Lecture11

1
San DIEGO SUPERCOMPUTER CENTER
NATIONAL PARTNERSHIP FOR ADVANCED COMPUTATIONAL
INFRASTRUCTURE
Point to Point Communications in MPI
• Basic operations of Point to Point (PtoP)
communication and issues of deadlock
• Several steps are involved in the PtoP
communication
• Sending process
– data is copied to the user buffer by the user
– User calls one of the MPI send routines
– System copies the data from the user buffer to the
system buffer
– System sends the data from the system buffer to the
destination processor

2
INFRASTRUCTURE
Point to Point Communications in MPI
• Receiving process
– User calls one of the MPI receive subroutines
– System receives the data from the source process, and
copies it to the system buffer
– System copies the data from the system buffer to the
user buffer
– User uses the data in the user buffer

3
INFRASTRUCTURE
sendbuf
Call send routine
Now sendbuf can be reused
Process 0 : User mode Kernel mode
Copying data from sendbuf to
systembuf
Send data from sysbuf to
dest
data
Process 1 : User mode Kernel mode
Call receive routine
receive data from src to
systembuf
Copying data from sysbuf
to recvbuf
sysbuf
sysbuf
recvbuf
Now recvbuf contains
valid data

4
INFRASTRUCTURE
Unidirectional communication
• Blocking send and blocking receive
if (myrank == 0) then
call MPI_Send(…)
elseif (myrank == 1) then
call MPI_Recv(….)
endif
• Non-blocking send and blocking receive
if (myrank == 0) then
call MPI_ISend(…)
call MPI_Wait(…)
else if (myrank == 1) then
call MPI_Recv(….)
endif

5
INFRASTRUCTURE
• Blocking send and non-blocking recv
if (myrank == 0 ) then
call MPI_Send(…..)
call MPI_Irecv (…)
call MPI_Wait(…)
endif
• Non-blocking send and non-blocking recv
call MPI_Isend (…)
call MPI_Wait (…)
call MPI_Irecv (….)
call MPI_Wait(..)
endif
Unidirectional communication

6
INFRASTRUCTURE
Bidirectional communication
• Need to be careful about deadlock when two processes exchange data with
each other
• Deadlock can occur due to incorrect order of send and recv or due to limited
size of the system buffer
sendbuf
recvbuf
Rank 0 Rank 1
recvbuf
sendbuf

7
INFRASTRUCTURE
• Case 1 : both processes call send first, then recv
call MPI_Send(….)
call MPI_Recv (…)
call MPI_Send(….)
call MPI_Recv(….)
endif
• No deadlock as long as system buffer is larger than send buffer
• Deadlock if system buffer is smaller than send buf
• If you replace MPI_Send with MPI_Isend and MPI_Wait, it is still the same
• Moral : there may be error in coding that only shows up for
larger problem size

8
INFRASTRUCTURE
• Following is free from deadlock
call MPI_Isend(….)
call MPI_Recv (…)
call MPI_Wait(…)
call MPI_Isend(….)
call MPI_Recv(….)
call MPI_Wait(….)
endif

9
INFRASTRUCTURE
• Case 2 : both processes call recv first, then send
call MPI_Recv(….)
call MPI_Send (…)
call MPI_Recv(….)
call MPI_Send(….)
endif
• The above will always lead to deadlock (even if you replace
MPI_Send with MPI_Isend and MPI_Wait)

10
INFRASTRUCTURE
• The following code can be safely executed
call MPI_Irecv(….)
call MPI_Send (…)
call MPI_Wait(…)
call MPI_Irecv(….)
call MPI_Send(….)
call MPI_Wait(….)
endif

11
INFRASTRUCTURE
• Case 3 : one process call send and recv in this order, and the other
calls in the opposite order
call MPI_Send(….)
call MPI_Recv(…)
call MPI_Recv(….)
call MPI_Send(….)
endif
• The above is always safe
• You can replace both send and recv on both processor with Isend
and Irecv

12
INFRASTRUCTURE
Scatter and Gather
A Ap0
p1
p2
p3
p0
p1
p2
p3
A
A
A
broadcast
scatterA B C D A
B
C
D
gather
A
B
C
D
A B C D
A B C D
A B C D
A B C D
all gather
p0
p1
p2
p3
p0
p1
p2
p3
p0
p1
p2
p3
p0
p1
p2
p3

13
INFRASTRUCTURE
Scatter Operation using MPI_Scatter
• Similar to Broadcast but sends a section of
an array to each processors
A(0) A(1) A(2) . . ………. A(N-1)
P0 P1 P2 . . . Pn-1
Goes to processors:
Data in an array on root node:

14
INFRASTRUCTURE
MPI_Scatter
• C
– int MPI_Scatter(&sendbuf, sendcnts, sendtype, &recvbuf,
recvcnts, recvtype, root, comm );
• Fortran
– MPI_Scatter(sendbuf,sendcnts,sendtype,
recvbuf,recvcnts,recvtype,root,comm,ierror)
• Parameters
– sendbuf is an array of size (number processors*sendcnts)
– sendcnts number of elements sent to each processor
– recvcnts number of element(s) obtained from the root processor
– recvbuf contains element(s) obtained from the root processor, may
be an array

