Parallel Computing: Lecture 16

Learning objectives

After this class, you should be able to:

1. Write MPI code using the following features: (i) MPI_Probe, (ii) MPI_Irecv, (iii) MPI_Isend, (iv) MPI_Test, and (v) MPI_Wait.
2. Given a type of data decomposition (such as row-wise block striped, column-wise block striped, etc), give an algorithm to perform matrix multiplication using that data decomposition, analyze its time complexity, and give its iso-efficiency and scalability functions.
3. Given matrices A and B, and the number of processors P, show the steps executed by Canon's algorithm to compute C = A*B.

Reading assignment

1. Chapter 11 (ignore the cache issues for now), also read the MPI standard for the MPI functions mentioned above (you may also refer to Chapter 9 for some of them).
2. None.

Exercises and review questions

• Questions on current lecture's material
  1. Give a parallel algorithm for matrix multiplication using a column-wise block-striped decomposition.
  2. Exercise 11.4 (b).
  3. (Programming) Write an MPI program that will run on two processors. Each process will exchange an array of 1000 integers with its neighbor using Isend/Irecv/Iwait. Before calling Iwait, please have each process sleep for 2 seconds. In a real program, you would not have the processes sleep. Instead, you would have them do some useful computation.

• Questions on next lecture's material
  1. None.