Class notes

Beware: because we are circling around the four tracks of shared memory parallelism, distributed memory parallelism, analysis, and applications, the notes below may not be in chronological order. Each section will have material added as it occurs.

• Architecture and Models
  • Basics of parallel computer architectures
  • Basic parallel timing models
  • Flynn's taxonomy
  • The SPMD model

• Threads and shared memory programming
  • Threaded programming for parallelism
  • Thread synchronization
  • Thread scheduling
  • Pthreads - POSIX standard threads
  • OpenMP compiler directive

• MPI Programming Basics
  • MPI overview and terms
  • Hello world in MPI
  • basic makefile usage
  • Collective communications in MPI
  • Non-blocking communications and modes
  • Scatter-gather operations
• Some Matlab utilities; on a Unix system put these into a directory ~/.matlab. On Windows, it depends on just what version of Windows you have. The most common is C:\\Documents\MATLAB, with your user name in the appropriate place if it is a shared system. You can also just create a folder and add that path to Matlab's search path: "userpath('MyCoolFolder')".
• discr_matlab.tar is a file with Matlab discretizer and solver. This replaces the older version by adding two more solver methods for the discretized PDE.
• bicgstab.tar has a complete PDE solver, in F2008
• The text file instructions has the questions to answer, and notes about the implementation, for the Bi-CGstab solver. If you choose to write a C version of the compressed sparse row matrix-vector product, also download the file c-version.tar which has the interface spec and a modified makefile and bicgstab.f90 file. Warning: I've tested a C version of the matrix-vector product for both COO and CSR data structures. And for the sizes I've tested (N = 2000 and N = 4000) the C version is 10-11% slower than the Fortran versions. This is a surprise to me, since the code is so simple and GCC and gfortran use the same backend optimizers and code generators. However, it does mean that if you do choose to implement CSR in C, you may also need to implement COO in C. Otherwise the comparison may not be meaningful.
• Matlab m-file looky.m has results from me running the code on mesh sizes N = 1000, 2000, ..., 8000. It creates four plots, which may be of interest. Read the m-file for details about just what the data is.

• Modified: Tue 06 Feb 2018, 06:18 PM
• Modified: Thu 08 Feb 2018, 07:20 AM adding basic MPI notes
• Modified: Tue 03 Apr 2018, 07:18 AM adding Matlab discretizer tar file
• Modified: Thu 26 Apr 2018, 01:54 PM to add final notes about bicgstab
• Last Modified: Fri 27 Apr 2018, 01:06 PM project