The global reduce functions (MPI_Reduce, MPI_Op_create, MPI_Op_free, MPI_Allreduce, MPI_Reduce_scatter, MPI_Scan) perform a global reduce operation (such as sum, max, logical AND, etc.) across all the members of a group. The reduction operation can be either one of a predefined list of operations, or a user-defined operation. The global reduction functions come in several flavors: a reduce that returns the result of the reduction at one node, an all-reduce that returns this result at all nodes, and a scan (parallel prefix) operation. In addition, a reduce-scatter operation combines the functionality of a reduce and a scatter operation.
.sp
MPI_Reduce combines the elements provided in the input buffer of each process in the group, using the operation op, and returns the combined value in the output buffer of the process with rank root. The input buffer is defined by the arguments sendbuf, count, and datatype; the output buffer is defined by the arguments recvbuf, count, and datatype; both have the same number of elements, with the same type. The routine is called by all group members using the same arguments for count, datatype, op, root, and comm. Thus, all processes provide input buffers and output buffers of the same length, with elements of the same type. Each process can provide one element, or a sequence of elements, in which case the combine operation is executed element-wise on each entry of the sequence. For example, if the operation is MPI_MAX and the send buffer contains two elements that are floating-point numbers (count = 2 and datatype = MPI_FLOAT), then recvbuf(1) = global max (sendbuf(1)) and recvbuf(2) = global max(sendbuf(2)).
.sp
.SH USE OF IN-PLACE OPTION
When the communicator is an intracommunicator, you can perform a reduce operation in-place (the output buffer is used as the input buffer). Use the variable MPI_IN_PLACE as the value of the root process \fIsendbuf\fR. In this case, the input data is taken at the root from the receive buffer, where it will be replaced by the output data.
.sp
Note that MPI_IN_PLACE is a special kind of value; it has the same restrictions on its use as MPI_BOTTOM.
.sp
Because the in-place option converts the receive buffer into a send-and-receive buffer, a Fortran binding that includes INTENT must mark these as INOUT, not OUT.
.sp
.SH WHEN COMMUNICATOR IS AN INTER-COMMUNICATOR
.sp
When the communicator is an inter-communicator, the root process in the first group combines data from all the processes in the second group and then performs the \fIop\fR operation. The first group defines the root process. That process uses MPI_ROOT as the value of its \fIroot\fR argument. The remaining processes use MPI_PROC_NULL as the value of their \fIroot\fR argument. All processes in the second group use the rank of that root process in the first group as the value of their \fIroot\fR argument. Only the send buffer arguments are significant in the second group, and only the receive buffer arguments are significant in the root process of the first group.
.sp
.SH PREDEFINED REDUCE OPERATIONS
.sp
The set of predefined operations provided by MPI is listed below (Predefined Reduce Operations). That section also enumerates the datatypes each operation can be applied to. In addition, users may define their own operations that can be overloaded to operate on several datatypes, either basic or derived. This is further explained in the description of the user-defined operations (see the man pages for MPI_Op_create and MPI_Op_free).
.sp
The operation op is always assumed to be associative. All predefined operations are also assumed to be commutative. Users may define operations that are assumed to be associative, but not commutative. The ``canonical'' evaluation order of a reduction is determined by the ranks of the processes in the group. However, the implementation can take advantage of associativity, or associativity and commutativity, in order to change the order of evaluation. This may change the result of the reduction for operations that are not strictly associative and commutative, such as floating point addition.
.sp
Predefined operators work only with the MPI types listed below (Predefined Reduce Operations, and the section MINLOC and MAXLOC, below). User-defined operators may operate on general, derived datatypes. In this case, each argument that the reduce operation is applied to is one element described by such a datatype, which may contain several basic values. This is further explained in Section 4.9.4 of the MPI Standard, "User-Defined Operations."
The following predefined operations are supplied for MPI_Reduce and related functions MPI_Allreduce, MPI_Reduce_scatter, and MPI_Scan. These operations are invoked by placing the following in op:
.sp
.nf
Name Meaning
--------- --------------------
MPI_MAX maximum
MPI_MIN minimum
MPI_SUM sum
MPI_PROD product
MPI_LAND logical and
MPI_BAND bit-wise and
MPI_LOR logical or
MPI_BOR bit-wise or
MPI_LXOR logical xor
MPI_BXOR bit-wise xor
MPI_MAXLOC max value and location
MPI_MINLOC min value and location
.fi
.sp
The two operations MPI_MINLOC and MPI_MAXLOC are discussed separately below (MINLOC and MAXLOC). For the other predefined operations, we enumerate below the allowed combinations of op and datatype arguments. First, define groups of MPI basic datatypes in the following way:
.sp
.nf
C integer: MPI_INT, MPI_LONG, MPI_SHORT,
MPI_UNSIGNED_SHORT, MPI_UNSIGNED,
MPI_UNSIGNED_LONG
Fortran integer: MPI_INTEGER
Floating-point: MPI_FLOAT, MPI_DOUBLE, MPI_REAL,
MPI_DOUBLE_PRECISION, MPI_LONG_DOUBLE
Logical: MPI_LOGICAL
Complex: MPI_COMPLEX
Byte: MPI_BYTE
.fi
.sp
Now, the valid datatypes for each option is specified below.
