Collective Comm

8/12/2019 Collective Comm

1/45

Collective Communication


2/45

Collective Communication

Collective communication is defined ascommunication that involves a group ofprocesses

More restrictive than point to point Data sent is same as the data received, i.e.

type, amount All processes involved make one call, no tag to

match operation Processes involved can return only when

operation completes blocking communication only

Standard Mode only


3/45

Collective Functions

Barrier synchronization across all group members Broadcast from one member to all members of a group Gather data from all group members to one member Scatter data from one member to all members of a group A variation on Gather where all members of the group receive

the result. (allgather) Scatter/Gather data from all members to all members of a

group (also called complete exchange or all-to-all) (alltoall) Global reduction operations such as sum, max, min, or user-

defined functions, where the result is returned to all groupmembers and a variation where the result is returned to onlyone member

A combined reduction and scatter operation Scan across all members of a group (also called prefix)


4/45



5/45



6/45

Collective Functions MPI_BARRIER

blocks the caller until all group membershave called it

returns at any process only after all groupmembers have entered the call

C int MPI_Barrier(MPI_Comm comm ) Input Parameter

comm: communicator (handle)

Fortran MPI_BARRIER(COMM, IERROR) INTEGER COMM, IERROR


7/45

Collective Functions MPI_BCAST

broadcasts a message from the process with rank root to allprocesses of the group, itself included

C int MPI_Bcast(void* buffer, int count, MPI_Datatype datatype, int

root, MPI_Comm comm ) Input Parameters

count: number of entries in buffer (integer) datatype: data type of buffer (handle)

root: rank of broadcast root (integer)


Input / Output Parameter buffer: starting address of buffer (choice)

Fortran

MPI_BCAST(BUFFER, COUNT, DATATYPE, ROOT, COMM, IERROR) BUFFER(*) INTEGER COUNT, DATATYPE, ROOT, COMM, IERROR


8/45

Collective Functions MPI_BCAST

A

AAA A


9/45

Collective Functions MPI_GATHER

Each process (root process included) sends the contents of itssend buffer to the root process.

The root process receives the messages and stores them inrank order

C

int MPI_Gather(void* sendbuf, int sendcount, MPI_Datatypesendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype,int root, MPI_Comm comm)

Input Parameters sendbuf: starting address of send buffer (choice)

sendcount: number of elements in send buffer (integer) sendtype: data type of send buffer elements (handle)

recvcount: number of elements for any single receive (integer,significant only at root)

recvtype: data type of recv buffer elements (significant only at root)(handle)

root: rank of receiving process (integer)



10/45


Output Parameter

recvbuf: address of receive buffer (choice,significant only at root)

Fortran MPI_GATHER(SENDBUF, SENDCOUNT,

SENDTYPE, RECVBUF, RECVCOUNT, RECVTYPE,ROOT, COMM, IERROR)

SENDBUF(*), RECVBUF(*)

INTEGER SENDCOUNT, SENDTYPE,RECVCOUNT, RECVTYPE, ROOT, COMM,IERROR


11/45


BA C D

DCA

A B C D

B


12/45

Collective Functions MPI_SCATTER

MPI_SCATTER is the inverse operation to MPI_GATHER C

int MPI_Scatter(void* sendbuf, int sendcount,MPI_Datatype sendtype, void* recvbuf, int recvcount,MPI_Datatype recvtype, int root, MPI_Comm comm)

Input Parameters sendbuf: address of send buffer (choice, significant only at

root) sendcount: number of elements sent to each process (integer,

significant only at root) sendtype: data type of send buffer elements (significant only

at root) (handle) recvcount: number of elements in receive buffer (integer) recvtype: data type of receive buffer elements (handle) root: rank of sending process (integer) comm: communicator (handle)


13/45


Output Parameter recvbuf: address of receive buffer (choice)

Fortran MPI_SCATTER(SENDBUF, SENDCOUNT,

SENDTYPE, RECVBUF, RECVCOUNT,RECVTYPE, ROOT, COMM, IERROR)


INTEGER SENDCOUNT, SENDTYPE,RECVCOUNT, RECVTYPE, ROOT, COMM,IERROR


14/45


A B C D

DCA

A B C D

B


15/45


MPI_ALLGATHER

MPI_ALLGATHER can be thought of as MPI_GATHER, butwhere all processes receive the result, instead of just the root.

