some OOB packet loss stats

An aspect of OOB that is rarely talked about is that UDP does lose
packets now and then.  We gathered some statistics from OOB sessions
that have succesfuly exchanged the LIME12 message.  All of the
sessions below have been active for at least 10 seconds after
receiving LIME12 so the loss of results could be due either to loss of
datagrams carrying QueryReply messages or LIME11 messages.  The data
was collected from random nodes in the network and it contains both
Ultrapeers and Leaves.

Overall success ratio  (In reality the success ratio is even lower
because LIME12 can get lost as well.)
9881 / 11399 = 86.683044%

Number of Response Datagrams in session : Total Number Received /
Total Number Requested = Success %

1 : 3953 / 4245 = 93.12132% samples: 4245
2 : 1389 / 1538 = 90.312096% samples: 769
3 : 837 / 912 = 91.77631% samples: 304
4 : 477 / 548 = 87.04379% samples: 137
5 : 420 / 510 = 82.35294% samples: 102
6 : 329 / 438 = 75.11416% samples: 73
7 : 249 / 273 = 91.20879% samples: 39
8 : 179 / 208 = 86.05769% samples: 26
9 : 154 / 180 = 85.55556% samples: 20
10 : 105 / 140 = 75.0% samples: 14

Its also worth noting that LimeWire sends one Response per datagram,
and as you see the more datagrams are exchanged the higher the
likelihood some get lost.

We have been discussing the issue in our Issue-tracking system:
https://www.limewire.org/jira/browse/LWC-1084 and we welcome any
comments there.  Some of the ideas on the table include erasure
encoding of the responses and sending duplicate LIME11 and LIME12s
(backwards compatible).

We look forward to your comments.

Zlatin Balevsky
Lime Wire LLC