.sp
.nf
Op Allowed Types
---------------- ---------------------------
MPI_MAX, MPI_MIN C integer, Fortran integer,
floating-point
MPI_SUM, MPI_PROD C integer, Fortran integer,
floating-point, complex
MPI_LAND, MPI_LOR, C integer, logical
MPI_LXOR
MPI_BAND, MPI_BOR, C integer, Fortran integer, byte
MPI_BXOR
.fi
.sp
\fBExample 1:\fR A routine that computes the dot product of two vectors that are distributed across a group of processes and returns the answer at process zero.
\fBExample 2:\fR A routine that computes the product of a vector and an array that are distributed across a group of processes and returns the answer at process zero.
.sp
.nf
SUBROUTINE PAR_BLAS2(m, n, a, b, c, comm)
REAL a(m), b(m,n) ! local slice of array
REAL c(n) ! result
REAL sum(n)
INTEGER n, comm, i, j, ierr
! local sum
DO j= 1, n
sum(j) = 0.0
DO i = 1, m
sum(j) = sum(j) + a(i)*b(i,j)
END DO
END DO
! global sum
CALL MPI_REDUCE(sum, c, n, MPI_REAL, MPI_SUM, 0, comm, ierr)
! return result at process zero (and garbage at the other nodes)
The operator MPI_MINLOC is used to compute a global minimum and also an index attached to the minimum value. MPI_MAXLOC similarly computes a global maximum and index. One application of these is to compute a global minimum (maximum) and the rank of the process containing this value.
MPI_REDUCE( in, out, 30, MPI_2DOUBLE_PRECISION, MPI_MAXLOC, root,
comm, ierr );
! At this point, the answer resides on process root
IF (myrank .EQ. root) THEN
! read ranks out
DO I= 1, 30
aout(i) = out(1,i)
ind(i) = out(2,i) ! rank is coerced back to an integer
END DO
END IF
.fi
.sp
\fBExample 5:\fR Each process has a nonempty array of values. Find the minimum global value, the rank of the process that holds it, and its index on this process.
.sp
.nf
#define LEN 1000
float val[LEN]; /* local array of values */
int count; /* local number of values */
int myrank, minrank, minindex;
float minval;
struct {
float value;
int index;
} in, out;
/* local minloc */
in.value = val[0];
in.index = 0;
for (i=1; i < count; i++)
if (in.value > val[i]) {
in.value = val[i];
in.index = i;
}
/* global minloc */
MPI_Comm_rank(MPI_COMM_WORLD, &myrank);
in.index = myrank*LEN + in.index;
MPI_Reduce( in, out, 1, MPI_FLOAT_INT, MPI_MINLOC, root, comm );
/* At this point, the answer resides on process root
*/
if (myrank == root) {
/* read answer out
*/
minval = out.value;
minrank = out.index / LEN;
minindex = out.index % LEN;
.fi
.sp
All MPI objects (e.g., MPI_Datatype, MPI_Comm) are of type INTEGER in Fortran.
.SH NOTES ON COLLECTIVE OPERATIONS
The reduction functions (
.I MPI_Op
) do not return an error value. As a result,
if the functions detect an error, all they can do is either call
.I MPI_Abort
or silently skip the problem. Thus, if you change the error handler from
.I MPI_ERRORS_ARE_FATAL
to something else, for example,
.I MPI_ERRORS_RETURN
,
then no error may be indicated.
The reason for this is the performance problems in ensuring that
all collective routines return the same error value.
Almost all MPI routines return an error value; C routines as the value of the function and Fortran routines in the last argument. C++ functions do not return errors. If the default error handler is set to MPI::ERRORS_THROW_EXCEPTIONS, then on error the C++ exception mechanism will be used to throw an MPI::Exception object.
Before the error value is returned, the current MPI error handler is
called. By default, this error handler aborts the MPI job, except for I/O function errors. The error handler may be changed with MPI_Comm_set_errhandler; the predefined error handler MPI_ERRORS_RETURN may be used to cause error values to be returned. Note that MPI does not guarantee that an MPI program can continue past an error.
