Computer Science 335
Parallel Processing and High Performance Computing

Fall 2021, Siena College

Home     Syllabus     Schedule     Examples

Agenda

• Announcements
  • Submission/grading updates
  • Lab 4: Point to Point Communication and Lab 5: Non-Blocking Messages are now past due
  • Lab 6: MPI on Stampede2, complete by early next week (no coding, just getting MPI programs running on Stampede2)
• Lab 7: Collective Communication out, likely the task mostly for Monday's class, due in a week
• Programming Project 3: Parallelizing Jacobi Iteration out
• Example: MPI implementation of Conway's Game of Life
  • Uses a distributed data structure: Each process maintains its own subset of the computational domain, in this case just some of the rows of the grid. Other processes do not know about the data on a given process. Only the data that is needed to compute the next generation, a one-cell overlap (sometimes called ghost cells or a halo), is exchanged between neighbors before each iteration.
  • When we need to get a global count of some statistic, such as the count of live cells at the start, we use a reduction.
  • The communication of cell data between iterations is done with two pairs of sends and receives. Here, we use nonblocking calls, then wait for their completion with the MPI_Waitall call.
  • This example also shows a technique for writing a file when the intended contents are distributed among a set of processes.

Terminology

• collective communication
• reduction
• ghost cells/halo

Examples

• mpilife: MPI parallelization of the Conway's Game of Life simulator