!~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ! Math and Computer Science Division, Argonne National Laboratory ! !----------------------------------------------------------------------- ! CVS twocmp.seqUnvn.F90,v 1.6 2007-12-19 17:13:17 rloy Exp ! CVS MCT_2_8_0 !BOP ------------------------------------------------------------------- ! ! !ROUTINE: twocomponentUneven.sequential ! ! !DESCRIPTION: Provide a simple example of using MCT to connect two components ! In this case the models are running sequentialy but the second model ! is only running on 1 processor. ! ! !INTERFACE: ! program twosequn ! ! !USES: ! !--- Get only the things needed from MCT use m_MCTWorld,only: MCTWorld_init => init use m_GlobalSegMap,only: GlobalSegMap use m_GlobalSegMap,only: MCT_GSMap_init => init use m_GlobalSegMap,only: MCT_GSMap_lsize => lsize use m_AttrVect,only : AttrVect use m_AttrVect,only : MCT_AtrVt_init => init use m_AttrVect,only : MCT_AtrVt_zero => zero use m_AttrVect,only : MCT_AtrVt_lsize => lsize use m_AttrVect,only : MCT_AtrVt_indexRA => indexRA use m_AttrVect,only : MCT_AtrVt_importRA => importRAttr use m_Rearranger,only: Rearranger use m_Rearranger,only: MCT_Rearranger_init => init use m_Rearranger,only: MCT_Rearrange => Rearrange implicit none include 'mpif.h' integer,parameter :: ngx = 6 ! points in x-direction integer,parameter :: ngy = 4 ! points in y-direction integer ier,world_group,model2_group,myrank2,myrank3 integer,dimension(:),pointer :: myids,mycomms,peloc2 integer,dimension(:,:),pointer :: GlobalId integer :: comm1,comm2,asize,mysize,i,myproc integer :: commsize integer,dimension(1) :: start1,length1,ranks integer,dimension(:),allocatable :: start2,length2 !----------------------------------------------------------------------- ! The Main program. ! We are implementing a single-executable, sequential-execution system. ! Because its sequential, communication occurs through the main using ! arguments. The second component is only running on 1 processor type(GlobalSegMap) :: GSmap1,GSmap2 type(AttrVect) :: av1,av2 type(Rearranger) :: Rearr call MPI_init(ier) call mpi_comm_size(MPI_COMM_WORLD, mysize,ier) if(mysize .gt. 12) then write(6,*)"Must run on less than 12 processors" stop endif call mpi_comm_rank(MPI_COMM_WORLD, myproc,ier) ! the first model is running on all the processors so give ! it a dubplicate of MPI_COMM_WORLD for its communicator call mpi_comm_dup(MPI_COMM_WORLD,comm1,ier) ! the second model is only running on one processor ! so use mpi_groups methods to define its communicator call mpi_comm_group(MPI_COMM_WORLD,world_group,ier) ! need a communicator that only has the first processor ranks(1)=0 ! define the group call mpi_group_incl(world_group,1,ranks,model2_group,ier) ! now define the communicator ! first initialize it comm2=MPI_COMM_NULL call mpi_comm_create(MPI_COMM_WORLD,model2_group,comm2,ier) ! don't need the groups anymore call mpi_group_free(world_group,ier) call mpi_group_free(model2_group,ier) ! allocate arrays for the ids and comms allocate(myids(2),mycomms(2)) ! Set the arrays to their values. myids(1)=1 myids(2)=2 mycomms(1)=comm1 mycomms(2)=comm2 ! now call the initm_ version of MCTWorld_init call MCTWorld_init(2,MPI_COMM_WORLD,mycomms,myids) ! first gsmap is the grid decomposed in one dimension ! there is 1 segment per processor length1(1)= (ngx * ngy)/mysize start1(1)= myproc * length1(1) + 1 write(6,*)'gsmap1', myproc,length1(1),start1(1) call