UW AMath High Performance Scientific Computing
 
AMath 483/583 Class Notes
 
Spring Quarter, 2013

Previous topic

Jacobi iteration using OpenMP with coarse-grain parallel block

Next topic

Bibliography and further reading

This Page

Jacobi iteration using MPIΒΆ

The code below implements Jacobi iteration for solving the linear system arising from the steady state heat equation using MPI. Note that in this code each process, or task, has only a portion of the arrays and must exchange boundary data using message passing.

Compare to:

The code:

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
! $UWHSPC/codes/mpi/jacobi1d_mpi.f90
!
! Domain decomposition version of Jacobi iteration illustrating
! coarse grain parallelism with MPI.
!
! The one-dimensional Poisson problem is solved, u''(x) = -f(x)
! with u(0) = alpha and u(1) = beta.
!
! The grid points are split up into ntasks disjoint sets and each task
! is assigned one set that it updates for all iterations.  The tasks 
! correspond to processes.
!
! The task (or process) number is called "me" in this code for brevity
! rather than proc_num.
!
! Note that each task allocates only as much storage as needed for its 
! portion of the arrays.
!
! Each iteration, boundary values at the edge of each grid must be
! exchanged with the neighbors.


program jacobi1d_mpi
    use mpi

    implicit none

    integer, parameter :: maxiter = 100000, nprint = 5000
    real (kind=8), parameter :: alpha = 20.d0, beta = 60.d0

    integer :: i, iter, istart, iend, points_per_task, itask, n
    integer :: ierr, ntasks, me, req1, req2
    integer, dimension(MPI_STATUS_SIZE) :: mpistatus
    real (kind = 8), dimension(:), allocatable :: f, u, uold
    real (kind = 8) :: x, dumax_task, dumax_global, dx, tol

    ! Initialize MPI; get total number of tasks and ID of this task
    call mpi_init(ierr)
    call mpi_comm_size(MPI_COMM_WORLD, ntasks, ierr)
    call mpi_comm_rank(MPI_COMM_WORLD, me, ierr)

    ! Ask the user for the number of points
    if (me == 0) then
        print *, "Input n ... "
        read *, n
    end if
    ! Broadcast to all tasks; everybody gets the value of n from task 0
    call mpi_bcast(n, 1, MPI_INTEGER, 0, MPI_COMM_WORLD, ierr)

    dx = 1.d0/real(n+1, kind=8)
    tol = 0.1d0*dx**2

    ! Determine how many points to handle with each task
    points_per_task = (n + ntasks - 1)/ntasks
    if (me == 0) then   ! Only one task should print to avoid clutter
        print *, "points_per_task = ", points_per_task
    end if

    ! Determine start and end index for this task's points
    istart = me * points_per_task + 1
    iend = min((me + 1)*points_per_task, n)

    ! Diagnostic: tell the user which points will be handled by which task
    print '("Task ",i2," will take i = ",i6," through i = ",i6)', &
        me, istart, iend


    ! Initialize:
    ! -----------

    ! This makes the indices run from istart-1 to iend+1
    ! This is more or less cosmetic, but makes things easier to think about
    allocate(f(istart-1:iend+1), u(istart-1:iend+1), uold(istart-1:iend+1))

    ! Each task sets its own, independent array
    do i = istart, iend
        ! Each task is a single thread with all its variables private
        ! so re-using the scalar variable x from one loop iteration to
        ! the next does not produce a race condition.
        x = dx*real(i, kind=8)
        f(i) = 100.d0*exp(x)               ! Source term
        u(i) = alpha + x*(beta - alpha)    ! Initial guess
    end do
    
    ! Set boundary conditions if this task is keeping track of a boundary
    ! point
    if (me == 0)        u(istart-1) = alpha
    if (me == ntasks-1) u(iend+1)   = beta


    ! Jacobi iteratation:
    ! -------------------

    do iter = 1, maxiter
        uold = u

        ! Send endpoint values to tasks handling neighboring sections
        ! of the array.  Note that non-blocking sends are used; note
        ! also that this sends from uold, so the buffer we're sending
        ! from won't be modified while it's being sent.
        !
        ! tag=1 is used for messages sent to the left
        ! tag=2 is used for messages sent to the right

        if (me > 0) then
            ! Send left endpoint value to process to the "left"
            call mpi_isend(uold(istart), 1, MPI_DOUBLE_PRECISION, me - 1, &
                1, MPI_COMM_WORLD, req1, ierr)
        end if
        if (me < ntasks-1) then
            ! Send right endpoint value to process on the "right"
            call mpi_isend(uold(iend), 1, MPI_DOUBLE_PRECISION, me + 1, &
                2, MPI_COMM_WORLD, req2, ierr)
        end if

