Revisit Pace-Filter
We should revisit the pace filters with respect to the following:
-
Window Size and filter-function should be pace-dependent (i.e. we should employ different ones for the system paces as for the network paces).
-
Filter-functions in general should be checked for prediction quality as well as runtime complexity. Maybe a 95-percentile filter is a good choice conceptually, but might be prohibitive in terms of runtime overhead.
-
Cooperation of pace-filters. Currently, we use multiple filters for the different components of a single pace (i.e. internal, external and dependent). Maybe they should be synchronized, i.e. they expire synchronously. The question is then obviously how we update the current three-tuple if a new three-tuple comes in (which tuple-component is used to apply MAX?).