|
High Performance Scientific Computing
AMath 483/583 Class Notes Spring Quarter, 2013 |
|
|
High Performance Scientific Computing
AMath 483/583 Class Notes Spring Quarter, 2013 |
Due Friday, May 31, 2013, by 11:00pm PDT. Note change in due date
The goals of this homework are to:
First make sure MPI works by trying out the sample programs in the directory $UWHPSC/codes/mpi.
Create a new subdirectory homework6 in your repository for the codes you create.
The sample program $UWHPSC/codes/mpi/copyvalue.f90 sets a value of i in Process 0 and passes this through a chain of processes up to Process num_procs - 1. There are print statements in this codes so that running it should give, for example with 4 processes specified:
$ mpif90 copyvalue.f90
$ mpiexec -n 4 a.out
Process 0 setting i = 55
Process 1 receives i = 55
Process 1 sends i = 55
Process 2 receives i = 55
Process 2 sends i = 55
Process 3 ends up with i = 55
Create a modified code copyvalue2.f90 that instead sets the value i = 55 in Process num_procs - 1 and passes the value backwards to Process 0.
Also change your version to subtract 1 from the value of i by each Process along the way. So the output should now look like:
$ mpif90 copyvalue2.f90
$ mpiexec -n 4 a.out
Process 3 setting i = 55
Process 2 receives i = 55
Process 2 sends i = 54
Process 1 receives i = 54
Process 1 sends i = 53
Process 0 ends up with i = 53
The directory $UWHPSC/codes/homework6 contains an MPI version of the quadrature codes, which simply uses the Trapezoid Rule to calculate an approximation to a single integral. Copy these files into your repository and then fix the file test.f90 so that it properly prints out the total number of f evaluations at the end. The sample version simply prints 0 and to do the right thing you will have to use MPI_REDUCE to sum up the values of feval_proc that each process computes.
There is a Makefile that can be used to run with different values of num_procs, and when working properly you code should give results similar to this:
$ make test
mpif90 -c functions.f90
mpif90 -c quadrature.f90
mpif90 -c test.f90
mpif90 functions.o quadrature.o test.o -o test.exe
mpiexec -n 4 test.exe
Using 4 processes
true integral: 6.00136745954910E+00
Process 0 with n = 1000 computes int_approx = 6.00088098873994E+00
Process 3 with n = 1000 computes int_approx = 6.00088098873994E+00
Process 1 with n = 1000 computes int_approx = 6.00088098873994E+00
Process 2 with n = 1000 computes int_approx = 6.00088098873994E+00
fevals by Process 3: 1000
fevals by Process 0: 1000
fevals by Process 1: 1000
fevals by Process 2: 1000
Total number of fevals: 4000
The default number of processes is 4, but you can change this, e.g.:
$ make test NUM_PROCS=6
mpiexec -n 6 test.exe
Using 6 processes
true integral: 6.00136745954910E+00
Process 0 with n = 1000 computes int_approx = 6.00088098873994E+00
Process 4 with n = 1000 computes int_approx = 6.00088098873994E+00
Process 5 with n = 1000 computes int_approx = 6.00088098873994E+00
fevals by Process 5: 1000
Process 1 with n = 1000 computes int_approx = 6.00088098873994E+00
fevals by Process 1: 1000
Process 2 with n = 1000 computes int_approx = 6.00088098873994E+00
fevals by Process 2: 1000
Process 3 with n = 1000 computes int_approx = 6.00088098873994E+00
fevals by Process 3: 1000
fevals by Process 0: 1000
fevals by Process 4: 1000
Total number of fevals: 6000
Note: In spite of the MPI_BARRIER calls in between blocks of print statements, the statements may appear in somewhat the wrong order as above because of the way print buffers work. Having multiple processes print to the screen (or to the same file) in an MPI code is generally a bad idea (and doesn’t even make sense if you’re running on a cluster of computers that don’t share the same file system or have access to the terminal). But it’s useful for our purposes in better understanding what’s being done.
The way test.f90 is written it is not using multiple processors in a sensible way. Each process is independently computing exactly the same thing – an approximation to the integral using the same number of points n across the full interval from a to b.
Create a new version test2.f90 that instead does the following:
When run with num_procs processors, split the interval from a to b up into num_procs - 1 pieces. Process 0 should act as the master process that splits up the interval and sends the subintervals out to each worker process 1, 2, ..., num_procs - 1. Process 0 should then receive a value int_sub from each worker process that is the Trapezoidal approximation the the integral over this subinterval computed with n points. Process 0 adds these up and prints out the resulting approximation to the integral over the full interval a to b.
Use this master-worker paradigm and a similar framework to the sample code $UWHPSC/codes/mpi/matrix1norm1.f90. Note that Process 0 does not compute an approximation on any subinterval, it just acts as the master.
For this problem it would be both easier and perhaps more sensible to instead have each process compute an approximation on a subinterval and then use MPI_REDUCE to combine these into the final full approximation. But the point of this exercise is to get some experience with the master-worker paradigm.
The master process should compute the left and right edge of the j’th subinterval and send these to Process j. These should be sent with a single call to MPI_SEND, e.g. in the code for the master process you might have lines like:
if (proc_num == 0) then
dx_sub = (b-a) / nsub
do j=1,nsub
ab_sub(1) = a + (j-1)*dx_sub
ab_sub(2) = a + j*dx_sub
call MPI_SEND(ab_sub, 2, MPI_DOUBLE_PRECISION, j, j, &
MPI_COMM_WORLD, ierr)
enddo
! followed by a loop to receive back each result.
Here ab_sub is an array of length 2 that holds the left and right edges of the j’th subinterval.
The code for each worker process should receive this information and then have a call of the form:
int_sub = trapezoid(f,ab_sub(1),ab_sub(2),n)
to compute the Trapezoid approximation on this subinterval with n points, and then send the result back to Process 0.
Sample results might look like the following:
$ make test2
mpif90 -c functions.f90
mpif90 -c quadrature.f90
mpif90 -c test2.f90
mpif90 functions.o quadrature.o test2.o -o test2.exe
mpiexec -n 4 test2.exe
Using 4 processes
true integral: 6.00136745954910E+00
fevals by Process 0: 0
fevals by Process 1: 1000
fevals by Process 2: 1000
fevals by Process 3: 1000
Trapezoid approximation with 3000 total points: 6.00131677608477E+00
Total number of fevals: 3000
Note that Process 0 does no function evaluations and each of the others use n = 1000 points for their subinterval.
With this version of the code, increasing the number of processes should improve the approximation to the integral.
The final two lines of the output shown above were printed with these statements:
if (proc_num==0) then
print '("Trapezoid approximation with ",i8," total points: ",es22.14)',&
nsub*n, int_approx
print '("Total number of fevals: ",i10)', fevals_total
endif
(Note that the format statement can be directly included in the print statement by enclosing it in quotes of a different sort than used in specifying the format.)
There is no additional 583 problem on this assignment.
Your homework6 directory should contain:
As usual, commit your results, push to bitbucket, and see the Canvas course page for the link to submit the SHA-1 hash code. These should be submitted by the due date/time to receive full credit.