Crossover Plots: #Interactions on X vs Cumulative Time on Y

[matthewhammer]

• Kyle, Mike and Matthew discussed this a couple of weeks ago
• Plot axis: #Interactions on X vs Cumulative Time on Y
• Plot lines: From-scratch versus Incremental (Nominal and Structural)
• These plots are computable from our current dataset, for each benchmark that we currently have.

[kyleheadley]

If I recall, we were looking for the crossover of (X-initial + X-incremental * interactions = Y-initial + Y-incremental * interactions)

which gives us: crossover-interactions = (Y-initial - X-inital)/(X-incremental - Y-incremental)

I'll add this to the next set of data I gather, and we can see how it looks.

[matthewhammer]

@kyleheadley: My use of X and Y referred to the axis of the plot.

What are the "X" and "Y" in your comment above referring to? The two versions (incremental and non-incremental)?

[kyleheadley]

Yes, sorry for the confusion. Did you want to plot unique interactions? The way I remember it, we were 'extending' a single interaction, to generate a crossover point for each test.

I was just going to include it as a data item in our table instead of overhead.

----- Reply message -----
From: "Matthew Hammer" [email protected]
To: "plum-umd/adapton.ocaml" [email protected]
Cc: "Kyle N Headley" [email protected]
Subject: [adapton.ocaml] Crossover Plots: #Interactions on X vs Cumulative Time on Y (#28)
Date: Wed, Apr 22, 2015 2:40 pm

@kyleheadleyhttps://github.com/kyleheadley: My use of X and Y referred to the axis of the plot.

What are the "X" and "Y" in your comment above referring to? The two versions (incremental and non-incremental)?

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/28#issuecomment-95296373.

[matthewhammer]

Yes, after each initial run, we want to "extend" a single interaction sequence, by imagining that we repeat some interaction type over and over. We get two lines from that, one for the incremental version, and one for lazyrecalc, or whatever is modeling the always-do-it-from-scratch version.

Presenting this in place of overhead in a table sounds good to me. If things go well, the cross-over and overhead are nearly the same number, as we've discussed before with Mike.

Plots:

In addition to the table column for "Crossover", it'd also be great to see plots for each benchmark, as we discussed with Mike. I like seeing these, since they meaningfully compare a lot of data: both versions' initial times, as well as both versions' edit+update times. I think they'll be useful to us for diagnostics, as well as communicating the results in talks. For example, I'm giving a talk at CMU in a few weeks, and I'd like to show some cross-over plots to summarize all of this numerical data in just a few pictures.

For those plots, the X and Y axis are as described above.

Ideally we'd see lots of plots, for each benchmark, for each edit type, there'd be a cross over plot.

Maybe Nick knows how to do this in R easily, to avoid lots of busywork in Excel? I'm not sure.