Performance Analysis: BatchTranslator

[browsermt]

Performance Analysis: BatchTranslator

[jerinphilip]

Using just WNGT20 and slight modifications to batch-translator, this notebook holds a preliminary analysis at the moment (will be updated as dev progresses). Next obvious task looks like:

• Have batcher operate in a random sample from WNGT20 mode to have enough samples for all (b, t) constructed from an, shuffling batches to get more of a realistic estimate.

[jerinphilip]

Obtained enough (b, t) to have some more datapoints. The triangle beneath the off-diagonal is expensive and takes longer to compute, hence currently just doing below b*t = constant. --mini-batch-words might actually be a decent heuristic to assume O(batch) = 1 and operate (since the same color path follows some hyperbola). Grid is (b, t). Colors indicate runtimes, with colormap 'jet'. red = longer, blue = shorter. We can possibly use the C(b, t) to come up with a better batch scheduling perhaps.

[jerinphilip]

• Account for burn in - Apparently the first batches are a bit off.
• Shift hack in Batcher + BatchTranslator to hack in Request, Response

[jerinphilip]

Abandoned, possibly results-to-efforts ratio is unappealing.