CSCI-UA.0480-051: Parallel Computing

CSCI-UA.0480-051: Parallel Computing

Final Exam (May 15th, 2023)

Total: 100 points

Problem 1

Suppose we have the following two DAGs. Each DAG represents a process. That is, DAG 1 is a process and DAG 2 is another process. The two DAGs are totally independent from each other. The table shows the time taken by each task in each process.

a. [8 points] What will be the minimum time taken by each process if we execute it alone on one core, two cores, three cores, and four cores? That is, execute DAG 1 on one core, then on two cores, then on three cores, and then on four cores. Do the same for DAG 2. (Hint: You can put the result in the form of a table with three columns: #cores, time for DAG 1, and time for DAG 2. your table will have four rows for one, two, three, and four cores). Assume each core does not have hyperthreading technology.

b. [10 points] Based on the results you calculated in part a above, which DAG benefited more from using more cores? The two DAGs look similar, yet one DAG benefited from more cores. What is the main characteristic that this DAG has and makes it benefit from more cores?

c. [10 points] Suppose DAG 1 is your media player and DAG 2 is your browser. You start these two processes exactly at the same time and you are using a machine with eight cores (with no hyperthreading technology). How long does each DAG take to finish? Justify your answer.

Problem 2

Suppose we have the following MPI+OpenMP piece of code. The whole program is not shown. The lines with dots (i.e. lines 2, 7,  10, and  12) contain some other code not needed for this problem. Some of the variables are used but not declared (e.g. num, procID, n, and arrays A and B). Assume they have been declared in a code not shown here. Assume this program runs on eight cores processor. Each core is a four- way hyperthreading. Each core has its own private level 1 cache. Each level 1 cache is 32KB. Each core also has its own level 2 cache of size 1MB each. There is a shared level 3 cache of size 8MB. The program is executed using the following command: (progname is the name of the program’s executable).

mpiexec -n 4 ./progname

a. [20 points] Fill in the right column of the following table with a short answer to the questions on the left. Please do not write any justification unless the question asks for it explicitly.

b. [8 points] Is there a situation where threads, in the above code, accessing the arrays A[] and B[] can cause the coherence protocol to start? If yes, what is the situation? If not, why not?

c. [8 points] Given the code above, how many virtual address spaces did the OS create? Justify your answer in 1-2 lines only.

d.  [8  points]  Is there a possibility that a process reaches line 27 before the other processes? If yes, will this cause a problem, and what is that problem? If not, why not?

e. [8 points] If one of the processes created for the above code crashed for some reason, do we risk having a deadlock? Justify.

Problem 3

For each question below, choose all correct answers. That is, a question may have one or more correct answers.

a. [4 points] Suppose a process wants to send data to a subset of processes . That process has the following options:

1. Split the communicator to smaller ones and use collective communication.

2. Make a series of send and receive to each one of the destination processes.

3. Use broadcast call.

4. Split each process into multiple threads and let threads communicate through shared memory.

b. [4 points] A warp in an NVIDIA GPU:

1. is transparent to the programmer.

2. consists of a maximum of 32 threads.

3. has each four threads share the same fetch and decode hardware.

4. suffer from thread divergence, also called branch divergence, possibility.

c.  [4 points] A block in CUDA can be split among two SMs.

1. This statement is always true.

2. This statement is true if the block has more threads than the number of SPs in the SM.

3. This statement is always false.

4. This statement is false only if the number of threads in the block is less than the number of SPs in the SM.

d. [4 points] The following characteristics are needed for a code to be GPU friendly.

1. computation intensive

2. independent computations

3. similar computations

4. large problem size

e. [4 points] Choose all the correct statements from the following one:

1. If one program has a higher speedup than another program, for the same number of cores, it means that the program with the higher speedup also has higher efficiency than the other one.

2. MPI can run on distributed memory machines and shared memory machines.

3. OpenMP can run on distributed memory machines and shared memory machines.

4. Power consumption/dissipation is the main reason we moved from single core to multicore.