15
INFRASTRUCTURE
Scatter Operation using MPI_Scatter
• Scatter with Sendcnts = 2
A(0) A(2) A(4) . . . A(2N-2)
A(1) A(3) A(5) . . . A(2N-1)
P0 P1 P2 . . . Pn-1
B(0) B(0) B(0) B(0)
B(1) B(1) B(1) B(1)
Goes to processors:
Data in an array on root node:

16
INFRASTRUCTURE
Gather Operation using MPI_Gather
• Used to collect data from all processors to
the root, inverse of scatter
• Data is collected into an array on root
processor
A(0) A(1) A(2) . . . A(N-1)
P0 P1 P2 . . . Pn-1
A0 A1 A2 . . . An-1
Data from various
Processors:
Goes to an array
on root node:

17
INFRASTRUCTURE
MPI_Gather
• C
– int MPI_Gather(&sendbuf,sendcnts, sendtype, &recvbuf,
recvcnts,recvtype,root, comm );
• Fortran
– MPI_Gather(sendbuf,sendcnts,sendtype,
recvbuf,recvcnts,recvtype,root,comm,ierror)
• Parameters
– sendcnts number of elements sent from each processor
– sendbuf is an array of size sendcnts
– recvcnts number of elements obtained from each processor
– recvbuf of size recvcnts*number of processors

18
INFRASTRUCTURE
Code for Scatter and Gather
• A parallel program to scatter data using
MPI_Scatter
• Each processor sums the data
• Use MPI_Gather to get the data back to the
root processor
• Root processor prints the global data
• See attached Fortran and C code

19
INFRASTRUCTURE
module mpi
!DEC$ NOFREEFORM
include "mpif.h“
!DEC$ FREEFORM
end module
! This program shows how to use MPI_Scatter and MPI_Gather
! Each processor gets different data from the root processor
! by way of mpi_scatter. The data is summed and then sent back
! to the root processor using MPI_Gather. The root processor
! then prints the global sum.
module global
integer numnodes,myid,mpi_err
integer, parameter :: mpi_root=0
end module
subroutine init
use mpi
use global
implicit none
! do the mpi init stuff
call MPI_INIT( mpi_err )
call MPI_COMM_SIZE( MPI_COMM_WORLD, numnodes, mpi_err )
call MPI_Comm_rank(MPI_COMM_WORLD, myid, mpi_err)

20
INFRASTRUCTURE
end subroutine init
program test1
use mpi
use global
implicit none
integer, allocatable :: myray(:),send_ray(:),back_ray(:)
integer count
integer size,mysize,i,k,j,total
call init
! each processor will get count elements from the root
count=4
allocate(myray(count))
! create the data to be sent on the root
if(myid == mpi_root)then
size=count*numnodes
allocate(send_ray(0:size-1))
allocate(back_ray(0:numnodes-1))
do i=0,size-1
send_ray(i)= i
enddo
endif

21
INFRASTRUCTURE
call MPI_Scatter( send_ray, count, MPI_INTEGER, &
myray, count, MPI_INTEGER, &
mpi_root, MPI_COMM_WORLD,mpi_err)
! each processor does a local sum
total=sum(myray)
write(*,*)"myid= ",myid," total= ",total
! send the local sums back to the root
call MPI_Gather( total, 1, MPI_INTEGER, &
back_ray, 1, MPI_INTEGER, &
mpi_root, MPI_COMM_WORLD,mpi_err)
! the root prints the global sum
if(myid == mpi_root)then
write(*,*)"results from all processors= ",sum(back_ray)
endif
call mpi_finalize(mpi_err)
end program

22
INFRASTRUCTURE
#include <mpi.h>
#include <stdio.h>
#include <stdlib.h>
/*! This program shows how to use MPI_Scatter and MPI_Gather
! Each processor gets different data from the root processor
! by way of mpi_scatter. The data is summed and then sent back
! to the root processor using MPI_Gather. The root processor
! then prints the global sum. */
/* globals */
int numnodes,myid,mpi_err;
#define mpi_root 0
/* end globals */
void init_it(int *argc, char ***argv);
void init_it(int *argc, char ***argv) {
mpi_err = MPI_Init(argc,argv);
mpi_err = MPI_Comm_size( MPI_COMM_WORLD, &numnodes );
mpi_err = MPI_Comm_rank(MPI_COMM_WORLD, &myid); }

23
INFRASTRUCTURE
int main(int argc,char *argv[]){
int *myray,*send_ray,*back_ray;
int count;
int size,mysize,i,k,j,total;
init_it(&argc,&argv);
/* each processor will get count elements from the root */
count=4;
myray=(int*)malloc(count*sizeof(int));
/* create the data to be sent on the root */
if(myid == mpi_root){
size=count*numnodes;
send_ray=(int*)malloc(size*sizeof(int));
back_ray=(int*)malloc(numnodes*sizeof(int));
for(i=0;i<size;i++)
send_ray[i]=i;
}
/* send different data to each processor */

24
INFRASTRUCTURE
mpi_err = MPI_Scatter( send_ray, count, MPI_INT, myray, count, MPI_INT,
mpi_root, MPI_COMM_WORLD);
/* each processor does a local sum */
total=0;
for(i=0;i<count;i++)
total=total+myray[i];
printf("myid= %d total= %dn ",myid,total);
/* send the local sums back to the root */
mpi_err = MPI_Gather(&total, 1, MPI_INT, back_ray, 1, MPI_INT,
/* the root prints the global sum */
total=0;
for(i=0;i<numnodes;i++)
total=total+back_ray[i];
printf("results from all processors= %d n ",total);
}
mpi_err = MPI_Finalize();}

