On Fri, Sep 18, 2015 at 9:49 AM, Jari Arkko <jari.arkko@xxxxxxxxx> wrote: > Donald, > >>> (Maybe this helps: I’m not actually sure why in a k-element set you order them based <something> mod k because that would seem to produce likely duplicates. Since your backup option in the case of duplicates is proper numeric sort, why just not do that and only that? E.g. "RBridges are sorted in byte string ascending order by their LAALP IDs, or if they are equal, by their System IDs considered as unsigned integers.” But it could also be that it is too early and I have not yet had enough Diet Coke…) >> >> I believe the idea is to quasi-randomize the order. The DF election is >> per VLAN and a goal is to spread the multicast traffic across the >> RBridges in the active-active edge group. > > It is a fine goal to randomise the order. > > My only observation of the current setup is that if you randomise a > k-element group through "mod k” operation, you will likely have > some number of collisions in the result. I don’t know enough > about math to calculate the percentage. But for the sake of > argument, if k=2 it seems that the likelihood of collision is 50%. > > And for every collision, your order becomes no longer random > but simply numerical order of the identifiers. In our degenerate > k=2 example it seems that in 50% of the cases you have a random > order and 50% of the cases you have numerical order. I’m sure > there would be other ways to randomise the order with less > collisions, if avoiding numerical order is important. Well, the way to randomize the order with quite low probability of collisions is to sort by the hash of (System IDj | LAALP IDi), for example SHA-1(System IDj | LAALP IDi). Ties could still be broken by System ID which is guaranteed to be unique but ties would be quite rare. This seems like a minor localized change. > Jari Thanks, Donald ============================= Donald E. Eastlake 3rd +1-508-333-2270 (cell) 155 Beaver Street, Milford, MA 01757 USA d3e3e3@xxxxxxxxx