[Bloat] tiny monsters: multicast packets

[Dave Taht <dave.taht@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 2534 bytes --]

So after my experiments [1] yesterday with the wndr3700v2 hardware[2], I
came away
even more convinced that the wireless world and the wired worlds should
not be bridged together.

All the AQMs out there assume that it takes the same period of time to
deliver a packet consisting of X bytes to the next hop. Wired more or less
does that.

Wireless breaks that assumption. It wasn't so bad, back in the early days of
802.11b - 802.11b ran as fast as 11Mbit, and multicast, at 2... for a ratio
of 5.5x1.

11g came around, and runs at 54Mbit, and - (if you don't run in mixed mode,
still supporting B), you can multicast at 6Mbit. But nearly everyone does
still run mixed mode, so multicast is stuck at 2... for a ratio of 26x1....
instead of a mere 9x1.

Soo... 11n has come around, and I shudder to think of what packet rate ratio
of "normal" vs "multicast packets" do to the assumptions of the rest of the
stack.

So with just a little multicast going through your wireless network, any
assumptions the higher level portions of the stack might make are invalid.
HTB? hah, uses fixed buckets... RED? a single multicast packet is a monster
packet, how's it supposed to find it in the swamp?

Worse, most multicast packets are statistically rare and needed for the
network to actually continue to function.

In my last 2 months of travel, I have seen multicast packets, such as ARP,
DHCP, MDNS, and now babel, all failing far, far, far more often than is
desirable. I have seen DHCP fail completely for hours at a time, I've seen
ARP take dozens of queries to resolve.

Nextly, It is trivial to trigger the symptoms of bufferbloat, with a
multicast stream.

Perhaps eBDP can handle multicast well, but certainly AQMs are going to have
headaches that are difficult to solve at ratios between normal and multicast
packets this poor.

Lastly, most home router vendors bridge wired and wireless together, sort of
like jamming together "jet engines and a vw bugs",  and I finally broke them
apart [3], to try and look at them separately - as even the switch is
displaying 100+ms of buffering when I slam it with 4 simultaneous iperf
streams [4]

1: https://lists.bufferbloat.net/pipermail/bloat-devel/2011-May/000156.html
2: http://www.bufferbloat.net/projects/bismark/wiki/Capetown
3: http://www.bufferbloat.net/issues/186
4: http://www.bufferbloat.net/projects/bismark-testbed/wiki/Experiment_-_QoS

-- 
Dave Täht
SKYPE: davetaht
US Tel: 1-239-829-5608
http://the-edge.blogspot.com

[-- Attachment #2: Type: text/html, Size: 3016 bytes --]