The difficulty is in creating a C function at runtime that will do this. A generic C function can reference a Java object through a reference held in some static variable. But if more than one user defined operation is needed, how does a particular invocation find the proper Java function object? One possibility is to assume a constant bound, $N$, on the number of permitted user-defined operations. The interface predefines $N$ different MPI_User_function functions, invoking call methods through $N$ different static variables. These MPI_User_function/static-variable pairs are managed as a pool. Calling the Op constructor involves copying a reference to its function object argument into one of the static variables in the pool and passing the corresponding MPI_User_function to the physical MPI_Op_create.(End of advice to implementors.)

void Op.finalize()

Destructor. Java binding of the MPI operation MPI_OP_FREE.

void Intracomm.allreduce(Object sendbuf, int sendoffset,
                         Object recvbuf, int recvoffset,
                         int count, Datatype datatype,
                         Op op)

sendbuf	send buffer array
sendoffset	initial offset in send buffer
recvbuf	receive buffer array
recvoffset	initial offset in receive buffer
count	number of items in send buffer
datatype	data type of each item in send buffer
op	reduce operation

Same as reduce except that the result appears in receive buffer of all process in the group. Java binding of the MPI operation MPI_ALLREDUCE.