CS 6504 
Fall 2015 
Accessible Parallel Computing

Wu Feng

This course will address making parallel computing in the context
of applications (or more immediately, the Berkeley dwarfs or motifs)
more accessible or more programmable. Overall, we will seek to address
programmability, portability, and performance in the context of parallel
computing and as an offshoot of our "Parallel Computing for the Masses"
grant from NSF.

I would anticipate the first 2-4 weeks of the course looking at studying
a range of programming models --- from the simple and visual, i.e., our
parallel Scratch/Snap! prototype to library-based, e.g., cuBLAS, cuSPARSE,
etc., to pragma-based, e.g., OpenMP or OpenACC, to primitive-based, i.e.,
CUDA or OpenCL. Underpinning the programming models is our OpenDwarfs
suite on github. While everyone would learn about the programming models
and the OpenDwarfs, I expect each student to identify an application
area that they have expertise for or want to have expertise in, e.g.,
a physicist taking the class may gravitate to the n-body dwarf and apply
it to his/her seismology project.

The paper reading portion of the course would occur about 1/4 to 1/3
of the way through the course, where we would look at the different
application areas and read and critique papers in parallel computing
in support of those areas, particularly at the abstraction level of a
dwarf, e.g., much of my own work with dynamic programming, structured
and unstructured grid, and n-body dwarfs, Berkeley's work with structured
grid dwarf, and so on. We would be critiquing the work and simultaneously
students will be leveraging this critiqued work in advanced parallel
computing to apply to their own advanced topics project.

Loosely, a syllabus would look as follows:

Weeks 1-4 Parallel programming models: parallel Scratch/Snap!,
library-based programming, MapReduce, OpenMP, OpenACC, CUDA/OpenCL in
the context of a n-body dwarf and/or a dense linear algebra dwarf from
our OpenDwarfs project

Weeks 5-15 Paper reading (including the Landscape of Parallel Architecture
and dwarfs), discussion, and occasionally directed hands-on application
of readings to code (likely towards the end of the semester).