Introduction to MPI
This page gathers together resources for the Introduction to MPI course by MJ Rutter for the proposed 2022 MPhil / CDT in the Centre for Scientific Computing. The copyright of the contents is held by its author, unless otherwise stated.
- Course Handout as slides.
Are available only to those registered with the course organisers. For those, a link to the lectures.
The files below use the convention that .c is C source code,
and .cc is C++ source code. The C should be comprehensible to
anyone who knows C++, so not all C examples have been translated. If
this course also teaches C++ programmers about
then that is a bonus.
The Beginning (lecture 2)
- hello.c (hello.cc)
- hello.f90 and hello.f08
- version.c (or version.f90)
- int1.c (trivial integration) (or int1.cc or int1.f90)
Save for the Hello examples, all .f90 examples use the Fortran 2008
MPI interface of
use mpi_f08. Your Fortran compiler
might, or might not, recognise f08 as a valid extension for Fortran
source. Gfortran does, ifort does not...
More Quadrature (lecture 3)
- gsl.c GSL quadrature, serial
- gsl_mpi.c GSL quadrature, MPI
- square.c pi/4, serial (or square.f90)
- square_mpi.c pi/4, MPI (or square_mpi.f90)
- int1_gather.c (or int1_gather.cc) trivial integration plus MPI_Gather
Calling the C library GSL from Fortran is sufficiently messy that those examples have been tidied away to a separate page as an example of using Fortran's iso_c_binding and GSL.
Mandelbrot Set (lecture 4)
(Note that C (C99) and C++ differ in their treatment of complex
numbers. The C examples below will not work with C++ compilers. Note
too that Preview on MacOS will read
.pam files. One way of
achieving this is to type
$ open mandel.pam -a /Applications/Preview.app/
- mandel.c Mandelbrot, serial (or mandel.cc)
- mandel_mpi.c Mandelbrot, MPI (or mandel_mpi.cc)
- mandel_mpi_io.c Mandelbrot, MPI, I/O on rank zero
- mandel_mpi_io_pp.c Mandelbrot, MPI, I/O on rank zero, point to point
- mandel_mpi_ms.c Mandelbrot, MPI, master-slave (or mandel_mpi_ms.cc)
Some prefer to see examples in Fortran, and, for them, mandel_mpi.f90 is provided. This uses mpi_gather to collect the whole array onto the root process before writing it out. It uses stream I/O (introduced in Fortran 2003). It could be modified to use the inquire statement and pos option to the write statement in order to let each MPI process write its own line, in a similar fashion to the C codes above.
As an extra example, combining MPI with GTK (with no proof that they are compatible), mgtk.c. This is a master-slave Mandelbrot set which displays its output via GTK. It is intended as an MPI example, not an example of good MPI+GTK programming. One can readily see idle slaves consuming CPU time in this example.
(One might well argue that modern computers are so fast that there is little reason to use the power of MPI for this problem, and even serial python codes such as this TK python Mandelbrot set are quite fast enough. On the other hand, the MPI code presented here was used to produce a 31,500x10,500 pixel image of part of the Mandelbrot set with maxiter set to 2048. This was then printed at 300dpi to produce a poster for a school. An attempt to produce an image of this size using Mathematica's MandelbrotSetPlot function failed, whereas the MPI code succeeded on a much more modest computer. One can waste much time admiring images of different parts of the Mandelbrot set. Regrettably many images on the web fail to include co-ordinates, so are hard to reproduce. But sometimes one can find a gallery of Mandelbrot images with co-ordinates.)
Laplace's Equation (lecture 5)
- laplace.c Laplace, serial
- laplace.f90 ditto, Fortran
- laplace_mpi.c Laplace, MPI (or laplace_mpi.cc)
- laplace_mpi_bsend.c Laplace, MPI, Bsend
- laplace_mpi_isend.c Laplace, MPI, Isend
- laplace_mpi_analytic.f90 Laplace, MPI, Isend, periodic boundary conditions and comparison with analytic result
as above, but using the older
use mpiinterface, showing the difference with requests and statuses
Miscellaneous (lecture 6)
Communicators (lecture 7)
- comms.f90 and comms.cc: synthetic examples of reductions on matrix (and these examples are not translations of each other, so one may wish to read both)
MPI-IO (lecture 7)
- mandel_mpi-io.c MPI_File_write_at
- mandel_mpi-io.f90 ditto, Fortran
- mandel_mpi-io_datatype.f90 (user defined datatype and file view to interleave rows)
- mandel_mpi-io_datatype_v2.f90 (user defined datatype and file view to interleave pixels)
Hardware (lecture 8)
- stream_mpi.f Stream benchmark, Fortran, MPI, directly copied from Dr McAlpin's Stream website. You will also need a timer function such as mysecond.f90
- timer_test.f90 a possible solution to testing timer resolution