집합통신 II

Chapter 8

Scatter

1. MPI_Scatter

scatter

• root rank의 isend 버퍼 값을 각 rank로 전달
• MPI_Gather()의 반대 연산

MPI_Scatter()

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

2. Example: MPI_Scatter()

C code

#include <stdio.h>
#include <mpi.h>
int main()
{
	int i, myrank;
	int isend[3]={0,}, irecv=0;
	MPI_Init(NULL,NULL);
	MPI_Comm_rank(MPI_COMM_WORLD,&myrank);
    
	if(myrank==0) for(i=0;i<3;i++) isend[i]=i+1;
	MPI_Scatter(isend,1,MPI_INT,&irecv,1,MPI_INT,0,MPI_COMM_WORLD);
	printf("myrank: %d irecv=%d\n",myrank,irecv);
    
	MPI_Finalize();
	return 0;
}

Result

3. MPI_Scatterv

scatterv

• 각 rank 별로 다른 크기의 data를 전달
• MPI_Gather()의 반대 연산
• 전체 input 값을 root에서 한번에 읽은 후 각 rank에 필요한 부분을 배분시킬 때 많이 사용

MPI_Scatterv()

int MPI_Scatterv( const void* sendbuf, const int sendcounts[], const int displs[], 
MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, 
int root, MPI_Comm comm )

Reduce

1. MPI_Reduce

reduce

• 각 프로세스의 input buffer의 요소들을 op 연산을 수행한 후 root 랭크의 output buffer에 결과를 반환
• Input buffer와 output buffer의 type과 요소의 수(배열의 크기)는 일치해야 함

MPI_Reduce()

int MPI_Reduce( const void* sendbuf, void* recvbuf, int count,
MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm )

• &sendbuf 다음에 바로 &recvbuf가 위치함

Operation & Data type

Operation	Data type(C)
MPI_SUM(sum), MPI_PROD(product), MPI_MAX(maximum), MPI_MIN(minimum)	MPI_INT, MPI_LONG, MPI_SHORT, MPI_UNSIGNED_SHORT, MPI_UNSIGNED, MPI_UNSIGNED_LONG, MPI_FLOAT, MPI_DOUBLE, MPI_LONG_DOUBLE
MPI_MAXLOC(max value and location), MPI_MINLOC(min value and location)	MPI_FLOAT_INT, MPI_DOUBLE_INT, MPI_LONG_INT, MPI_2INT, MPI_SHORT_INT, MPI_LONG_DOUBLE_INT
MPI_LAND(logical AND), MPI_LOR(logical OR), MPI_LXOR(logical XOR)	MPI_INT, MPI_LONG, MPI_SHORT, MPI_UNSIGNED_SHORT, MPI_UNSIGNED, MPI_UNSIGNED_LONG
MPI_BAND(bitwise AND), MPI_BOR(bitwise OR), MPI_BXOR(bitwise XOR)	MPI_INT, MPI_LONG, MPI_SHORT, MPI_UNSIGNED_SHORT, MPI_UNSIGNED, MPI_UNSIGNED_LONG, MPI_BYTE

• MPI_2INT = { MPI_INT, MPI_INT }

2. Example

C code

#include <stdio.h>
#include <mpi.h>
int main()
{
	int nrank,ista,iend,i;
	double A[9]={0.0,}, sum=0.0, tsum=0.0;
	MPI_Init(NULL,NULL);
	MPI_Comm_rank(MPI_COMM_WORLD,&nrank);
    
	ista=nrank*3; iend=ista+2;
	for(i=ista;i<iend+1;i++) A[i]=i+1;
	sum=0.0;
	for(i=ista;i<iend+1;i++) sum += A[i];
	MPI_Reduce(&sum,&tsum,1,MPI_DOUBLE,MPI_SUM, 0, MPI_COMM_WORLD);
	if(nrank==0) printf("rank(%d):sum=%.2f\n",nrank,tsum);
    
