Question on client/server model in 4.2

[Don Libes]

Using the 4.2 implementation of TCP, each server is restricted to a
very limited number of simultaneous conversations.  One alternative
is to simulate many virtual circuits with sendto and recvfrom,
however I would guess that's very slow.

The IP protocol has its own drawbacks with its reliability and short
packet size.

What am I supposed to use if I want a server with 50 (or even 100)
clients and (reasonably) unlimited-size message lengths?  I am
willing to sacrifice speed.  Is the client/server (using circuits)
model wrong for clients > 32 and if so why?

Also, has anyone actually done any benchmarking of how much more
slowly communications occur when a connection has to be established
for each message?  If not, I will do this.