25
INFRASTRUCTURE
Output of previous code on 4 procs
ultra:/work/majumdar/examples/mpi % bsub -q hpc -m ultra -I -n 4 ./a.out
Job <48051> is submitted to queue <hpc>.
<<Waiting for dispatch ...>>
<<Starting on ultra>>
myid= 1 total= 22
myid= 2 total= 38
myid= 3 total= 54
myid= 0 total= 6
results from all processors= 120
( 0 through 15 added up = (15) (15 + 1) /2 = 120)

26
INFRASTRUCTURE
Global Sum with MPI_Reduce
2d array spread across processors
A0+A1+A2 B0+B1+B2 C0+C1+C2NODE 0
NODE 1
NODE 2
X(0) X(1) X(2)
A0 B0 C0
A1 B1 C1
A2 B2 C2
NODE 0
NODE 1
NODE 2
X(0) X(1) X(2)

27
INFRASTRUCTURE
MPI_Allgather and MPI_Allreduce
• Gather and Reduce come in an "ALL"
variation
• Results are returned to all processors
• The root parameter is missing from the call
• Similar to a gather or reduce followed by a
broadcast

28
INFRASTRUCTURE
Global Sum with MPI_Allreduce
2d array spread across processors
A0 B0 C0
A1 B1 C1
A2 B2 C2
X(0) X(1) X(2)
NODE 0
NODE 1
NODE 2
A0+A1+A2 B0+B1+B2 C0+C1+C2
A0+A1+A2 B0+B1+B2 C0+C1+C2
A0+A1+A2 B0+B1+B2 C0+C1+C2
X(0) X(1) X(2)
NODE 0
NODE 1
NODE 2

29
INFRASTRUCTURE
All to All communication with MPI_Alltoall
• Each processor sends and receives data
to/from all others
• C
– int MPI_Alltoall(&sendbuf,sendcnts, sendtype, &recvbuf,
recvcnts, recvtype, MPI_Comm);
• Fortran
– call MPI_Alltoall(sendbuf,sendcnts,sendtype,
recvbuf,recvcnts,recvtype,comm,ierror)

30
INFRASTRUCTURE
b0 b1 b2 b3
c0 c1 c2 c3
d0 d1 d2 d3
a0 a1 a2 a3
a1 b1 c1 d1
a2 b2 c2 d2
a3 b3 c3 d3
a0 b0 c0 d0MPI_AlltoallP0
P1
P2
P3
P0
P1
P2
P3

31
INFRASTRUCTURE
All to All with MPI_Alltoall
• Parameters
– sendcnts # of elements sent to each processor
– sendbuf is an array of size sendcnts
– recvcnts # of elements obtained from each processor
– recvbuf of size recvcnts
• Note that both send buffer and receive
buffer must be an array of size of the
number of processors
• See attached Fortran and C codes

32
INFRASTRUCTURE
module mpi
!DEC$ NOFREEFORM
include "mpif.h“
!DEC$ FREEFORM
end module
! This program shows how to use MPI_Alltoall. Each processor
! send/rec a different random number to/from other processors.
module global
integer numnodes,myid,mpi_err
integer, parameter :: mpi_root=0
end module
subroutine init
use mpi
use global
implicit none
! do the mpi init stuff
call MPI_Comm_rank(MPI_COMM_WORLD, myid, mpi_err)
end subroutine init

33
INFRASTRUCTURE
program test1
use mpi
use global
implicit none
integer, allocatable :: scounts(:),rcounts(:)
integer ssize,rsize,i,k,j
real z
call init
! counts and displacement arrays
allocate(scounts(0:numnodes-1))
allocate(rcounts(0:numnodes-1))
call seed_random
! find data to send
do i=0,numnodes-1
call random_number(z)
scounts(i)=nint(10.0*z)+1
Enddo
write(*,*)"myid= ",myid," scounts= ",scounts

34
INFRASTRUCTURE
! send the data
call MPI_alltoall( scounts,1,MPI_INTEGER, &
rcounts,1,MPI_INTEGER, MPI_COMM_WORLD,mpi_err)
write(*,*)"myid= ",myid," rcounts= ",rcounts
call mpi_finalize(mpi_err)
end program
subroutine seed_random
use global
implicit none
integer the_size,j
integer, allocatable :: seed(:)
real z
call random_seed(size=the_size) ! how big is the intrisic seed?
allocate(seed(the_size)) ! allocate space for seed
do j=1,the_size ! create the seed
seed(j)=abs(myid*10)+(j*myid*myid)+100 ! abs is generic
enddo
call random_seed(put=seed) ! assign the seed
deallocate(seed)
end subroutine

35
INFRASTRUCTURE
#include <mpi.h>
#include <stdio.h>|
#include <stdlib.h>
/*! This program shows how to use MPI_Alltoall. Each processor
! send/rec a different random number to/from other processors. */
/* globals */
#define mpi_root 0
/* end module */
void seed_random(int id);
void random_number(float *z);
mpi_err = MPI_Comm_rank(MPI_COMM_WORLD, &myid);
}

