mbuf underrun, zombie process

[Bánhalmi Csaba]

hi!

i hope i'm writin' to the right list this time..
my problem is the following: my server has high traffic on 100mbit dedicated
line (around 10-11,5mb/s all day)
with approx ~100-120 tcp connection. the nic is an intel pro 100, used with
polling.
now..when i set my data proxy's tcp window size to optimal, the network
traffic dies in 2-3hours and
the process become zombie, can't even kill.
when i use small or large tcp window size the data proxy is stable, however
the traffic is not good, dsl
users complain, high speed copy is slow, etc.
i don't really know what to do, what sysctl should increase or anything, if
anyone had same problem,
please post..
my machine is a pc, amd 2000+, around 50-60% used cpu, 512 ram and this is
the only high load process.
i changed couple sysctl settings:

kern.ipc.nmbclusters=32768
kern.polling.enable=1
kern.ipc.somaxconn=4096
net.inet.udp.maxdgram=32768
net.inet.tcp.inflight.enable=0
net.inet.ip.portrange.randomized=0

ps: i tried to increase nmbclusters (65536, 131072), increased somaxconn
(8192) and
tcp.recvspace/sendspace to 65536, but didn't helped.