I wasn't aware of the treatment of multicast packets as less than best effort in wireless transmission. That is not exactly intuitive, given that radio is inherently broadcast.
Regarding relative loss probability, we may not be talking about the same thing. I am referring to the increased probability of losing at least one packet when the packets are replicated. In the wired world, things that determine the probability of dropping one packet (corruption, congestion, etc.) likely determine the probability of dropping others. What I was saying is that - if there is a .0001 % chance of dropping a unicast packet, and (assuming the same probability for each multicast packet) a .0001 % chance of dropping each of 3 replicated multicast packets, there is about a .0003 % probability of dropping at least one of the 3 multicast packets. This is only an approximation because the probabilities are not precisely additive. For values this small, however, it is a reasonably close approximation. This might not be obvious in deployments, given the relatively small amount of real multicast traffic and the very large amount of data that will need to be accumulated for this difference to be seen as more than statistical noise. There is also the counting difficulty; in gathering the data, it would be necessary to correlate multicast replications to determine the total probability for multicast loss for the group. Without this correlation, each multicast packet loss would have to be considered independently of corresponding replicants. It seems possible to me that you are referring to the probability of packet loss for multicast packets independently of each other. This, too, is not intuitive given that losing any multicast packet in a multicast service represents a failure of the service. -- Eric -----Original Message----- From: Mikael Abrahamsson [mailto:swm...@swm.pp.se] Sent: Thursday, August 06, 2015 10:27 AM To: Eric Gray Cc: Glenn Parsons; Alia Atlas; Acee Lindem (acee); Toerless Eckert (eckert); Homenet; Dan Romascanu (droma...@avaya.com) Subject: RE: [homenet] Despair Importance: High On Thu, 6 Aug 2015, Eric Gray wrote: > It strikes me as something of a mistake generally to assume that > multicast is as reliable as unicast. > > Unicast reliability depends on the mechanism(s) used to ensure > reliability. Unicast traffic tends to get lost every now and then. Nobody doubts that packets get lost, but the general tendency since IP networking was invented, was that multicast delivery of packets wasn't especially worse than unicast. Packets get lost, but generally less than 1% get lost, and multicast and unicast are affected equally. > All the same factors that affect unicast packet delivery also affect > delivery of each packet with multicast. Hence multicast reliability > should be worse than unicast reliability by an amount roughly > proportional to the amount of packet replication necessary to support > it. Hm, care to elaborate? That seems a lot worse than my experience in deploying networks would tell me. > Each replicated packet is as likely to be lost as any unicast packet. > Loss of one or more packets should be expected to be more likely with > multiple packets than with a single packet. But it's still only a single packet per link. > Multicast reliability, even when considered at the link level and > assuming replication is not required in transmission of multicast > packets onto the link itself, is only slightly better. As > full-duplex, point-to-point connectivity becomes increasingly likely > (fat yellow cables are relatively rare any more), data replication > still occurs - just not at the level where a router sending packets > onto the link is likely to be aware of it. Correct, as of 20 years ago or something we do not use 10base5 so L2 devices do L2 replication. > Hence it is interesting in this discussion that we are talking about > an assumption that seems broken at the start. > > Have I missed something? Well, 802.11 treats multicast (and broadcast) packets as a second rate citizen, I am not aware of any other L1/L2 technology that does this. 3GPP uses basically a point-to-point tunnel, so unicast and multicast is treated in very similar fashion without multicast being at a disadvantage. So IETF needs to sit down and work out a strategy on how its protocols should work going forward, if everybody who designs protocols in the IETF should be told that multicast and broadcast "doesn't work properly", and act accordingly. What probably needs to happen is that over time, the IETF should try to use less multicast, but on the other hand, 802.11 really needs to make sure that multicast works a lot better than it does today. -- Mikael Abrahamsson email: swm...@swm.pp.se _______________________________________________ homenet mailing list homenet@ietf.org https://www.ietf.org/mailman/listinfo/homenet
