C2/C3 states cause poor TCP performance with re(4)

[Aragon Gouveia]

Hi,

I've recently been hunting down a TCP performance issue with re(4) on a 
system that has C3 states enabled.  In short, outgoing TCP traffic shows 
good performance, but incoming TCP traffic has very unstable throughput 
with substantial packet loss evident in a tcpdump.

The same was observed with C2 states enabled.  Limiting to C1 states 
brought performance back to normal.

Can anyone confirm if this is reproducible with drivers other than 
re(4)?  Am I expecting too much in the hope of achieving both C3 states 
and good performance out of re(4)?  Any pointers on getting there?

This system is tuned using mav@'s tuning guide:

http://wiki.freebsd.org/TuningPowerConsumption


Thanks,
Aragon