Lecture 17

Learning objectives

After this class, you should be able to:

1. Optimize parallel code using double buffering.
2. Optimize large loops to reduce instruction cache misses.
3. Use domain decomposition techniques to distribute data assigned to different processes or threads in a parallel computation.
4. Use dynamic load balancing techniques to assign roughly equal work to each process or thread in a parallel computation.

Reading assignment

1. Read the paper on graph partitioning under the 'Course Library' tab on Blackboard.

Exercises and review questions

• Exercises and review questions on current lecture's material

  1. Consider an 8x8 array of cells, which are numbered in row major order. Give the order in which cells are visited by the Hilbert space-filling curve.
  2. Suggest alternatives to the edge-cut metric as a measure of graph partition quality. Post your answer on the discussion board.
  3. Show the Laplacian of a 4x4 mesh.
  4. Show how the Spectral method would bisect the graph of slide 39.

• Preparation for the next lecture
  1. None.

---

Last modified: 17 Mar 2010