Compile and execute the program in the file compute_pi_mpi.c, which computes an estimate of using the parallel algorithm discussed in class. The program is available on the shared Google Drive for this class. It should be compiled and executed on either ada.tamu.edu or terra.tamu.edu.
Load the Intel software stack prior to compiling and executing the code.
module load intel/2017A
To compile, use the command:
mpiicc -o compute_pi_mpi.exe compute_pi_mpi.c
To execute the program, use
mpirun –np <p> ./compute_pi_mpi.exe <n>
where <n> represents the number of intervals and <p> represents the number of processes.
The output of a sample run is shown below.
mpirun -np 4 compute_pi_mpi.exe 100000000
n = 100000000, p = 4, pi = 3.1415926535897749, relative error = 5.80e-15, time (sec) = 0.0608
The run time of the code should be measured when it is executed in dedicated mode. Use the batch file compute_pi_mpi.job, to execute the code in dedicated mode using the following command on ADA:
bsub < compute_pi_mpi.job
On Terra, you will need to use compute_pi.terra_job, and the corresponding command is:
sbatch compute_pi.terra_job
Execute the code for n=108 with p chosen to be 2k, for k = 0, 1, …, 6. Specify ptile=4 in the job file. Using the experimental data obtained from these experiments, answer the following questions.
1. (10 points) Plot execution time versus p to demonstrate how time varies with the number of processes. Use a logarithmic scale for the x-axis.
2. (10 points) Plot speedup versus p to demonstrate the change in speedup with p.
3. (5 points) Using the definition: efficiency = speedup/p, plot efficiency versus p to demonstrate how efficiency changes as the number of processes is increased.
4. (5 points) What value of p minimizes the parallel runtime?
5. (10 points) With n=109 and p=64, determine the value of ptile that minimizes the total_time. Plot time versus ptile to illustrate your experimental results for this question.
6. (10 points) Repeat the experiments with p=64 for n=102, 104, 106 and 108.
a. Plot the speedup observed w.r.t. p=1 versus n.
b. Plot the relative error versus n to illustrate the accuracy of the algorithm as a function of n.
Submission: Upload a single PDF or MSWord document with your answers to ecampus.
