COMP5426 Distributed
mming Distributed Memory
ective Com
Copyright By PowCoder代写 加微信 powcoder
MPI specifies a set of collective comm.
One-to-All: One process contributes to the processes receive the result
All-to-One: All processes
receives the result
munication
MPI_Scatter, MPI_Scatter
ather, MPI_Gatherv, MPI
l-to-All: All processes contribute to th d all processes receive the result
PI_Allgather, MPI_Allgathe
contribute to the
_Scan, …
Functions result and all
llreduce, …
datatype, i
The MPI broadcast routine (one-to-all) is:
int MPI_Bcast(void *buf, int count, MPI_D
source, MPI_Comm c
must call MPI_Bcast()
The MPI reduction operation (all-to-one) is:
int MPI_Reduce(void *sendbuf, void *recvbuf, int
MPI_Datatype datatype, MPI_
, MPI_Comm comm)
The MPI All reduction operation (all-to-all) is:
int MPI_Allreduce(void *sendbuf, void *recvbuf,
int count,
MPI_Comm comm)
MPI_Datatype
datatype, MPI_Op
MPI_LAND MPI_BAND
MPI_BXOR MPI_MAXLOC
MPI_MINLOC
ined Reductio
Logical AND Bit-wise AND
Logical OR
Logical XOR
Bit-wise XOR
max value-location
min value-location
Operations
C integers
C integers
C integers
C integers
C integers C integers
C integers
C integers
C integers
C integers Data-pairs
Data-pairs
and floating point
and floating point
and floating point
and floating point
ined Reductio
Operations
The operation MPI
_MAXLOC combin
(v, l) and
returns the pair (v, l) such that v is the
maximum among all v ‘s
and l is the corresponding l (if there are more than one, it is
smallest among all these l ‘s).
MPI_MINLOC does the same, except for
minimum valu
ined Reductio
Operations
data-pairs used w
MPI_MAXLOC and
MPI_MINLOC reduction operations
C Datatype
pair of int
PI_LONG_INT
long and int
PI_LONG_DO
long double and
PI_FLOAT_INT
uble and int
/* each process has a d
double ain;
ined Reductio
MPI_Allreduce( in, root, MPI_CO
ouble: ain */
Operations
out, 1, MPI_DOUBLE_INT, MPI_MAX MM_WORLD );
The MPI gather
ather(void
MPI_Datatype senddatatype,
int target, MP
o-one) is:
MPI_Datatype
t sendcount,
cvdatatype,
The MPI All gather operat
PI_Allgather(void
int recvcount, MPI_Dataty
ion (all-to-all) is:
*sendbuf, int s
MPI_Datatype senddatatype,
MPI_Allgather
void *recvbuf,
cvdatatype,
The MPI gatherv o
peration (all-
therv(void
MPI_Datatype sendtype
*recvcounts,
MPI_Datatype recvty
t sendcount,
, void *recvbuf,
pe, int roo
The MPI gatherv o
int sendarray[100];
peration (all-
therv(void
MPI_Datatype sendtype
*recvcounts,
MPI_Datatype recvty
int stride = 150;
int *displs, *rcounts;
…rbuf = (int *) malloc(numprocs* for (int i=0; i