Improve cascade plots based on feedback at P3HPC21 #9
Conversation
This commit makes several changes to the cascade plots: - The "cliff" connecting the last platform to the x-axis is made optional (and disabled by default). This declutters the plot and makes it simpler to explain. - Each platform is assigned a single character identifier that is rendered on top of the legend and platform map. This improves accessibility and makes choosing good colors less critical. - Minor assorted formatting changes to improve readability of axis and platform titles. Co-authored-by: Jason Sewall <[email protected]>
|
An example of an improved plot: While reviewing these changes I remembered that there's a lot of magic in the Jupyter notebooks, and they don't diff very nicely. It would probably improve the user experience and make it easier for us to keep track of changes if we could move some of the new code (e.g. all the |
|
Looks much nicer, but wouldn't it be better to swap the placement of the two legends, so that the one for the lines is near the lines, and the one for the boxes is below the boxes? |
|
p.s. there was an article on using Git with Jupyter notebooks recently https://www.fast.ai/2022/08/25/jupyter-git/ which may be useful. |
|
Strangely, I ran across that very article independently yesterday. Thanks for getting this out there, @Pennycook. I think @JimCownie's suggestion about switching the axes of those two legends is a fine one. Some magic might be needed to make very big platform lists fit. |
|
FWIW, I also have some changes outside this function to ensure that if you are creating multiple plots of different (but intersecting sets of data) the encoding choices are common. So, say you have a number of plots comparing some set of compilers, each on a different machine, you want icx always to be orange, and LLVM always blue, etc. even if icx isn't available for all of the platforms (because they're Arm machines). |
| @@ -152,7 +199,15 @@ | |||
| } | |||
| ], | |||
| "source": [ | |||
| "fig = plt.figure(figsize=(12, 5))\n", | |||
| "fig = plt.figure(figsize=(16, 6.5))\n", | |||
| "app_order = [\"OpenMP\", \"Kokkos\", \"CUDA\", \"OpenACC\", \"OpenCL\"]\n", | |||
There was a problem hiding this comment.
Note to self: This is the approach to calling the script to ensure stable ordering of the applications so the application colours stay the same. The script is relatively unmodified, but invoked through the notebook with a list of applications and their orderings. By default, the script will order the applications with the most supported platforms first.
This commit makes several changes to the cascade plots:
The "cliff" connecting the last platform to the x-axis is made
optional (and disabled by default). This declutters the plot
and makes it simpler to explain.
Each platform is assigned a single character identifier that is
rendered on top of the legend and platform map. This improves
accessibility and makes choosing good colors less critical.
Minor assorted formatting changes to improve readability of axis and
platform titles.
Co-authored-by: Jason Sewall [email protected]