Department of Computer Science	University of San Francisco

You should speak to me or send me email regarding your proposed project by Monday, April 14.

Projects can be team efforts. However, no one project can have more than two students working on it.

Duplicate projects won't be allowed, and projects will be assigned on a first-come, first-served basis. The first student (or team of students) to propose a project will be assigned the project. So let me know about your project as soon as possible.

Here's a list of ideas for possible projects. You don't have to choose a project from this list, but you must get your project approved.

1. Implement Gaussian elimination with submatrix partitioning. Compare pipelined and non-pipelined implementations. Compare block, cyclic, and block-cyclic distributions. Predict performance and compare actual and predicted performance. Compare your solver with the ScaLAPACK solver.
2. Implement a distributed memory preconditioned GMRES solver for sparse linear systems. Predict performance and compare actual and predicted performance. Compare your solver with the PETSc GMRES solver.
3. Implement serial and distributed-memory parallel versions of Strassen's algorithm for matrix multiplication. Compare the performance of your parallel implementation to parallel matrix multiplication available in ScaLAPACK.
4. Implement WaTor using MPI and dynamic load balancing. Discuss performance.
5. Implement a distributed memory parallel branch-and-bound solution of the travelling salesman problem using dynamic load balancing. Discuss performance.
6. Write a distributed memory parallel program for repartitioning a distributed graph so that the weight of the edge-set that's cut is minimized. Include code for redistributing the graph. Compare the performance of distributed sparse matrix-vector multiplication before and after the redistribution. Include the cost of the redistribution.
7. Parallel sorting. Implement a variety of distributed memory parallel sorting algorithms. Discuss their relative performance.
8. Write an MPI program that uses asynchronous iteration with conjugate gradients to solve sparse systems of equations. Compare the performance of your solver to a ``conventional'' CG solver.
9. Explore latency and bandwidth of shared memory communication between threads when the threads are assigned to different cores of the same processor and when the threads are assigned to cores on different processors. How does the performance of AMD systems (grolsch, penguin) compare to the performance of Intel systems (spaten, stella)? How does the addition of communicating threads affect performance?
10. Install one or more of the software transactional memory systems on grolsch and spaten or stella. Implement one or more of the dwarves with the transactional memory software, Pthreads, and OpenMP. How does the performance of the systems compare?
11. Implement a significant parallel algorithm on the cluster using OpenMP for intranode communication and MPI for internode communication. Also implement the algorithm using only MPI. How does the performance of the mixed implementation compare with the MPI-only implementation? How did the difficulty of implementation of the two systems compare?