36
INFRASTRUCTURE
int *sray,*rray;
int *scounts,*rcounts;
int ssize,rsize,i,k,j;
float z;
scounts=(int*)malloc(sizeof(int)*numnodes);
rcounts=(int*)malloc(sizeof(int)*numnodes);
/*! seed the random number generator with a
! different number on each processor*/
seed_random(myid);
/* find data to send */
for(i=0;i<numnodes;i++){
random_number(&z);
scounts[i]=(int)(10.0*z)+1;
}
printf("myid= %d scounts=",myid);
printf("%d ",scounts[i]);
printf("n");

37
INFRASTRUCTURE
/* send the data */
mpi_err = MPI_Alltoall( scounts,1,MPI_INT,
rcounts,1,MPI_INT, MPI_COMM_WORLD);
printf("myid= %d rcounts=",myid);
printf("%d ",rcounts[i]);
printf("n");
mpi_err = MPI_Finalize();}
void seed_random(int id){
srand((unsigned int)id);}
void random_number(float *z){
int i;
i=rand();
*z=(float)i/32767;
}

38
INFRASTRUCTURE
Output of previous code on 4 procs
ultra:/work/majumdar/examples/mpi % bsub -q hpc -m ultra -I -n 4 a.out
Job <48059> is submitted to queue <hpc>.
<<Waiting for dispatch ...>>
<<Starting on ultra>>
myid= 1 scounts= 6 2 4 6
myid= 1 rcounts= 7 2 7 3
--------------------------------------------
1 7 4 7 1 6 1 6
6 2 4 6 7 2 7 3
1 7 4 4 4 4 4 4
6 3 4 3 7 6 4 3

39
INFRASTRUCTURE
The variable or “V” operators
• A collection of very powerful but difficult
to setup global communication routines
• MPI_Gatherv: Gather different amounts of data
from each processor to the root processor
• MPI_Alltoallv: Send and receive different
amounts of data form all processors
• MPI_Allgatherv: Gather different amounts of data
from each processor and send all data to each
• MPI_Scatterv: Send different amounts of data to
each processor from the root processor
• We discuss MPI_Gatherv and MPI_Alltoallv

40
INFRASTRUCTURE
MPI_Gatherv
• C
– int MPI_Gatherv (&sendbuf, sendcnts, sendtype,
&recvbuf, &recvcnts, &rdispls,recvtype, comm);
• Fortran
– MPI_Gatherv (sendbuf, sendcnts, sendtype, recvbuf,
recvcnts, rdispls, recvtype, comm, ierror)
• Parameters:
– Recvcnts is now an array
– Rdispls is a displacement
· See attached codes

41
INFRASTRUCTURE
MPI_Gatherv
rank 0 = root rank 1 rank 2
1 2 3
sendbuf 2 3
sendbuf 3
sendbuf
recvcnts[0] 1 0 = rdispls[0]
2 2
3 4
3 5
recvbuf

42
INFRASTRUCTURE
MPI_Gatherv code
Sample program:
include ‘mpif.h’
integer isend(3), irecv(6)
integer ircnt(0:2), idisp(0:2)
data icrnt/1,2,3/ idisp/0,1,3/
call mpi_init(ierr)
call mpi_comm_size(MPI_COMM_WORLD, nprocs,ierr)
call mpi_comm_rank(MPI_COMM_WORLD,myrank,ierr)
do I = 1,myrank+1
isend(I) = myrank+1
enddo
iscnt = myrank + 1
call MPI_GATHERV(isend,iscnt,MPI_INTEGER,irecv,ircnt,idisp,MPI_INTEGER
& 0,MPI_COMM_WORLD, ierr)
if (myrank .eq. 0) then
print *, ‘irecv =‘, irecv
endif
call MPI_FINALIZE(ierr)
end
Sample execution:
% bsub –q hpc –m ultra –I –n 3 ./a.out
% 0: irecv = 1 2 2 3 3 3

43
INFRASTRUCTURE
#include <mpi.h>
#include <stdio.h>
#include <stdlib.h>
/*! This program shows how to use MPI_Gatherv. Each processor sends a
! different amount of data to the root processor. We use MPI_Gather
! first to tell the root how much data is going to be sent.*/
/* globals */
#define mpi_root 0
/* end of globals */
}

44
INFRASTRUCTURE
int *will_use;
int *myray,*displacements,*counts,*allray;
int size,mysize,i;
mysize=myid+1;
myray=(int*)malloc(mysize*sizeof(int));
for(i=0;i<mysize;i++)
myray[i]=myid+1;
/* counts and displacement arrays are only required on the root */
counts=(int*)malloc(numnodes*sizeof(int));

45
INFRASTRUCTURE
displacements=(int*)malloc(numnodes*sizeof(int));
}
/* we gather the counts to the root */
mpi_err = MPI_Gather((void*)myray,1,MPI_INT,
(void*)counts, 1,MPI_INT,
mpi_root,MPI_COMM_WORLD);
/* calculate displacements and the size of the recv array */
displacements[0]=0;
for( i=1;i<numnodes;i++){
displacements[i]=counts[i-1]+displacements[i-1];
}
size=0;
for(i=0;i< numnodes;i++)
size=size+counts[i];
allray=(int*)malloc(size*sizeof(int));
}

