KILLs, server or operator generated, have long been a plea on IRC networks.
Some years ago, I was keeping some stats on how many KILLs were generated,
but stopped. A year after The Great Split, I felt like it
was time to have a new look at this topic and see if, as I thought,
the problem has been improved, or if I was just dreaming.
It is difficult to do any comparison with the past as I don't have any
good statistics left. It seems that the net hit 20k users in
February 96. An old file shows that in October 96, I had recorded
1726818 server kills, and 16239 operator kills over 137 days.
|Server Kills||Operator Kills|
It's difficult to use these numbers as there is no track of the
average number of users during that period. They are provided
for your information.
We have two networks to look at, they have both chosen a different path
in fighting nickname collisions (server kills). It's extremely difficult
to do any valuable comparison between them for various reasons:
- lack of past statistics.
- difficulty of getting accurate statistics (especially on EFnet).
- different userbase (meaning possibly different abuse types).
- different userbase size (EFnet peaks at 33K, IRCNet at 17k).
The two different features in short:
- TimeStamping: The number of kills is reduced by 50% as
only one of the 2 clients is killed. The number of collisions
- Nick Delay: Collisions are unchanged but avoided if possible
by forbidding some nick changes during splits. (Efficiency goes
from 0% to 100% depending on circumstances, average unknown).
- Since EFnet only sends each KILL in one direction, it's
difficult to estimate the real number of server kills occurring.
If server kills are equally distributed on the network, one could
simply assume that the probability to be on the side that sees the
KILL is 50%, in which case the number of KILLs seen is half of the
total number of KILLs.
- On IRCnet, each KILL affects (usually) 2 users, however the
simplistic "1 kill = 1 collision" formula is valid.
Because I'm just doing this out of curiosity, I have only taken
a 24 hour sample on both nets (from approximatively 20:30 EDT Monday
97/09/01). This isn't the best but should be enough for the
conclusions I plan to not draw.
|Same user server kills||N/A||>=167|
- Operator kills: There isn't much to say, both nets have
a different attitude regarding these. EFnet's number of /kill has
increased significantly, while IRCNet has seen the opposite change.
About a third of all KILLs on EFnet now come from an operator.
- Server kills: It's obvious in both cases, TS and ND work
and have led to a significant reduction of nickname collisions. I'll
try to go a little bit further, but the above is the only
- If we follow the reasoning described above, we then see that
EFnet has almost 3 times as many collisions as IRCnet, for twice as
many users (818*2=1636 and 580*2.8=1636). (If the number of
collisions is a linear function of the number of users, then we
could say that TS and ND have a similar efficiency). (2 IFs here)
- Why EFnet still has so many collisions is something I don't
understand. Unlike ND, TS makes it impossible to collide someone
if that person doesn't get disconnected and doesn't change nicknames.
- users are stupid and keep trying? (I can't believe it)
- these are just unlucky users? (it seems to be too many to be true)
- Chris Behrens (Comstud) says that he is "really not sure that TS
is working 100% for nicknames".
- a little bit of 1, 2 and eventually 3 ?
- About 30% of IRCNet kills are collisions due to someone changing
servers. (Something that TS gracefully deals with)
Both methods are very efficient, and since they seem to work in opposite
ways, I can already see people thinking (again) "let's combine them".
Well, The Great Split was partly caused because nobody agreed to do
this, I believe it'd be really sad to see it happen now. However,
it isn't the reason why I still think it isn't worth it:
- the gain would be minimum.
- the cost of adding TS would be high (bandwidth & memory wise) [Sorry,
I don't have any number on this anymore] (However, ND's cost is quite
low: about 300Kb on IRCNet)
- both methods use a different philosophy and I believe it's good
to explore various options. (IMO)
Christophe Kalt <email@example.com>