This is a very useful way of thinking about responses and latency, and a great example of a domain with a lot of very specialized knowledge (industrial engineering and logistics) that probably doesn't overall have enough cross-talk into CS given how much we talk about queues, latency, and workers.

I'm going to have to brush up on my IE one of these days to try to figure out if there are some useful insights to potentially be gleaned from models that account for server-level concurrency more directly, along the lines of https://en.wikipedia.org/wiki/M/M/c_queue

Instead of giving the entire dataset to the model (which in this case already identifies a performance limit visually), give a partial set and then see if these algorithms can predict the remaining known values. That should be the real test of whether prediction works.

Here's what I found last time I did it: https://github.com/brendangregg/PerfModels/blob/master/scale...

ie, I'm splitting the dataset into "model input" and "extra measurements". USL was not really working.

I think that the 10th data point in your model is an outlier.

Could you test it again

1) by adding 11th data point, 2) without 10th data point, 3) with 10th data point corrected?

Everyone interested in this should have a read of the Neil Gunther's own books: http://www.perfdynamics.com/books.html