After an exciting week-end of debugging, we've found a bug in the smoothing algorithm that could significantly slow down convergence in some cases. If you're running head with smoothing enabled, I strongly recommend upgrading.
http://git.wifi.pps.univ-paris-diderot.fr/?p=babeld.git;a=commit;h=e8fe07 Many thanks to the person who found the issue and suggested the fix. The issue is as follows. Suppose we have a nice, stable route R1: R1: metric = 256, smoothed_metric = 256 Now the current node moves away, and a new route appears, while the old one gets retracted: R1: metric = infinity, smoothed_metric = 256 R2: metric = 1024, smoothed_metric = 32768 In this case, we want to switch to R2 straight away, without waiting for the smoothed metric to converge. So upon a retraction, we now slam the smoothed metric to infinity straight away: R1: metric = infinity, smoothed_metric = infinity R2: metric = 1024, smoothed_metric = 32768 Since we only do that when a metric reaches infinity, it will only inhibit smoothing for retracted routes, not for routes that oscillate with finite metric. -- Juliusz _______________________________________________ Babel-users mailing list [email protected] http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/babel-users
