Computing, networks and machine learning

Service latency, classifier evaluation, training diagnostics and hardware performance models. Systems data is heavy-tailed almost by definition, which is why the tail percentile rather than the mean is plotted, and why the roofline and scaling charts are log-log: the interesting behaviour is a change of slope, and only a log axis makes a slope change visible.

Service latency percentiles over a day

Service latency percentiles over a day

ROC and precision-recall for an imbalanced problem

ROC and precision-recall for an imbalanced problem

Confusion matrix, normalised by true class

Confusion matrix, normalised by true class

Training curves for three learning rates

Training curves for three learning rates

Roofline model of kernel performance

Roofline model of kernel performance

A training run, revealed epoch by epoch

A training run, revealed epoch by epoch