MCT_GSMap_init(GSMap1,start1,length1,0,comm1,1) ! second gsmap is the grid on one processor ! for GSMap init to work, the size of the start and length arrays ! must equal the number of local segments. So I must allocate ! size zero arrays on the other processors. if(myproc .eq. 0) then allocate(start2(1),length2(1)) length2(1) = ngx*ngy start2(1) = 1 else allocate(start2(0),length2(0)) endif call MCT_GSMap_init(GSMap2,start2,length2,0,comm1,2) write(6,*)'gsmap2', myproc,GSMap2%ngseg,GSmap2%gsize,GSmap2%start(1), & GSmap2%pe_loc(1),GSmap2%length(1) ! initialize an Av on each GSMap call MCT_AtrVt_init(av1,rList="field1:field2",lsize=MCT_GSMap_lsize(GSMap1,comm1)) ! Use comm1 because lsize of GSMap2 on comm1 will return 0 on non-root processors. ! We need av2 to be full-sized on proc 0 and 0 size on other processors. call MCT_AtrVt_init(av2,rList="field1:field2",lsize=MCT_GSMap_lsize(GSMap2,comm1)) ! create a rearranger. Use the communicator which contains all processors ! involved in the rearrangement, comm1 call MCT_Rearranger_init(GSMap1,GSMap2,comm1,Rearr) !-------------end of initialization steps ! Start up model1 which fills av1 with data. call model1(comm1,av1) ! print out Av data do i=1,MCT_AtrVt_lsize(av1) write(6,*) "model 1 data", myproc,i,av1%rAttr(1,i),av1%rAttr(2,i) enddo ! rearrange data from model1 so that model2 can use it. call MCT_Rearrange(av1,av2,Rearr) ! pass data to model2 (which will print it out) ! model2 should only run on one processor. if(myproc .eq. 0) then call model2(comm2,av2) endif ! all done call MPI_Barrier(MPI_COMM_WORLD,ier) if (myproc==0) write(6,*) 'All Done' call mpi_finalize(ier) contains !----------------------------------------------------------------------- !----------------------------------------------------------------------- ! !ROUTINE: subroutine model1(comm1,mod1av) ! the first model implicit none integer :: comm1,mysize,ier,asize,myproc integer :: fieldindx,avsize,i integer,dimension(1) :: start,length real,pointer :: testarray(:) type(GlobalSegMap) :: GSmap type(AttrVect) :: mod1av !--------------------------- ! find local rank and size call mpi_comm_size(comm1,mysize,ier) call mpi_comm_rank(comm1,myproc,ier) write(6,*)"model1 myproc,mysize",myproc,mysize avsize = MCT_AtrVt_lsize(mod1av) write(6,*)"model 1 myproc, av size", myproc,avsize ! Fill Av with some data ! fill first attribute the direct way fieldindx = MCT_AtrVt_indexRA(mod1av,"field1") do i=1,avsize mod1av%rAttr(fieldindx,i) = float(i+ 20*myproc) enddo ! fill second attribute using Av import function allocate(testarray(avsize)) do i=1,avsize testarray(i)= cos((float(i+ 20*myproc)/24.) * 3.14) enddo call MCT_AtrVt_importRA(mod1av,"field2",testarray) end subroutine model1 !----------------------------------------------------------------------- !----------------------------------------------------------------------- ! !ROUTINE: subroutine model2(comm2,mod2av) implicit none integer :: comm2,mysize,ier,asize,myproc integer :: i type(AttrVect) :: mod2av !--------------------------- ! find local rank and size call mpi_comm_size(comm2,mysize,ier) call mpi_comm_rank(comm2,myproc,ier) write(6,*)"model2 myproc,mysize",myproc,mysize asize = MCT_AtrVt_lsize(mod2av) write(6,*)"model 2 myproc, av size", myproc,asize ! print out Av data do i=1,asize write(6,*) "model 2 data after", myproc,i,mod2av%rAttr(1,i),mod2av%rAttr(2,i) enddo end subroutine model2 end