消息传递并行编程环境MPI.doc

Size: px

Start display at page:

Download "消息传递并行编程环境MPI.doc"

役卓
5 years ago
Views:

1 973 MPI PETS MPI MPI MPI MPI 2 MPI PETS PETS 1

3 MPI MPI MPI 1 MPI MPI MPI 2 - u=f MPI 3

4 1 proess CPU 4

5 send reeive barrier redution CPU soket, 4 : API FORTRAN C C++ 5

6 PVM MPICH 1.2 LAMMPI MPI Message Passing Interfae MPI MPI MPICH LAMMPI MPICH 6

7 6 SPMD MPMD SPMD 7 7

8 CPU CPU CPU CPU CPU CPU M M M M M M 8

9 MPI MPI MPI MPI MPI MPICH LAMMPI MPI 1.2 MPI 2.0 MPI MPP UNIX Windows NT ; 1000 MPI MPI PVM 9

10 10

11 MPI 1 MPI 2 MPI SPMD MPI rank P 0 1 P-1 MPI P

12 program example1 inlude mpif.h!! MPI integer status(mpi_status_size),my_rank,p,soure,dest,tag,ierr,data MPI all MPI_Init(ierr) all MPI_Comm_rank(MPI_COMM_WORLD,my_rank,ierr) all MPI_Comm_size(MPI_COMM_WORLD,p,ierr) data=0 tag = 5 soure= my_rank-1 if(soure.eq.-1) soure=p-1 dest =my_rank+1 if(dest.eq.p) dest=0 if(my_rank.eq.0) then else all MPI_Send(data,1,MPI_INTEGER,dest,tag,MPI_COMM_WORLD,ierr) all MPI_Rev(data,1,MPI_INTEGER,soure,tag,MPI_COMM_WORLD,status,ierr) all MPI_Rev(data,1,MPI_INTEGER,soure,tag,MPI_COMM_WORLD,status,ierr) data=data+1 all MPI_Send(data,1,MPI_INTEGER,dest,tag,MPI_COMM_WORLD,ierr) endif all MPI_Bast(data,1,MPI_INTEGER,0,MPI_COMM_WORLD,ierr) if(my_rank.eq.0) then endif if(data.eq.p-1) then print *, Suessful, data=,data else print *, Failure, data=,data endif all MPI_Finalize(ierr) end 12

13 : mpif77 -o exam.e example.f : mpirun np 4 exam.e MPI 4 exam.e MPI Init() MPI Comm_rank() my_rank=0 myrank=1 myrank=2 my_rank=3 Comm_size() p=4 p=4 p=4 p=4 tag=5,data=0 tag=5,data=0 tag=5,data=0 tag=5,data=0 soure=3 soure=0 soure=1 soure=2 dest=1 dest=2 dest=3 dest=0 send() rev() rev() rev() rev() data=data+1 send() data=data+1 send() data=data+1 send() Broadast() send() rev() rev() rev() output data Finalize() MPI 13

14 3 MPI rank MPI_Comm_rank() ; (Communiator) 1) 2) MPI 3) MPI MPI_COMM_WORLD MPI_Init() 4) MPI_Comm_size() ( MPI ; data envelope FORTRAN C INTEGER REAL DOUBLE PRECISION COMPLEX LOGICAL CHARACTER MPI MPI_INTEGER MPI_REAL MPI_DOUBLE_PRECISION MPI_COMPLEX MPI_LOGICAL MPI_CHARACTER ; 14

15 MPI : MPI FORTRAN MPI INTEGER MPI handle MPI MPI MPI MPI_COMM_WORLD FORTRAN MPI INTEGER mpirun np xxx xxx ; 15

16 read(*,*) 0 0 write * * 4 MPI FORTAN 0 C C ierr=mpi_comm_rank(myrank) F : MPI_Comm_rank(myrank,ierr) 5. MPI : man MPI_Xxxx,, man MPI_Comm_rank, man MPI_Send, man MPI_rev. 6. MPI MPI man 16

17 7. MPI Call MPI_Init ( ) MPI, MPI_COMM_WORLD Call MPI_Comm_rank ( ) Call MPI_Comm_size ( ) 1 2 MPI Call MPI_Finalize ( ) End 17