46
INFRASTRUCTURE
/* different amounts of data from each processor */
/* is gathered to the root */
mpi_err = MPI_Gatherv(myray, mysize, MPI_INT,
allray,counts,displacements,MPI_INT,
for(i=0;i<size;i++)
printf("%d ",allray[i]);
printf("n");
}
mpi_err = MPI_Finalize();
}
ultra% bsub –q hpc –m ultra –I –n 3 ./a.out
1 2 2 3 3 3

47
INFRASTRUCTURE
MPI_Alltoallv
• Send and receive different amounts of data
form all processors
• C
– int MPI_Alltoallv (&sendbuf, &sendcnts, &sdispls,
sendtype, &recvbuf, &recvcnts, &rdispls, recvtype,
comm );
• Fortran
– Call MPI_Alltoallv(sendbuf, sendcnts, sdispls,
sendtype, recvbuf, recvcnts, rdispls,recvtype,
comm,ierror);
• See attached code

48
INFRASTRUCTURE
MPI_Alltoallv
rank0 rank1 rank2
sendnts[0] 1 4 7 0=sdispls[0]
sendcnts[1] 2 5 8 1=sdispls[1]
2 5 8 2
sendcnts[3] 3 6 9 3=sdispls[2]
3 6 9 4
3 6 9 5
sendbuf sendbuf
recvcnts[0] 1 2 3 0=rdispls[0]
recvcnts[1] 4 2 3 1
recvcnts[3] 7 5 3 2
recvbuf 5 6 3=rdispls[1]
8 6 4
8 6 5
recvbuf 9 6=rdispls[2]
9 7
9 8

49
INFRASTRUCTURE
MPI_Alltoallv
proc#
recvcnts 0 1 2
0 1 2 3
1 1 2 3
2 1 2 3
proc#
rdispls 0 1 2
0 0 0 0
1 1 2 3
2 2 4 6

50
INFRASTRUCTURE
MPI_Alltoallv
Program alltoallv
integer isend(6), irecv(9)
integer iscnt(0:2), isdsp(0:2), ircnt(0), irdsp(0:2)
data isend/1,2,2,3,3,3/
data iscnt/1,2,3/ isdsp/0,1,3/
call MPI_INIT(ierr)
call MPI_COMM_SIZE(MPI_COMM_WORLD,nprocs, ierr)
call MPI_COMM_RANK(MIP_COMM_WORLD,myrank, ierr)
do i = 1,6
isend(i) = isend(i) + nprocs*myrank
enddo
do i = 0, nprocs – 1
ircnt(i) = myrank + 1
irdsp(i) = i* (myrank + 1)
enddo
print*, ‘isend=‘, isend
call MP_FLUSH(1)
call MPI_ALLTOALLV(isend,iscnt,isdsp,MPI_INTEGER,irecv, ircnt, irdsp,MPI_INTEGER,
MPI_COMM_WORLD, ierr)
print*, ‘irecv=‘,irecv
end

51
INFRASTRUCTURE
MPI_Alltoallv
Sample execution of mpialltoallv program:
% bsub –q hpc –m ultra –I –n 3
% 0: isend = 1 2 2 3 3 3
1: isend = 4 5 5 6 6 6
2: isend = 7 8 8 9 9 9
0: irecv = 1 4 7 0 0 0 0 0 0
1: irecv = 2 2 5 5 8 8 0 0 0
2: irecv = 3 3 3 6 6 6 9 9 9

52
INFRASTRUCTURE
#include <mpi.h>
#include <stdio.h>
#include <stdlib.h>
/*
! This program shows how to use MPI_Alltoallv. Each processor
! send/rec a different and random amount of data to/from other
! processors.
! We use MPI_Alltoall to tell how much data is going to be sent.
*/
/* globals */
#define mpi_root 0
/* end module */

53
INFRASTRUCTURE
void seed_random(int id);
void random_number(float *z);
}
int *sray,*rray;
int *sdisp,*scounts,*rdisp,*rcounts;
int ssize,rsize,i,k,j;
float z;
scounts=(int*)malloc(sizeof(int)*numnodes);
rcounts=(int*)malloc(sizeof(int)*numnodes);
sdisp=(int*)malloc(sizeof(int)*numnodes);
rdisp=(int*)malloc(sizeof(int)*numnodes);
/*

54
INFRASTRUCTURE
! seed the random number generator with a
! different number on each processor
*/
seed_random(myid);
/* find out how much data to send */
random_number(&z);
scounts[i]=(int)(10.0*z)+1;
}
printf("myid= %d scounts=",myid);
printf("%d ",scounts[i]);
printf("n");
/* tell the other processors how much data is coming */
mpi_err = MPI_Alltoall( scounts,1,MPI_INT,
rcounts,1,MPI_INT,
MPI_COMM_WORLD);

55
INFRASTRUCTURE
/* write(*,*)"myid= ",myid," rcounts= ",rcounts */
/* calculate displacements and the size of the arrays */
sdisp[0]=0;
sdisp[i]=scounts[i-1]+sdisp[i-1];
}
rdisp[0]=0;
rdisp[i]=rcounts[i-1]+rdisp[i-1];
}
ssize=0;
rsize=0;
ssize=ssize+scounts[i];
rsize=rsize+rcounts[i];
}