The jth block of data sent from each process is received byevery process and placed in the jth block of the buffer recvbuf.

C

int MPI_Allgather(void* sendbuf, int sendcount, MPI_Datatypesendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype,MPI_Comm comm)

Input Parameters sendbuf: starting address of send buffer (choice)

sendcount: number of elements in send buffer (integer) sendtype: data type of send buffer elements (handle)

recvcount: number of elements received from any process (integer)

recvtype: data type of receive buffer elements (handle)



16/45


MPI_ALLGATHER

Output Parameter

recvbuf: address of receive buffer (choice)

Fortran

MPI_ALLGATHER(SENDBUF, SENDCOUNT,SENDTYPE, RECVBUF, RECVCOUNT, RECVTYPE,COMM, IERROR)


INTEGER SENDCOUNT, SENDTYPE,RECVCOUNT, RECVTYPE, COMM, IERROR


17/45


MPI_ALLGATHER

BA C D

DCA

A B C D

B

A B C D A B C DA B C D

MPI_ALLGATHER


18/45

Collective Functions MPI_ALLTOALL

Extension of MPI_ALLGATHER to the case where eachprocess sends distinct data to each of the receivers.The jth block sent from process i is received by process

j and is placed in the ith block of recvbuf C

int MPI_Alltoall(void* sendbuf, int sendcount,MPI_Datatype sendtype, void* recvbuf, int recvcount,MPI_Datatype recvtype, MPI_Comm comm)

Input Parameters sendbuf: starting address of send buffer (choice) sendcount: number of elements sent to each process (integer) sendtype: data type of send buffer elements (handle)

recvcount: number of elements received from any process(integer) recvtype: data type of receive buffer elements (handle) comm: communicator (handle)


19/45


Output Parameter recvbuf: address of receive buffer (choice)

Fortran MPI_ALLTOALL(SENDBUF, SENDCOUNT,

SENDTYPE, RECVBUF, RECVCOUNT,RECVTYPE, COMM, IERROR)


INTEGER SENDCOUNT, SENDTYPE,RECVCOUNT, RECVTYPE, COMM,IERROR


20/45


E F G HA B C D I J K L M N O P

A B C D

A E I M

E F G H

B F J N

I J K L

C G K O

M N O P

D H L P

Rank 0 Rank 1 Rank 2 Rank 3

MPI_ALLTOALL


21/45

Collective Functions MPI_REDUCE

MPI_REDUCE combines the elements provided in the inputbuffer (sendbuf) of each process in the group, using theoperation op, and returns the combined value in the outputbuffer (recvbuf) of the process with rank root

C int MPI_Reduce(void* sendbuf, void* recvbuf, int count,

MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm) Input Parameters

sendbuf: address of send buffer (choice)

count: number of elements in send buffer (integer)

datatype: data type of elements of send buffer (handle) op: reduce operation (handle)

root: rank of root process (integer)


Output Parameter recvbuf: address of receive buffer (choice, significant only at root)


22/45


Fortran MPI_REDUCE(SENDBUF, RECVBUF, COUNT, DATATYPE, OP, ROOT, COMM,

IERROR) SENDBUF(*), RECVBUF(*) INTEGER COUNT, DATATYPE, OP, ROOT, COMM, IERROR

Predefined Reduce Operations [ 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 (return the max and an integer,

which is the rank storing the max value)

[ MPI_MINLOC] min value and location


23/45



A B C D E F G H I J K L M N O P

AoEoIoM


In this case, root = 1

if count = 2, therewill be BoFoJoN inthe 2nd element ofthe array


24/45


MPI_ALLREDUCE

Variants of the reduce operations wherethe result is returned to all processes inthe group

The all-reduce operations can be

implemented as a reduce, followed by abroadcast. However, a directimplementation can lead to betterperformance.

C int MPI_Allreduce(void* sendbuf, void* recvbuf,

int count, MPI_Datatype datatype, MPI_Op op,MPI_Comm comm)


25/45


MPI_ALLREDUCE

Input Parameters sendbuf: starting address of send buffer (choice) count: number of elements in send buffer (integer) datatype: data type of elements of send buffer

(handle) op: operation (handle) comm: communicator (handle)

Output Parameter recvbuf: starting address of receive buffer (choice)

Fortran

MPI_ALLREDUCE(SENDBUF, RECVBUF, COUNT,DATATYPE, OP, COMM, IERROR) SENDBUF(*), RECVBUF(*) INTEGER COUNT, DATATYPE, OP, COMM, IERROR


26/45


MPI_ALLREDUCE


A B C D E F G H I J K L M N O P

AoEoIoM



27/45


MPI_REDUCE_SCATTER

Variants of each of the reduce operations where the result isscattered to all processes in the group on return.

