Over the last week it seems to me that some improvement has gone into TorFlow and the BWauths are performing better. In particular the weightings calculated by each BWauth have stabilized and become less erratic. Several relays in the same region of a large network as my relay are now showing a roughly matching pattern of ascending measures, from smallest to largest: longclaw, gabelmoo, maatuska, moria1. One would expect this but it was not true in the past.
One fast relay I watch, 'pixelminer' had a tendency to be yanked from one extreme to another (e.g. 10k to 120k and back) and now has a placid consensus running in the middle of the range (still apparent in the {presumably stabilized} component values).
But all is not well. . .
Gabelmoo just crashed the weight for the relay here from 7150 to 945 in a single step. No way this is correct for a cleanly running 9375 Mbyte/sec relay with a 24-hour average load of 2175 Mbyte/sec per Blutmagie and 24-hour self measured peak of 5165. Peak easily runs all the way to the maximum. Have never seen such a low value reported by an individual BWauth. Currently have
longclaw Bandwidth=5165 Measured=3250 gabelmoo Bandwidth=5165 Measured=945 maatuska Bandwidth=5165 Measured=9730 moria1 Bandwidth=5165 Measured=25400
Recently set BandwidthRate/Burst to 8906000 (95% of actual bandwidth) per Mike Perry's 2009 research that shows this prevents TorFlow circuits from failing and prevents exponential TCP backoff.
tor-relays@lists.torproject.org