Summary of MPI commands Luis Basurto

Summary of MPI commands

Luis Basurto

Large scale systems

• Shared Memory systems

– Memory is shared among processors

• Distributed memory systems

– Each processor has its own memory

MPI

• Created in 1993 as an open standard by large scale system users and creators.

• Each system provider implements MPI for its systems.

• Currently at version MPI-2.0

• Other implementations of MPI such as MPICH,OpenMPI.

How many commands?

• 130+ commands

• 6 basic commands (we will cover 11)

• C and Fortran bindings

How does an MPI program work?

Start program on n processors

For i=0 to n-1

Run a copy of program on processor i

Pass messages between processors

End For

End Program

What are messages?

• Simplest message: an array of data of one type.

• Predefined types correspond to commonly used types in a given language

– MPI_REAL (Fortran), MPI_FLOAT (C)

– MPI_DOUBLE_PRECISION (Fortran), MPI_DOUBLE (C)

– MPI_INTEGER (Fortran), MPI_INT (C)

• User can define more complex types and send packages.

Before we start

Include MPI in our program

• In C/C++

#include “mpi.h”

• In Fortran

include 'mpif.h'

• In C MPI calls are functions

MPI_Init();

• In Fortran they are subroutines

call MPI_Init(ierror)

A note about Fortran

• All calls to MPI include an extra parameter, an error code of type integer.

• Used to test the success of the function (i.e. The function executed correctly).

Basic Communication

• Data values are transferred from one processor to another

– One processor sends the data

– Another receives the data

• Synchronous

– Call does not return until the message is sent or received

• Asynchronous

– Call indicates a start of send or receive, and another call is made to determine if finished

MPI_init()

• Initializes the MPI environment

• Every MPI program must have this.

• C

– MPI_Init();

• If using command line arguments

– MPI_Init( &argc, &argv );

• Fortran

– call MPI_Init(ierror)

MPI_Finalize()

• Stops the MPI environment

• Every MPI program must have this at the end.

• C

MPI_Finalize ( );

• Fortran

call MPI_Finalize(ierr)

MPI_Comm_size()

• Returns the size of the communicator (number of nodes) that we are working with.

• CMPI_Comm_size ( MPI_COMM_WORLD, &p );

• Fortrancall MPI_COMM_SIZE(MPI_COMM_WORLD, p, ierr )

MPI_Comm_rank()

• Return the zero based rank (id number) of the node executing the program.

• CMPI_Comm_rank ( MPI_COMM_WORLD, &id );

• Fortrancall MPI_COMM_RANK(MPI_COMM_WORLD, my_rank, ierr )

A note con communicators

• MPI_COMM_WORLD is the default communicator (all nodes in the cluster)

• Communicators can be created dynamically in order to assign certain tasks to certain nodes (processors).

• Inter communicator message passing is possible.

MPI_Send()

• C

MPI_Send(void *buf, int count, MPI_Datatype dtype, int dest, int tag, MPI_Comm comm);

• Fortran

Call MPI_Send(buffer, count, datatype, destination,tag,communicator, ierr)

MPI_Recv()

• C

MPI_Recv(void *buf, int count, MPI_Datatype dtype, int src,int tag, MPI_Comm comm,

MPI_Status *stat);

• Fortran

Call MPI_Recv(buffer, count, datatype,source, tag, communicator,status, ierr)

MPI_Bcast()

• Send message to all nodes

• C

MPI Bcast(void * buf, int count, MPI_Datatype dtype, int root, MPI Comm comm);

• Fortran

CALL MPI_BCAST(buff, count, MPI_TYPE, root, comm, ierr)

MPI_Reduce()

• Receive message from all nodes, do operation on every element.

• C

MPI_Reduce(void *sbuf, void* rbuf, int count, MPI_Datatype dtype, MPI_Op op, int root, MPI Comm comm);

• Fortran

CALL MPI_REDUCE(sndbuf, recvbuf,count, datatype,operator,root,comm,ierr)

MPI_Barrier()

• Used as a synchronization barrier, every node that reaches this point must wait until all nodes reach it in order to proceed.

• C

MPI_Barrier(MPI_COMM_WORLD);

• Fortran

call MPI_Barrier(MPI_COMM_WORLD,ierr)

MPI_Scatter()

• Parcels out data from the root to every member of the group in linear order by node

• C

MPI_Scatter(void *sbuf, int scount, MPI_Datatype sdtype,void *rbuf, int rcount, MPI_Datatype rdtype,int root, MPI_Comm comm)

• Fortran

CALL MPI_SCATTER(sndbuf,scount,datatype, recvbuf,rcount,rdatatype,root,comm, ierr)

MPI_Scatter

Node 0

Node 1

Node 2

Node 3

MPI_Gather()

• C

MPI_Gather(void *sbuf, int scount, MPI_Datatype sdtype,void *rbuf, int rcount, MPI_Datatype rdtype,int root, MPI_Comm comm)

Fortran

CALL MPI_GATHER(sndbuf,scount,datatype, recvbuf,rcount,rdatatype,root,comm,ierr)

Deadlock

• The following code may provoke deadlock

if(rank==0) {

MPI_COMM_WORLD.Send(vec1,vecsize,MPI::DOUBLE,1,0);

MPI_COMM_WORLD.Recv(vec2,vecsize,MPI::DOUBLE,1,MPI::ANY_TAG);

}

if(rank==1) {

MPI_COMM_WORLD.Send(vec3,vecsize,MPI::DOUBLE,0,0);

MPI_COMM_WORLD.Recv(vec4,vecsize,MPI::DOUBLE,0,MPI::ANY_TAG);

}

Bcast

• Must be called by all nodes, the following code will not work

if(rank==0) {

MPI_Bcast(&value, 1, MPI_int,0, MPI_comm_world);

}

else {

/* Do something else */

}

Questions