MPI_REDUCE_SCATTER first does an element-wise reductionon vector of count=irecvcounts[i] elements in the sendbuffer defined by sendbuf, count and datatype.

Next, the resulting vector of results is split into n disjointsegments, where n is the number of members in the group.Segment i contains recvcounts[i] elements.

The ith segment is sent to process i and stored in the receivebuffer defined by recvbuf, recvcounts[i] and datatype.

The MPI_REDUCE_SCATTER routine is functionally equivalentto: A MPI_REDUCE operation function with count equal to thesum of recvcounts[i] followed by MPI_SCATTERV with

sendcounts equal to recvcounts. However, a directimplementation may run faster.


28/45


MPI_REDUCE_SCATTER

C int MPI_Reduce_scatter(void* sendbuf, void* recvbuf, int

*recvcounts, MPI_Datatype datatype, MPI_Op op, MPI_Commcomm)

Input Parameters sendbuf: starting address of send buffer (choice) recvcounts: integer array specifying the number of elements in result

distributed to each process. Array must be identical on all calling processes. datatype: data type of elements of input buffer (handle) op: operation (handle) comm: communicator (handle)


Fortran MPI_REDUCE_SCATTER(SENDBUF, RECVBUF, RECVCOUNTS,

DATATYPE, OP, COMM, IERROR) SENDBUF(*), RECVBUF(*) INTEGER RECVCOUNTS(*), DATATYPE, OP, COMM, IERROR


29/45


MPI_REDUCE_SCATTER

A B C DRank 0recvcounts = 1

E F G H

I J K L

M N O P

Rank 1

recvcounts = 2

Rank 2recvcounts = 0

Rank 3recvcounts = 1

AoEoIoM

A B C D

E F G H

I J K L

M N O P

BoFoJoN

CoGoKoO

DoHoLoP


30/45

Collective Functions MPI_SCAN

MPI_SCAN is used to perform a prefix reductionon data distributed across the group. Theoperation returns, in the receive buffer of theprocess with rank i, the reduction of the values in

the send buffers of processes with ranks 0,...,i(inclusive). The type of operations supported,their semantics, and the constraints on send andreceive buffers are as for MPI_REDUCE.

C

int MPI_Scan(void* sendbuf, void* recvbuf, intcount, MPI_Datatype datatype, MPI_Op op,MPI_Comm comm )


31/45


Input Parameters sendbuf: starting address of send buffer (choice) count: number of elements in input buffer (integer) datatype: data type of elements of input buffer

(handle) op: operation (handle) comm: communicator (handle)


Fortran MPI_SCAN(SENDBUF, RECVBUF, COUNT, DATATYPE,

OP, COMM, IERROR) SENDBUF(*), RECVBUF(*) INTEGER COUNT, DATATYPE, OP, COMM, IERROR


32/45


A B C DRank 0

E F G H

I J K L

M N O P

Rank 1

Rank 2

Rank 3

AoEoIoM

A B C D

E F G H

I J K L

M N O P

AoEoI

AoE

A


33/45

Example MPI_BCAST

To demonstrate how to useMPI_BCAST to distribute an array toother process


34/45

Example MPI_BCAST (C)

/* // root broadcast the array to all processes */

#include #include

#define SIZE 10

main( int argc, char** argv)

{ int my_rank; // the rank of each proc int array[SIZE]; int root = 0; // the rank of root int i; MPI_Comm comm = MPI_COMM_WORLD;

MPI_Init(&argc, &argv); MPI_Comm_rank(comm, &my_rank);

if (my_rank == 0) { for (i = 0; i < SIZE; i ++) { array[i] = i; } }


35/45

Example MPI_BCAST (C)

else { for (i = 0; i < SIZE; i ++) {

array[i] = 0; } }

printf("Proc %d: (Before Broadcast) ", my_rank); for (i = 0; i < SIZE; i ++)