56
INFRASTRUCTURE
/* allocate send and rec arrays */
sray=(int*)malloc(sizeof(int)*ssize);
rray=(int*)malloc(sizeof(int)*rsize);
for(i=0;i<ssize;i++)
sray[i]=myid;
/* send/rec different amounts of data to/from each processor */
mpi_err = MPI_Alltoallv( sray,scounts,sdisp,MPI_INT,
rray,rcounts,rdisp,MPI_INT,
MPI_COMM_WORLD);
printf("myid= %d rray=",myid);
for(i=0;i<rsize;i++)
printf("%d ",rray[i]);
printf("n");
mpi_err = MPI_Finalize();
}

57
INFRASTRUCTURE
void seed_random(int id) {
srand((unsigned int)id);
}
void random_number(float *z){
int i;
i=rand();
*z=(float)i/32767;
}
Ultra output from 3 procs run:
0:myid= 0 scounts=1 7 4
0:myid= 0 rray=0 1 1 1 1 1 1 2
1:myid= 1 rray=0 0 0 0 0 0 0 1 1 2 2 2 2 2 2 2
2:myid= 2 rray=0 0 0 0 1 1 1 1 2 2 2 2

58
INFRASTRUCTURE
Derived types
• C and Fortran 90 have the ability to define
arbitrary data types that encapsulate reals,
integers, and characters.
• MPI allows you to define message data
types corresponding to your data types
• Can use these data types just as default
types

59
INFRASTRUCTURE
Derived types, Three main classifications:
• Contiguous Vectors: enable you to send
contiguous blocks of the same type of data
lumped together
• Noncontiguous Vectors: enable you to send
noncontiguous blocks of the same type of
data lumped together
• Abstract types: enable you to (carefully)
send C or Fortran 90 structures, don't send
pointers

60
INFRASTRUCTURE
Derived types, how to use them
• Three step process
– Define the type using
• MPI_Type_contiguous for contiguous vectors
• MPI_Type_vector for noncontiguous vectors
• MPI_Type_struct for structures
– Commit the type using
• MPI_Type_commit
– Use in normal communication calls
• MPI_Send(buffer, count, MY_TYPE,
destination,tag, MPI_COMM_WORLD, ierr)

61
INFRASTRUCTURE
MPI_Type_contiguous
• Defines a new data type of length count elements
from your old data type
• C
– MPI_Type_contiguous(int count, old_type, &new_type)
• Fortran
– Call MPI_TYPE_CONTIGUOUS(count, old_type,
new_type, ierror)
• Parameters
– Old_type: your base type
– New_type: a type count elements of Old_type
• See attached codes

62
INFRASTRUCTURE
MPI_TYPE_CONTIGUOUS
Sample program - Fortran:
program type_contiguous
integer ibuf(20)
call MPI_INIT(ierr)
call MPI_COMM_SIZE(MPI_COMM_WORLD,nprocs,ierr)
call MPI_COMM_RANK(MPI_COMM_WORLD,myrank,ierr)
if (myrank .eq. 0) then
do i = 1,20
ibuf(i) = I
enddo
endif
call MPI_TYPE_CONTIGUOUS(3,MPI_INTEGER,inewtype, ierr)
call MPI_TYPE_COMMIT(inewtype, ierr)
call MPI_BCAST(ibuf,3,inewtype,0,MPI_COMM_WORLD, ierr)
print*, ‘ibuf=‘,ibuf
end
Sample execution:
% bsub –q hpc –m ultra –I in 2 a.out
% 0 : ibuf =1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
1: ibuf = 1 2 3 4 5 6 7 8 9 0 0 0 0 0 0 0 0 0 0 0

63
INFRASTRUCTURE
MPI_Type_contiguous#include <stdio.h>
#include "mpi.h“
#include <math.h>
int main(argc,argv)
int argc;
char *argv[];{
int myid, numprocs, i , buffer[20];
MPI_Status status;
MPI_Datatype inewtype ;
MPI_Init(&argc,&argv);
MPI_Comm_size(MPI_COMM_WORLD,&numprocs);
MPI_Comm_rank(MPI_COMM_WORLD,&myid);
if (myid == 0) { for (i=0; i<20; i++)
buffer[i]=i ;}
if (myid == 1) { for (i=0; i<20; i++)
buffer[i]=0 ;}
MPI_Type_contiguous(3,MPI_INT,&inewtype);
MPI_Type_commit(&inewtype) ;
MPI_Bcast(buffer,3,inewtype,0,MPI_COMM_WORLD);
for(i=0;i<20;i++)
printf("%d ",buffer[i]);
printf("n");
MPI_Finalize(); }
Output on two processors :
0 1 2 3 4 5 6 7 8 0 0 0 0 0 0 0 0 0 0 0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

64
INFRASTRUCTURE
MPI_Type_vector
• Defines a datatype which consists of count blocks
each of length blocklength and stride
displacement between blocks
• C
– MPI_Type_vector(count, blocklength, stride, old_type,
*new_type)
• Fortran
– Call MPI_TYPE_VECTOR(count, blocklength, stride,
old_type, new_type, ierror)
• See attached codes

