[Nel] TCP vs. UDP
Daniel Miller
miller@nevrax.com
Sun, 8 Jul 2001 18:44:09 +0200
I've been reading this thread with great interest. There's clearly a good
case for using UDP for low priority or short life-span data. In this case
the app software on top of NeL must clearly be designed with the fact that
some data isn't going to get through in mind.
The problem is particularly interesting because:
1. The limiting factor on the amount of data we transmit looks like being
the cost of output bandwidth for the servers.
- as far as I can ascertain, at today's costs we're looking at something
like 8-16 kilobits per client per second maximum output from the servers
(including lost packets)
2. The complexity of the scenes that we are trying to convey to the clients
is rising fast.
- view distances are increasing, as are screen resolutions which means that
you can see more dynamic world content at any given time (characters,
creatures, objects...)
- It is clear that for MMORPGs we wil soon need to be able to display crouds
or armies of 100s or even 1000s of characters and creatures at once.
- Character animation is becoming more complicated to reflect a mix of
actions performed simultaneously (such as speaking with emotional facial
expression and manipulating an object in one's hands while sitting down) -
which means more information to describe each character's state.
- and so on...
These two points combined mean that, whichever way one looks at it, a lot of
scene information will have to be filtered out or sent at low frequency. If
one isn't careful low frequency sends can obviously be very sensitive to
packet loss - which leads me to the point of this posting: In order to work
the problem we need to have a good understanding of true internet behaviour
and to have a good set of test data for simulating it.
The trouble right now is that I have no hard data to use to model packet
loss or delivery latency over time. If anybody knows of any studies that
have been done or has any of their own data I'd be very interested. In
particular I'm interested in moderately bad connections that exhibit both
good behaviour and bad behaviour over time.
Regards to all,
Daniel