{ printf("%d ", array[i]); } printf("\n");

MPI_Bcast(array, SIZE, MPI_INT, root, comm);

printf("Proc %d: (After Broadcast) ", my_rank); for (i = 0; i < SIZE; i ++) { printf("%d ", array[i]); } printf("\n");

MPI_Finalize(); return 0; }


36/45

Example MPI_BCAST (Fortran)

C /* C * root broadcast the array to all processes C */

PROGRAM main INCLUDE 'mpif.h'

PARAMETER (SIZE = 10) INTEGER my_rank, ierr, root, i INTEGER array(SIZE) INTEGER comm INTEGER arraysize

root = 0 comm = MPI_COMM_WORLD arraysize = SIZE


37/45

Example MPI_BCAST (Fortran)

CALL MPI_INIT(ierr) CALL MPI_COMM_RANK(comm, my_rank, ierr)

IF (my_rank.EQ.0) THEN DO i = 1, SIZE array(i) = i

END DO ELSE DO i = 1, SIZE array(i) = 0 END DO END IF

WRITE(6, *) "Proc ", my_rank, ": (Before Broadcast)", (array(i), i=1, SIZE) CALL MPI_Bcast(array, arraysize, MPI_INTEGER, root, comm, ierr)

WRITE(6, *) "Proc ", my_rank, ": (After Broadcast)", (array(i), i=1, SIZE)

call MPI_FINALIZE(ierr) end


38/45

Case Study 1 MPI_SCATTER and

MPI_REDUCE

Master distributes (scatters) anarray across processes. Processesadd their elements, then combine

sum in master through a reductionoperation.

Step 1 Proc 0 initializes a 16 integers array

Proc 0: {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,12, 13, 14, 15, 16}


39/45


MPI_REDUCE

Step 2

Scatter array among all processes

Proc 0: {1, 2, 3, 4}

Proc 1: {5, 6, 7, 8}

Proc 2: {9, 10, 11, 12}

Proc 3: {13, 14, 15, 16}

Step 3 Each process does some calculations


40/45


MPI_REDUCE

Step 4 Reduce to Proc 0 Proc 0: Total Sum

C mpi_scatter_reduce01.c

Compilation: mpicc mpi_scatter_reduce01.c o mpi_scatter_reduce01

Run: mpirun np 4 mpi_scatter_reduce01

Fortran mpi_scatter_reduce01.f

Compilation: mpif77 mpi_scatter_reduce01.f o mpi_scatter_reduce01

Run: mpirun np 4 mpi_scatter_reduce01


41/45

Case Study 2 MPI_GATHER

Matrix Multiplication

760

686

612

538

20161912188174

20151911187173

20141910186172

2013199185171

20

19

18

17

161284

151173

141062

13951

Algorithm:

{4x4 matrix A} x {4x1 vector x} = product

Each process stores a row of A and a single entry ofx

Use 4 gather operations to place a full copy of x ineach process, then perform multiplications


42/45



Step 1: Initialization Proc 0: {1 5 9 13}, {17} Proc 1: {2 6 10 14}, {18} Proc 2: {3 7 11 15}, {19} Proc 3: {4 8 12 16}, {20}

Step 2: Perform 4 times MPI_GATHER to gather the column

matrix to each process Proc0: {1 5 9 13}, {17 18 19 20}

Proc1: {2 6 10 14}, {17 18 19 20} Proc2: {3 7 11 15}, {17 18 19 20} Proc3: {4 8 12 16}, {17 18 19 20}


43/45



Step 3:

Perform multiplication

Proc 0: 1x17+5x18+9x19+13x20=538

Proc 1: 2x17+6x18+10x19+14x20=612 Proc 2: 3x17+7x18+11x19+15x20=686

Proc 3: 4x17+8x18+12x19+16x20=760

Step 4:

Gather all processinner product into master

process and display the result


44/45



C mpi_gather01.c Compilation:

mpicc mpi_gather01.c o mpi_gather01

Run: mpirun np 4 mpi_gather01

Fortran mpi_gather01.f Compilation:

mpif77 mpi_gather01.f o mpi_gather01 Run:

mpirun np 4 mpi_gather01


45/45

END

Date post:	03-Jun-2018
Category:	Documents
Upload:	ing-eduardo-p-hdez
View:	221 times
Download:	0 times

Collective Comm

Documents