        ! Accept incoming endpoint values from other tasks.  Note that
        ! these are blocking receives, because we can't run the next step
        ! of the Jacobi iteration until we've received all the
        ! incoming data.

        if (me < ntasks-1) then
            ! Receive right endpoint value
            call mpi_recv(uold(iend+1), 1, MPI_DOUBLE_PRECISION, me + 1, &
                1, MPI_COMM_WORLD, mpistatus, ierr)
        end if
        if (me > 0) then
            ! Receive left endpoint value
            call mpi_recv(uold(istart-1), 1, MPI_DOUBLE_PRECISION, me - 1, &
                2, MPI_COMM_WORLD, mpistatus, ierr)
        end if

        dumax_task = 0.d0   ! Max seen by this task

        ! Apply Jacobi iteration on this task's section of the array
        do i = istart, iend
            u(i) = 0.5d0*(uold(i-1) + uold(i+1) + dx**2*f(i))
            dumax_task = max(dumax_task, abs(u(i) - uold(i)))
        end do

        ! Take global maximum of dumax values
        call mpi_allreduce(dumax_task, dumax_global, 1, MPI_DOUBLE_PRECISION, &
            MPI_MAX, MPI_COMM_WORLD, ierr)
        ! Note that this MPI_ALLREDUCE call acts as an implicit barrier,
        ! since no process can return from it until all processes
        ! have called it.  Because of this, after this call we know
        ! that all the send and receive operations initiated at the
        ! top of the loop have finished -- all the MPI_RECV calls have
        ! finished in order for each process to get here, and if the
        ! MPI_RECV calls have finished, the corresponding MPI_ISEND
        ! calls have also finished.  Thus we can safely modify uold
        ! again.

        ! Also periodically report progress to the user
        if (me == 0) then
            if (mod(iter, nprint)==0) then
                print '("After ",i8," iterations, dumax = ",d16.6,/)', &
                    iter, dumax_global
            end if
        end if

        ! All tasks now have dumax_global, and can check for convergence
        if (dumax_global < tol) exit
    end do

    print '("Task number ",i2," finished after ",i9," iterations, dumax = ",&
            e16.6)', me, iter, dumax_global


    ! Output result:
    ! --------------

    ! Note: this only works if all processes share a file system
    ! and can all open and write to the same file!

    ! Synchronize to keep the next part from being non-deterministic
    call mpi_barrier(MPI_COMM_WORLD, ierr)

    ! Have each task output to a file in sequence, using messages to
    ! coordinate

    if (me == 0) then    ! Task 0 goes first
        ! Open file for writing, replacing any previous version:
        open(unit=20, file="heatsoln.txt", status="replace")
        write(20,*) "          x                  u"
        write(20, '(2e20.10)') 0.d0, u(0)    ! Boundary value at left end

        do i = istart, iend
            write(20, '(2e20.10)') i*dx, u(i)
        end do

        close(unit=20)
        ! Closing the file should guarantee that all the ouput 
        ! will be written to disk.
        ! If the file isn't closed before the next process starts writing,
        ! output may be jumbled or missing.

        ! Send go-ahead message to next task
        ! Only the fact that the message was sent is important, not its contents
        ! so we send the special address MPI_BOTTOM and length 0.
        ! tag=4 is used for this message.

        if (ntasks > 1) then
            call mpi_send(MPI_BOTTOM, 0, MPI_INTEGER, 1, 4, &
                          MPI_COMM_WORLD, ierr)
            endif

    else
        ! Wait for go-ahead message from previous task
        call mpi_recv(MPI_BOTTOM, 0, MPI_INTEGER, me - 1, 4, &
                          MPI_COMM_WORLD, mpistatus, ierr)
        ! Open file for appending; do not destroy previous contents
        open(unit=20, file="heatsoln.txt", status="old", access="append")
        do i = istart, iend
            write(20, '(2e20.10)') i*dx, u(i)
        end do

        ! Boundary value at right end:
        if (me == ntasks - 1) write(20, '(2e20.10)') 1.d0, u(iend+1)    

        ! Flush all pending writes to disk
        close(unit=20)

        if (me < ntasks - 1) then
            ! Send go-ahead message to next task
            call mpi_send(MPI_BOTTOM, 0, MPI_INTEGER, me + 1, 4, &
                          MPI_COMM_WORLD, ierr)
        end if
    end if

    ! Notify the user when all tasks have finished writing
    if (me == ntasks - 1) print *, "Solution is in heatsoln.txt"

    ! Close out MPI
    call mpi_finalize(ierr)

end program jacobi1d_mpi