18 MPI 1 point-to-point MPI

19 INTEGER A INTEGER A A=100 A=100 MPI_Send(A,1,.) MPI_Isend(A,1, ) A=200 A=200 : A=100 A=100 A=200 MPI_Isend(A,1, flag, ) MPI_Wait(flag, ) A=200 : A=100 A=100 19

20 3 MPI_Send buf ount datatype dest tag omm ierr Real*8 integer buf : (Fortran, ) integer ount integer datatype integer dest integer tag integer omm integer ierr buf : : : : real *8 a(100,100) integer b(60,60) a(5,20) : a(54,20) 2 all MPI_Send( a (5,20),50,MPI_DOUBLE_PRECISION,2, & 99999,MPI_COMM_WORLD,ierr ) b(20,40) : b(39,40) 5 all MPI_Send( b (20,40),20,MPI_DOUBLE_PRECISION,5, & 99999,MPI_COMM_WORLD,ierr ) 20

21 MPI_Rev buf ount datatype dest tag omm status, ierr Real*8 integer buf : (Fortran, ) integer ount integer datatype integer dest integer tag integer omm buf : : : integer status(mpi_status_size) : ; integer ierr : real *8 a(100,100) integer b(60,60) a(5,20) : a(54,20) all MPI_Rev( a (5,20),50,MPI_DOUBLE_PRECISION,2, & 99999,MPI_COMM_WORLD,status,ierr ) b(20,40) : b(39,40) all MPI_Rev( b (20,40),20,MPI_DOUBLE_PRECISION,5, & 99999,MPI_COMM_WORLD,status,ierr ) 21

22 4 Colletive Communiation MPI MPI_Barrier ( omm, ierr ) : : MPI_Bast(buf,ount,dtype,root,omm,ierr) root other MPI_Gather(bufs,bufr,ount,dtype,root,omm,ierr) all root MPI_Allgather() all all :MPI_Gatherv() all root 22

23 :MPI_Allgatherv() all All :MPI_Satter(bufs,bufr,ount,dtype,root,omm,ierr) root All : MPI_Satterv() root All : MPI_Alltoall() All All : MPI_Alltoallv() (global redution) : : MPI_Redue(sbuf,rbuf,ount,dtype,op,root, omm,ierr); op : MPI_SUM, MPI_MIN, MPI_MAX, MPI_PROD 12 ; 23

24 :, 0 subroutine par_blas1(m,a,b,,omm) real a(m),b(m)! loal slie of array real! result real sum integer m,omm,i,ierr!loal sum sum=0.0d0 do i=1,m sum=sum+a(i)*b(i) enddo! global sum all MPI_Redue(sum,,1,MPI_REAL,MPI_SUM,0, & omm,ierr) : MPI_Allredue(),, MPI_Redue() ; : MPI_Redue_satter(), MPI_San() 24

25 P i 100 A 100 i +1 mod P i -1 mod P 100 B 100 C =A B C

26 MPI 1 MPI MPI_INTEGER,MPI_REAL,MPI_DOUBLE_PRECI SION,MPI_CHARACTER real a(1000) : a(5:9) all MPI_Send(a(5), 5, MPI_ REAL,..) OK : a(5),a(7),a(9),a(11),a(13),a(15) do i=5, 15, 2 all MPI_Send(a(i),1,MPI_REAL,.) OK enddo : all MPI_Type_vetor(5,1,2,MPI_REAL,newtype,ierr) all MPI_Type_ommit(newtype, ierr) all MPI_Send(a(5), 1, newtype,.) all MPI_Type_free(newtype,ierr) 26

27 MPI_Type_Commit() MPI_Type_free() 2 3*3 9 Cartesion i,j (i-1,j), (i+1,j), (i,j-1), (i,j+1) ,0 3 1,0 1 0,1 4 1,1 2 0,2 5 1,2 0 2,0 1 2,1 2 2,2 3 I/O / MPICH

