Name:____________________________

SS# ______________________________

CS 320

Introduction to Parallel Programming

Final Exam

Due: Tuesday May 12,1998, at 9:00 am

Bring exam to 3317 DCL

NOTE: Because this is a take-home exam, we cannot answer any questions. So, if there are any ambiguities, please state your assumptions.

1. [5 pts.] What is the effect of Moore's Law on the popularization of parallel computing?
2. [5pts.] Briefly discuss how supercomputers contribute to lower costs and enhance the quality of designs in the aerospace industry.
3. [5pts.] (a) Define locality in the context of multicomputers. Why is locality important?
4. (b) Give two examples of SIMD machines.
5. [5pts.] (a) Mention one advantage of the array notation over the parallel loop notation. (b) Mention a disadvantage.
6. [5pts] What is a Fortran program unit?
7. [5 pts.] Under which circumstances is the following where statement incorrect? Assume that subscripts are always within the bounds of the array.

where (a(1:n+1) > b(m:m+n)) x(1:100)=y(1:100)+1

1. [10 pts.]Consider the code segment

...

t=a+b

y=c*d

t=t+y

do i=1,n

x(i)=x(i)+t

end do

print *,x

end

.

1. The following translation to parallel form is incorrect:

...

c$omp parallel private(y)

t=a+b

y=c*d

t=t+y

c$omp pdo

do i=1,n

x(i)=x(i)+t

end do

c$omp end pdo nowait

c$omp end parallel

print *,x

end

1. Explain why the code is incorrect. Then, give two different correct versions of the previous translation following the following rules: (a) do not remove any of the OpenMp directives, (b) do not change or delete any of the statements, and (c) do not change the location of any of the statements or OpenMp directives.
2. [10 pts.] Consider the code segment

c$omp parallel

c$omp pdo

do i=1,n

a(i)=b(i)+c(i)

do while (a(i) > x(i))

a(i) = a(i) - x(i)

end do

end do

c$omp end pdo nowait

c$omp pdo

do i=1,n

c(i) = d(i) + e(i)

end do

c$omp end pdo nowait

c$omp end parallel

(a) Explain why each of the nowait directives could improve performance.

(b) How would you change the second loop so that the first nowait causes the parallel program to contain races?

1. [5 pts.] Consider the code segment:

c$omp parallel do

do i=1,n

x(k(i))=x(k(i)+1)+B(i)

end do

c$omp end parallel

(a) Are races possible in this loop? Explain (b) Is the loop a valid translation of its sequential counterpart? Explain.

1. [5 pts.] Transform the following loop to parallel form:

do i=1,n

k=2*k

x=b(i)+c(i)

a(k)=x+sin(x)

end do

1. [10 pts.] Consider the loop:

do i=3,n

do j=2,n

a(i,j)=a(i+2,j-1)+a(i-1,j)+a(i,j-1)

end do

end do

(a) Sketch the dependence graph of the unrolled two dimensional iteration space. Clearly mark anti dependences in the graph.

(b)Transform the loop into a forall vector operation with the maximum parallelism possible.

1. [10 pts.]Assume a vector A is distributed across N processors, one element per processor. Using send/receive as the only communication primitives, write an MPI program that computes in processor i the value:

1. [10pts.] Transform the following loop into vector parallel form:

do i=1,n

if (x(i).le.x(i-1)) x(i)=x(i-1)

end do

Is the resulting algorithm consistent? Explain.

1. [5pts.] Assuming that N is very large, and using only loop interchanging (if necessary), transform the following Fortran loop into the best possible parallel version. Justify your answer.

do i=1,n

do j=1,n

a(i,j)=a(i,j)+1

end do

end do

1. [5pts.] Give a brief intuitive reason why asynchronous iterative algorithms may converge faster than their synchronous equivalents.