LANS Informal Seminar: Junchao Zhang and Barry Smith

Seminar Title: IPM: A Post-MPI Programming Model (Part II)
Speaker: Junchao Zhang and Barry Smith, Computational Mathematicians, Argonne National Laboratory

Date/Time: 2016-04-20 15:00
Location: Bldg 240 rm 4301

Description:
Part I: The MPI parallel programming model has been a very successful parallel programming model for over twenty years. Though many alternative parallel programming models have been proposed over the years, both before and after the golden age of MPI, none has gained much traction because none have been better than MPI across the board. IPM is a programming model that attempts to be at least as good as MPI at everything and much better than MPI at certain things. One aspect that MPI is not good at is working with dynamic data structures that are often resizing and thus require the user to manage repartitioning and movement of the data structures across MPI processes. We will introduce the IPM programming model, discuss how it may be used to write an adaptive mesh refinement/unrefinement PDE simulation code, and compare its performance with the libMesh package. Humor will be used to make the presentation engaging.

Part II: Managing (irregular) ghost regions usually is the most difficult part of distributed parallel programming. However, it is a crucial part since it links to data structures of the program and data dependance of the computation; it tells secretes about parallelism and communication in the code. Unfortunately, many programming models for distributed memory, including HPF, MPI, PGAS etc, have done a bad job to help users on that. In this presentation, I will talk how I implemented a finite element AMR code with no explicit ghost regions using IPM.