65
INFRASTRUCTURE
module mpi
!DEC$ NOFREEFORM
include "mpif.h"
!DEC$ FREEFORM
end module
!Shows how to use MPI_Type_vector to send noncontiguous blocks of data
!and MPI_Get_count and MPI_Get_elements to see the number of elements sent
program do_vect
use mpi
! include "mpif.h"
integer , parameter :: size=24
integer myid, ierr,numprocs
real*8 svect(0:size),rvect(0:size)
integer i,bonk1,bonk2,numx,stride,extent
integer MY_TYPE
integer status(MPI_STATUS_SIZE)
call MPI_INIT( ierr )
call MPI_COMM_RANK( MPI_COMM_WORLD, myid, ierr )
call MPI_COMM_SIZE( MPI_COMM_WORLD, numprocs, ierr )
stride=5
numx=(size+1)/stride

66
INFRASTRUCTURE
extent = 1
if(myid == 1)write(*,*)"numx=",numx," extent=",extent," stride=",stride
call MPI_Type_vector(numx,extent,stride,MPI_DOUBLE_PRECISION,MY_TYPE,ier
r)
call MPI_Type_commit(MY_TYPE, ierr )
if(myid == 0)then
do i=0,size
svect(i)=i
enddo
call MPI_Send(svect,1,MY_TYPE,1,100,MPI_COMM_WORLD,ierr)
endif
if(myid == 1)then
do i=0,size
rvect(i)=-1
enddo
call MPI_Recv(rvect,1,MY_TYPE,0,100,MPI_COMM_WORLD,status,ierr)
endif
if(myid == 1)then
call MPI_Get_count(status,MY_TYPE,bonk1, ierr )
call MPI_Get_elements(status,MPI_DOUBLE_PRECISION,bonk2,ierr)

67
INFRASTRUCTURE
write(*,*)"got ", bonk1," elements of type MY_TYPE"
write(*,*)"which contained ", bonk2," elements of type MPI_DOUBLE_PREC
ISION"
do i=0,size
if(rvect(i) /= -1)write(*,'(i2,f4.0)')i,rvect(i)
enddo
endif
call MPI_Finalize(ierr )
end program
! output
! numx= 5 extent= 1 stride= 5
! got 1 elements of type MY_TYPE
! which contained 5 elements of type MPI_DOUBLE_PRECISION
! 0 0.
! 5 5.
! 10 10.
! 15 15.
! 20 20.

68
INFRASTRUCTURE
Shows how to use MPI_Type_vector to send noncontiguous blocks of data
and MPI_Get_count and MPI_Get_elements to see the number of elements sent
*/
#include <stdio.h>
#include "mpi.h"
#include <math.h>
int main(argc,argv)
int argc;
char *argv[];
{
int myid, numprocs,mpi_err;
#define SIZE 25
double svect[SIZE],rvect[SIZE];
int i,bonk1,bonk2,numx,stride,extent;
MPI_Datatype MPI_LEFT_RITE;
MPI_Status status;

69
INFRASTRUCTURE
stride=5;
numx=(SIZE+1)/stride;
extent=1;
if(myid == 1){
printf("numx=%d extent=%d stride=%dn",stride,numx,extent,stride);
}
mpi_err=MPI_Type_vector(numx,extent,stride,MPI_DOUBLE,&MPI_LEFT_
RITE);
mpi_err=MPI_Type_commit(&MPI_LEFT_RITE);
if(myid == 0){
for (i=0;i<SIZE;i++)
svect[i]=i;
MPI_Send(svect,1,MPI_LEFT_RITE,1,100,MPI_COMM_WORLD);
}
if(myid == 1){
rvect[i]=-1;
MPI_Recv(rvect,1,MPI_LEFT_RITE,0,100,MPI_COMM_WORLD,&status);
}

70
INFRASTRUCTURE
if(myid == 1){
MPI_Get_count(&status,MPI_LEFT_RITE,&bonk1);
MPI_Get_elements(&status,MPI_DOUBLE,&bonk2);
printf("got %d elements of type MY_TYPEn",bonk1);
printf("which contained %d elements of type MPI_DOUBLEn",bonk2);
if(rvect[i] != -1)printf("%d %gn",i,rvect[i]);
}
MPI_Finalize();
}
/*
output
numx=5 extent=5 stride=1
got 1 elements of type MY_TYPE
which contained 5 elements of type MPI_DOUBLE
0 0
5 5
10 10
15 15
20 20
*/

71
INFRASTRUCTURE
MPI_Type_struct
• Defines a MPI datatype which maps to a
user defined derived datatype
• C
– int MPI_Type_struct(count, &array_of_blocklengths,
&array_of_displacement, &array_of_types, &newtype);
• Fortran
– Call MPI_TYPE_STRUCT(count, array_of_blocklengths,
array_of_displacement, array_of_types, newtype,ierror)

72
INFRASTRUCTURE
MPI_Type_struct
• Parameters:
– [IN count] # of old types in the new type (integer)
– [IN array_of_blocklengths] how many of each type in
new structure (integer)
– [IN array_of_types] types in new structure (integer)
– [IN array_of_displacement] offset in bytes for the
beginning of each group of types (integer)
– [OUT newtype] new datatype (handle)
– Call MPI_TYPE_STRUCT(count, array_of_blocklengths,
array_of_displacement,array_of_types, newtype,ierror)
– Ierr = MPI_Type_struct(count, &array_of_blocklengths,
&array_of_displacement, &array_of_types, &newtype);

