Introduction to Programming with OpenMP
---------------------------------------

Practical Exercise 6 (Tasks)
----------------------------

Use similar commands to compile and run the program as in practical
exercise 2.  These exercises need minimal changes, but they are a bit
tricky - the hints describe what to watch out for.


Question 1
----------

1.1 The programs Programs/Carmichael.f90 and Programs/Carmichael.c
search for Carmichael numbers (see Wikipedia), using a very crude and
approximate approach.  Number theorists will blench at their crudeness,
so please ignore that.

Take one of them and parallelise it using tasks.  Divide the work of the
loop in the main program between tasks, creating 90% as many of them as
there are OpenMP threads.  In procedure Search, perform the weak and
strong tests in separate tasks, wait for them to complete.  Also remember
to synchronise the array update.


Question 2
----------

2.1 The programs Programs/Quicksort.f90 and Programs/Quicksort.c contain
a recursive quicksort for double precision floating-point.  The
algorithm is good, but there are better ways to code quicksort than by
using recursion - it is written that way to make this exercise easier to
code.  Also, this is not the best sort for SMP performance.

Take one of them and parallelise it using tasks.  You should also not
create a task if it has very little work to do - the specimen answer
used a criterion of 1000 elements, but that was picked out of a hat and
is not optimal.  In the last lecture, you will learn a better way to do
the last than by simply using if statements.

Perceptive people will also note that it potentially generates a huge
number of tasks - as commented, that will work if you use just the
techniques taught in this course.  Because of that, a very wide range of
values of the criterion will give similar performance.

For reasons mentioned in the course, Fortran users will need to use
taskwait where C ones will not (and will use a dirty trick, instead).
In both cases, this is because the OpenMP specification assumes a data
model that is incompatible with both languages, and this is exposed when
it comes to array arguments.