Parallel Computing (CMSC416/CMSC616)

Assignment 5: MPI+OpenMP (only for 616 students)

Due: November 21, 2024 @ 11:59 PM Eastern Time

For this assignment, you will implement a hybrid MPI+OpenMP version of Assignment 2. There are three parts/steps to this assignment.

Part I: Implementing a 2D decomposition in MPI

First you will implement a 2D decomposition in MPI using non-blocking routines. You will add two arguments to the command line at the end that enable providing the X and Y dimensions of the MPI virtual grid. As an example, if we are running with 64 MPI processes, we should be able to use a 8x8 or 16x4 or 4x16 virtual grid of processes. The modified command line will like this: 


          mpirun -np <# of processes> ./life <data-file-name> <# of generations> <X_limit> <Y_limit> <# of processes in X> <# of processes in Y>
You can assume that X_limit and Y_limit will be powers of 2 as will be the number of processes you will be running on. You can also assume that you will be running the program on a minimum of 16 processes and X_limit and Y_limit are much larger than the number of processes in each dimension.

Part II: Using OpenMP in the compute loop

Next you will create a hybrid MPI+OpenMP version of the program implemented in Part I by adding support for OpenMP in the sequential compute region of the program. This is the code region that is executed sequentially by each MPI process in the Part I implementation.

Part III: Evaluating MPI processes vs. OpenMP threads

Finally, once you have a correctly working implementation from Part II, you will study the impact of using a varying number of processes vs. threads on performance. You will use 2 nodes of zaratan for these studies. On one extreme, you can use 256 MPI processes (and 1 thread/node) on 2 nodes, and on the other extreme, you can create 16 MPI processes, 8 on each node, and create 16 OpenMP threads per MPI process. And you can do anything in between.

What to Submit

You must submit the following files and no other files:

  • life-nonblocking-2d.[c,cpp,f77,f90]: parallel version using non-blocking Isend/Irecv routines, where the file extension depends on the language used for the implementation
  • life-nonblocking-hybrid.[c,cpp,f77,f90]: parallel version using non-blocking Isend/Irecv routines and OpenMP, where the file extension depends on the language used for the implementation
  • Makefile that will compile your programs successfully on zaratan when using mpicc or mpicxx. Make sure that the executable names are life-nonblocking-2d and life-nonblocking-hybrid, and do not include the executable in the tarball.
  • You must also submit a report (pdf) with performance results comparing processes vs. threads and your reasoning for the observations. The line plot should present the execution times to run the parallel version on the input file life.1024x1024.data (for 16, 32, 64, 128, and 256 processes and the corresponding no. of OpenMP threads), running on a 1024x1024 board for 500 iterations. In the report, you should mention:
    • how was the initial data distribution done
    • any considerations for implementing the hybrid version
    • what are the performance results, and are they what you expected
You should put the code, Makefile and report in a single directory (named LastName-FirstName-assign5), compress it to .tar.gz (LastName-FirstName-assign5.tar.gz) and upload that to gradescope.

Tips

  • Zaratan primer
  • Use the compiler flag -g while debugging but -O2 when collecting performance numbers.
  • MPI_Wtime example
  • Use the --exclusive flag with sbatch when collecting execution times.

Grading

The project will be graded as follows:

Component Percentage
2D MPI only version runs correctly 40
Hybrid version runs correctly 30
Performance evaluation and writeup 30
NOTE: If your program does not compile when submitted on gradescope, you get 0 points. If your program does not run correctly, you do NOT get any points.

Web Accessibility