73
INFRASTRUCTURE
Derived Data type Example
Consider the data type or structure consisting of
3 mpi double
10 mpi integer
2 mpi character
Creating the MPI data structure matching this C/Fortran
structure is a three step process
• Fill the descriptor arrays:
B - blocklengths
T - types
D - displacements
• Use MPI_Type_struct to create the MPI data structure
• Commit the new data type using MPI_Type_commit

74
INFRASTRUCTURE
Derived Data type Example
• To create the MPI data structure
matching this C/Fortran structure
– Fill the descriptor arrays:
• B - blocklengths
• T - types
• D - displacements
• Then use MPI_Type_struct

75
INFRASTRUCTURE
Derived Data type Example (continued)
Fortran :
! t contains the types that
! make up the structure
t(1)=MPI_DOUBLE_PRECISION
t(2)=MPI_INTEGER
t(3)=MPI_CHARACTER
! b contains the # of each type
b(1)=3;b(2)=10;b(3)=2
! d contains the byte offset of
! the start of each type
d(1)=0;d(2)=24;d(3)=64
call MPI_TYPE_STRUCT(3,b,d,t,
MPI_CHARLES,mpi_err)
MPI_CHARLES is our new data type
C :
/* t contains the types that
make up the structure*/
t[0]=MPI_DOUBLE
t[1]=MPI_INT
t[2]=MPI_CHAR
/*b contains the # of each type */
b[0]=3;b[1]=10;b[2]=2
/* d contains the byte offset of
the start of each type*/
d[0]=0;d[1]=24;d[2]=64
ierr = MPI_Type_struct(3,&b,&d,&t,
MPI_CHARLES,mpi_err)

76
INFRASTRUCTURE
MPI_Type_commit
• Before we use the new data type we call
MPI_Type_commit
• C
– MPI_Type_commit(&MPI_CHARLES)
• Fortran
– Call MPI_Type_commit(MPI_CHARLES,ierr)

77
INFRASTRUCTURE
Communicators
• A communicator is a parameter in all MPI
message passing routines
• A communicator is a collection of
processors that can engage in
communication
• MPI_COMM_WORLD is the default
communicator that consists of all processors
• MPI allows you to create subsets of
communicators

78
INFRASTRUCTURE
Why Communicators?
• Isolate communication to a small number of
processors
• Useful for creating libraries
• Different processors can work on different
parts of the problem
• Useful for communicating with "nearest
neighbors"

79
INFRASTRUCTURE
MPI_Comm_split
• Provides a short cut method to create a
collection of communicators
• All processors with the "same color" will be
in the same communicator
• Index gives rank in new communicator
• Fortran
– call MPI_COMM_SPLIT(OLD_COMM, color, index,
NEW_COMM, mpi_err)
• C
– MPI_Comm_split(OLD_COMM, color, index, &NEW_COMM)

80
INFRASTRUCTURE
MPI_Comm_split
• Split odd and even processors into 2 communicators
Program comm_split
include "mpif.h"
Integer color,zero_one
call MPI_COMM_RANK( MPI_COMM_WORLD, myid, mpi_err )
color=mod(myid,2) !color is either 1 or 0
call MPI_COMM_SPLIT(MPI_COMM_WORLD,color,myid,NEW_COMM,mpi_err)
call MPI_COMM_RANK( NEW_COMM, new_id, mpi_err )
call MPI_COMM_SIZE( NEW_COMM, new_nodes, mpi_err )
Zero_one = -1
If(new_id==0)Zero_one = color
Call MPI_Bcast(Zero_one,1,MPI_INTEGER,0, NEW_COMM,mpi_err)
If(zero_one==0)write(*,*)"part of even processor communicator"
If(zero_one==1)write(*,*)"part of odd processor communicator"
Write(*,*)"old_id=", myid, "new_id=", new_id
Call MPI_FINALIZE(mpi_error)
End program

81
INFRASTRUCTURE
MPI_Comm_split
• Split odd and even processors into 2 communicators
0: part of even processor communicator
0: old_id= 0 new_id= 0
2: part of even processor communicator
1: part of odd processor communicator
3: part of odd processor communicator

82
INFRASTRUCTURE
#include "mpi.h"
#include <math.h>
int main(argc,argv)
int argc;
char *argv[];
{
int myid, numprocs;
int color,Zero_one,new_id,new_nodes;
MPI_Comm NEW_COMM;
color=myid % 2;
MPI_Comm_split(MPI_COMM_WORLD,color,myid,&NEW_COMM);
MPI_Comm_rank( NEW_COMM, &new_id);
MPI_Comm_size( NEW_COMM, &new_nodes);
Zero_one = -1;
if(new_id==0)Zero_one = color;
MPI_Bcast(&Zero_one,1,MPI_INT,0, NEW_COMM);
if(Zero_one==0)printf("part of even processor communicator n");
if(Zero_one==1)printf("part of odd processor communicator n");
printf("old_id= %d new_id= %dn", myid, new_id);
MPI_Finalize();
}

Lecture11

Recommended

Recommended

More Related Content

Viewers also liked

Viewers also liked (20)

Similar to Lecture11

Similar to Lecture11 (20)

More from tt_aljobory

More from tt_aljobory (18)

Lecture11