1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
|
Multithreaded calculation
-------------------------
Benchmarks for the :ref:`threaded` algorithm are shown here, for unmasked ``z``, a problem size ``n``
of 1000 and a ``total_chunk_count`` of 40 for up to 6 threads.
.. image:: ../_static/threaded_lines_simple_light.svg
:class: only-light
.. image:: ../_static/threaded_lines_simple_dark.svg
:class: only-dark
.. image:: ../_static/threaded_filled_simple_light.svg
:class: only-light
.. image:: ../_static/threaded_filled_simple_dark.svg
:class: only-dark
For the ``simple`` dataset contour calculations are faster with more threads but it does not scale
particularly well. The speedup with 6 threads is 2.4-2.5 for both :meth:`~.ContourGenerator.lines`
and :meth:`~.ContourGenerator.filled`. This problem dataset is perhaps not
computationally expensive enough to justify the use of multiple threads.
.. image:: ../_static/threaded_lines_random_light.svg
:class: only-light
.. image:: ../_static/threaded_lines_random_dark.svg
:class: only-dark
.. image:: ../_static/threaded_filled_random_light.svg
:class: only-light
.. image:: ../_static/threaded_filled_random_dark.svg
:class: only-dark
For the ``random`` dataset contour calculations scale much better with increasing number of threads
as long as one of the ``ChunkCombined...`` line or fill types is being used.
Using 6 threads the speedup is 4.2-4.5 for :meth:`~.ContourGenerator.lines` and 4.6-5.2 for
:meth:`~.ContourGenerator.filled`.
The :class:`~.LineType` and :class:`~.FillType` options that do not scale well are those that return individual
NumPy arrays for each line or polygon rather than combined arrays for each chunk. This is because
the allocation of a new NumPy array can only be performed by one thread at a time, so the larger the
number of arrays that are generated, the greater the likelihood that other threads are left waiting
before they can allocate arrays.
.. note::
Whether it is worth using :ref:`threaded` rather than :ref:`serial` for a particular problem
depends on the complexity of the dataset and what the calculated contours are to be used for.
If they are only needed for rendering using the `Matplotlib`_ Agg renderer, then for complicated
problems the rendering time usually far exceeds the calculation time so a reduction in
calculation time may not be of much real-world benefit.
|
