PLOT: v1+ (and if possible for older versions too, mainly 0.4+) avg-mteps

[John Owens]

avg-mteps
Not interested in any parameters used here, just plain ol' mteps and runtime for every dataset for every primitive. Tooltips are appreciated.
Note for pagerank, the time would be divided by max-iter.

[John Owens]

Same thing as #46, but MTEPS. Please tell me how to fix or tell me to push to the docs @neoblizz.

Note the JSON has no MTEPS for pr or tc; we have no data there.

SSSP:
BC:

[Muhammad Osama]

Can you use the following for pagerank mteps? I will fix it in code:

m_teps = (double)num-edges / (elapsed * 1000.0);

[Muhammad Osama]

Also, a few things:

• Does this also include Chuck's changes? So, best performance INCLUDING v1.1.x results.
• Is there a way to show this as a really long bar-chart of some sort? Or have horizontal-spacing between the GPU markers so we can see what the difference is easily?

[John Owens]

This is

roots = [
    "../gunrock-output/v1-0-0/sssp",
    "../gunrock-output/v1-0-0/bc",
    "../gunrock-output/v1-0-0/tc",
    "../gunrock-output/v1-0-0/pr",
    "../gunrock-output/v1-0-0/bfs",
]

only. If you want another directory (@crozhon's) lemme know.

[John Owens]

I can fix PR MTEPS.

[John Owens]

We could do a bar chart but it would be really really wide. I'll think about that!

[Muhammad Osama]

We could do a bar chart but it would be really really wide. I'll think about that!

I figured, but if there's a better way to add some separation to datasets and have mini plots in that giant plot. Anything that shows the difference between the GPUs is nice.

[Muhammad Osama]

This is

roots = [
    "../gunrock-output/v1-0-0/sssp",
    "../gunrock-output/v1-0-0/bc",
    "../gunrock-output/v1-0-0/tc",
    "../gunrock-output/v1-0-0/pr",
    "../gunrock-output/v1-0-0/bfs",
]

only. If you want another directory (@crozhon's) lemme know.

Please include these:

gunrock-output/launch_bounds_comparison/*
gunrock-output/cuda_arch_comparison/*

[John Owens]

OK. I'll do that for all of them. But, I'm pretty sure it's not a complete set of runs in the way that the v-1-0-0 are.

[Chuck Rozhon]

Yes they are not complete sets. But everything in ‘cuda_arch_comparsion’ is in dev branch.

[Muhammad Osama]

That's fine, we can eventually get to a complete set after all your changes are pushed @crozhon. I would just like to see the entire v1.x.x runs anyways.

[John Owens]

@neoblizz My current recipe is "copy edges-queued into edges-visited, compute m-teps, then normalize runtime by the number of iterations". I assume edges-queued is global (across entire computation), not per iteration, so I need to do the m-teps compute before I normalize runtime. Tell me yea or nay.

[John Owens]

So I am hoping this is good enough and I don't have to turn this into a bar. We now only have one dot per GPU. It is feasible to jitter these in x but it's more involved in terms of the programming.

PR:
BC:
SSSP:

[Muhammad Osama]

@neoblizz My current recipe is "copy edges-queued into edges-visited, compute m-teps, then normalize runtime by the number of iterations". I assume edges-queued is global (across entire computation), not per iteration, so I need to do the m-teps compute before I normalize runtime. Tell me yea or nay.

That's correct.

This looks good. Just missing BFS now.

[John Owens]

Oh yeah forgot (BFS).

[John Owens]

DOBFS:
BFS:

[Muhammad Osama]

Also can go live! 🎉

[John Owens]

Are you cool with what we have now?

[Muhammad Osama]

Yes!