28 MPI 2 1 Dirihlet - u (x) = f(x) x Ω =(0,1) u(0)=0.0, u(1)=0.0, f(x)=4.0 :,Jaobi npro=4, n = 21, ns = (n-1)/npro+1 = program example impliit real*8 (a-h,o-z) inlude mpif.h parameter ( n = 21, npro = 4, ns = (n-1)/npro+1) parameter ( errtol=1.e-4,nitmax=100) dimension u(ns), f(ns), solution(ns),uold(0:ns)! u(0) : left dummy grid point! u(ns+1) : right dummy grid point integer status(mpi_status_size),size enter into MPI all MPI_Init(ierr) all MPI_Comm_rank(MPI_COMM_WORLD,myrank,ierr) all MPI_Comm_size(MPI_COMM_WORLD,size,ierr) if(size.ne.npro) then print *, +++ errors for number of proess distributions +++ goto 888 endif assign initial values h = 1.0d0/(n-1) xst = myrank*(ns-1)*h nloal=ns-1! disrete step size! start x-oordinate of loal domain! loal number of grid point 28

29 ist=1 if(myrank.eq.0) ist=ist+1 do i=1,ns f(i)=4.0d0 u(i)=0.0d0 uold(i)=0.0d0 xauxi=xst+(i-1)*h solution(i)= -2*xauxi*(xauxi-1.0d0) enddo u(0)=0.0d0 nlp=myrank-1! left proess nrp=myrank+1! right proess if(nlp.lt.0) nlp=mpi_proc_null if(nrp.gt.npro-1) nrp=mpi_proc_null! null proess, null ommuniation operation Jaobi iterating nit =0! urrent iterations 10 ontinue swapping dummy elements along pseudo-boundary all MPI_Send(uold(1),1,MPI_DOUBLE_PRECISION, & nlp,nit, MPI_COMM_WORLD,ierr) all MPI_Send(uold(nloal),1, MPI_DOUBLE_PRECISION, & nrp,nit,mpi_comm_world,ierr) all MPI_Rev( uold(nloal+1),1,mpi_double_precision, & nrp,nit, MPI_COMM_WORLD,status,ierr) all MPI_Rev( uold(0),1, MPI_DOUBLE_PRECISION, & nlp,nit, MPI_COMM_WORLD,status, ierr) iterating and onvergen heking error=0.0d0 do i=ist,nloal u(i)=(h*h*f(i)+uold(i-1)+uold(i+1))/2 xauxi= dabs(u(i)-solution(i)) if(dabs(xauxi).gt.error) error=xauxi enddo maximum error all MPI_Allredue(error,xauxi,1,MPI_DOUBLE_PRECISION,MPI_MAX, & MPI_COMM_WORLD,ierr) error=xauxi if(error.gt.errtol) then do i=ist,nloal 29

30 uold(i)=u(i) enddo if(nit.lt.nitmax) then nit=nit+1 if(myrank.eq.0) print *,"nit=",nit," error=",error goto 10 endif endif if(myrank.eq.0) then if(nit.le.nitmax.and.error.le.errtol) then write(*,100) nit 100 format(1x, Suessfully touh the exat solution after nit =, & i4, iterations ) else write(*,200) nit 200 format(1x, Fail to touh the exat solution after nit =, & endif endif i4, iterations ) 888 all MPI_Finalize(ierr) end 1. LINUX : mpif77 o exam.e example.f 2 mpirun -np 4 exam.e 30

31 MPI 1 V 1.0 V1.2 V1.2 I/O V OpenMP MPI OpenMP SMP + SMP MPI SMP OpenMP MPI OpenMP MPI OpenMP ; 31

32 3 MPI OS 32

33 NERSC ODE Argonal PETS Parallel Extensible Tooltiks for Sientifi Computing, LLNL NS ASCI Red % 25% SC 99 Gondar Bell, MPI ; 33

34 1 C M L =A M N *B N L A C B B C Dirihlet - u = f Jaobi 34

第7章-并行计算.ppt

第7章-并行计算.ppt EFEP90 10CDMP3 CD t 0 t 0 To pull a bigger wagon, it is easier to add more oxen than to grow a gigantic ox 10t 0 t 0 n p Ts Tp if E(n, p) < 1 p, then T (n) < T (n, p) s p S(n,p) = p : f(x)=sin(cos(x))