Add Blog Post on Autotuning for Kokkos with APEX

[vlkale]

This PR adds a blog post on auto-tuning for Kokkos APEX. It includes discussion of its new features in Kokkos 4.5. It is a collaboration between @khuck and I, and the ideas and key points for the blog post have been discussed with @crtrott. I discussed with @dalg24 today.

It would be good to get this out and online by end of this week or otherwise in early January. I will be gone starting next week through Jan 9th, and in any case @khuck wanted to have it released very soon if possible and assuming this PR is approved.

I want to note that I followed the instructions to create the blog post through the README provided. When I got to the last step to test the addition, I couldn't see the blog post, i.e., I tried to build the website locally with npm run start, but I don't see the website locally. In terms of the text, everything has been reviewed by @khuck and the blog posts directions have been run by @crtrott.

[vlkale]

Here is what I see when I build the website locally on the command-line

and in my browser (using Google Chrome) at http://localhost:1313/

[khuck]

@vlkale it took me some poking around, but because the post date is after today's date, it won't render. If you change the date to 2024-12-10, it will show up. Also, we should try to find/make a higher resolution image with the two logos together...

[khuck]

@vlkale oh, and you forgot to upload APEX-tuning.jpeg

[vlkale]

@vlkale oh, and you forgot to upload APEX-tuning.jpeg

Thanks for that catch - actually, the problem is that file should be referenced in the text as kokkos-blog-post9.png. This is what I understand is their naming convention for all images for blog posts, having looked at the other blog posts' references to image files.

[vlkale]

@vlkale it took me some poking around, but because the post date is after today's date, it won't render. If you change the date to 2024-12-10, it will show up. Also, we should try to find/make a higher resolution image with the two logos together...

Wow - thanks for that! I also now see the role of each element in that header better after looking at the Markdown's documentation. I have fixed the date to today.

Here is the Google Slide source from which the synergized logo/thumbnail was created:
https://docs.google.com/presentation/d/1w1JDszvUE9z1UuMF8U2aQ2DTjPCBkbIZeya1LmdD-no/edit?usp=sharing

I just pasted the logos images I downloaded from the web and put them in the Google Slide. From that, I downloaded the slide as a PNG. I think the source itself needs a clearer image. I can look for a Kokkos logo image.

If others have suggestions for higher-resolution image for Kokkos logo, let me know.

[vlkale]

I have tested the website locally and it works. Here are two screenshots to show how it looks.

@khuck I have updated the Google Slides source for the APEX-Kokkos thumbnail to make it a higher resolution. I found and re-used the logo image of Kokkos in the assets/img directory that was already there.

[dalg24]

Please segregate the images and figures you are adding into their own subdirectory.

[dalg24]

What is "TuningInterface" referring to?

[dalg24]

Suggested change

Kokkos includes a Tuning API (TuningInterface) that can be used to construct a tuning context around a computational kernel, declare input variables that define the context state, declare output variables to be tuned, and request output variables when the kernel is executed. The Kokkos Tools infrastructure provide integrated support to utilize this API during Kokkos application execution, i.e., online, rather than offline [2]. Together, Kokkos Tools and the Tuning API is used in APEX to tune at runtime Kokkos kernel parameters running in any execution space / policy combination. We note through this Kokkos auto-tuning capability from APEX allows for (a) switching its tuning heuristics between Kokkos Execution Spaces (i.e. choose between serial or OpenMP depending on the problem size, etc.) or execution policies and (b) auto-tuning any arbitrary parameter within an application that uses Kokkos - solver choices, algorithmic parameters, tolerances, etc.

Kokkos includes a Tuning API (TuningInterface) that can be used to construct a tuning context around a computational kernel, declare input variables that define the context state, declare output variables to be tuned, and request output variables when the kernel is executed. The Kokkos Tools infrastructure provide integrated support to utilize this API during Kokkos application execution, i.e., online, rather than offline [2]. Together, Kokkos Tools and the Tuning API is used in APEX to tune at runtime Kokkos kernel parameters running in any execution space / policy combination. We note through this Kokkos auto-tuning capability from APEX allows for (a) switching its tuning heuristics between Kokkos Execution Spaces (i.e. choose between Serial or OpenMP depending on the problem size, etc.) or execution policies and (b) auto-tuning any arbitrary parameter within an application that uses Kokkos - solver choices, algorithmic parameters, tolerances, etc.

[dalg24]

I wonder if we can come up with a code snippet illustrating how one declare tunable variables and whether it would help here. Thoughts?

[vlkale]

I wonder if we can come up with a code snippet illustrating how one declare tunable variables and whether it would help here. Thoughts?

Yes, I think we could show the matrix multiplication code that is being experimented with. The code and setup with the Tuning API is a bit long (about 50 lines of code) and all of it is needed. @khuck and I discussed it but decided it's not needed to be shown here - maybe we can go into more detail in that Wiki Post we linked in the Outcomes Section.

[vlkale]

What is "TuningInterface" referring to?

This is what @khuck and I used to refer to the set of functions in Kokkos::Tools that helps declare tuning parameter variables and tuning contexts in which those variables should be tuned, e.g., the function Kokkos::Tools::declareTuningVariable() is part of that interface.

I think the text TuningInterface could actually be taken out and we can link to the Wiki Post for more information.

Also, maybe this can documented better in Kokkos or Kokkos Tools right now - this GitHub issue is probably the best documentation for its use: kokkos/kokkos-tools#90

[khuck]

I wonder if we can come up with a code snippet illustrating how one declare tunable variables and whether it would help here. Thoughts?

Well... that's one of the issues with the current API, it is kind of complicated to create a variable...

[dalg24]

Be more specific, this title will look confusing in 1y, even more so in 3.
Try something like "Kokkos 4.5 Release Introduces New Auto-Tuning Features" or whatnot.

[vlkale]

I used your suggestion for the more specific title. Perhaps we can say "...Introduces New Online Auto-Tuning Features" to be even more specific. As described in the blog, this Online Auto-Tuning relies on Kokkos Tools infrastructure. I wonder if that makes it too long.

@khuck Feel free to suggest a different title.

[dalg24]

Is the git-submodule part important? I would not mention it.
Also I don't quite get what you mean by "a stable version released in 4.5". There is currently still no Kokkos Tools releases in lock step with Kokkos and Kokkos Kernels releases.

[dalg24]

You mean for the URL to appear rather than making "Wiki post" an hyper reference?

[dalg24]

You added a whole lot of duplicated images with various quality and format. Was it intentional?

[dalg24]

Did you mean to commit changes to that file?

[vlkale]

No, looks like that was done by mistake. I have restored the original version of package-lock.json. The diff only has three files added/changed, i.e., the new markdown file for the text of the blog post, the thumbnail image, and the plot in the third section of the post.