	MPI_Finalize();
	return 0;
}

• rank 0: sum = 1+2+3 = 6
• rank 1: sum = 4+5+6 = 15
• rank 2: sum = 7+8+9 = 24

Result

3. MPI_Allreduce

allreduce

• 각 프로세스의 input buffer의 요소들을 op 연산을 수행한 후 전체 프로세스에 그 결과값을 전달

MPI_Allreduce()

int MPI_Allreduce(const void* sendbuf, void* recvbuf, int count, 
MPI_Datatype datatype, MPI_Op op, MPI_Comm comm )

• MPI_Reduce() 연산에서 root argument가 빠진 형태

AlltoAll

1. MPI_Alltoall

alltoall

• 각 rank 순서대로 scatter 연산을 수행하는 형태
• 각 rank의 i번째 data들을 rank i로 전송

MPI_Alltoall()

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

2. Example: MPI_Alltoall

C code

#include <stdio.h>
#include <mpi.h>
int main()
{
	int isend[3],irecv[3];
	int i,nprocs,myrank;
	MPI_Init(NULL,NULL);
	MPI_Comm_size(MPI_COMM_WORLD,&nprocs);
	MPI_Comm_rank(MPI_COMM_WORLD,&myrank);
    
	for(i=0;i<nprocs;i++) isend[i]=1+i+nprocs*myrank;
	printf("myrank:%d isend=%d %d %d\n",myrank,isend[0],isend[1],isend[2]);
	MPI_Alltoall(isend,1,MPI_INT, irecv,1,MPI_INT,MPI_COMM_WORLD);
	printf("myrank:%d irecv=%d %d %d\n",myrank,irecv[0],irecv[1],irecv[2]);
    
	MPI_Finalize();
	return 0;
}

Result

3. MPI_Alltoallv

alltoallv

• MPI_Alltoall()에 유연성을 추가
• 송신 데이터의 위치는 sdispls에 지정
• 수신 측의 수신 데이터의 위치는 rdispls에 추가됨

MPI_Alltoallv()

int MPI_Alltoallv(const void* sendbuf, const int sendcounts[], const int sdispls[], 
MPI_Datatype sendtype, void* recvbuf,const int recvcounts[], const int rdispls[], 
MPI_Datatype recvtype, MPI_Comm comm )

4. Example: MPI_Alltoallv

C code

#include <stdio.h>
#include <mpi.h>
int main()
{
	// Hard Coding
	int isend[6]={1,2,2,3,3,3}, irecv[9]={0,};
	int iscnt[3]={1,2,3}, isdsp[3]={0,1,3}, ircnt[3], irdsp[3];
	int myrank,nprocs,i;
    
	MPI_Init(NULL,NULL);
	MPI_Comm_size(MPI_COMM_WORLD,&nprocs);
	MPI_Comm_rank(MPI_COMM_WORLD,&myrank);
    
	for(i=0;i<6;i++) isend[i]=nprocs*myrank+isend[i];
	for(i=0;i<3;i++){
		if(myrank==0) ircnt[i]=1,irdsp[i]=i;
		else if(myrank==1) ircnt[i]=2, irdsp[i]=2*i;
		else if(myrank==2) ircnt[i]=3, irdsp[i]=3*i;
		}
	MPI_Alltoallv(isend,iscnt,isdsp,MPI_INT,irecv,ircnt,irdsp,MPI_INT,MPI_COMM_WORLD);
	printf("myrank %d irecv=",myrank);
	for(i=0;i<9;i++) printf(" %d",irecv[i]);
	printf("\n");
    
	MPI_Finalize();
	return 0;
}

Result

Barrier, Wtime

Barrier

int MPI_Barrier(MPI_Comm comm)

int MPI_Ibarrier(MPI_Comm comm, MPI_Request *request)

• communicator내의 모든 프로세스가 호출한 지점에 도달할 때까지 진행을 멈춤

Wtime

double MPI_Wtime(void)

• Wall clock time을 return

Example

…
Double s_time, e_time, t_time;
MPI_Barrier(MPI_COMM_WORLD);
s_time=MPI_Wtime(); // 여기까지 걸린 시간
// 계산
MPI_Barrier(MPI_COMM_WORLD);
e_time=MPI_Wtime(); // 여기까지 걸린 시간
t_time = e_time – s_time;

• t_time = e_time – s_time은 실제 계산하는 데에 걸린 시간