Network performance measurements of -current

As part my funded TCP/IP optimization work I'm doing lots of measurements
and profiling with an Agilent N2X network tester and calibrated traffic
generator.

The following data shall serve as baseline of the current performance we
get out of FreeBSD 7-current.  More to come tomorrow though.

OS: FreeBSD 7-current as of 20051222-1600 UTC
KERNEL: Generic kernel, minus WITNESS and INVARIANTS, plus HWPMC, HZ=1000
HARDWARE: Dual Opteron 852 2.6Ghz, Tyan S2882 Mobo with AMD-8131 PCI-X tunnel
HARDWARE: dual Broadcom Gigabit BMC5704C PCI-X-133 ("bge")
HARDWARE: dual Intel Gigabit 82546EB PCI-X-133 ("em")

Uniprocessor kernel

 bge:
  normal forwarding bge0->bge1:   _at_64/326kpps/166us/402kpps(30%Loss)/194us
  normal forwarding bge0->bge1:   _at_1500/81kpps/520us
  normal forwarding bge0->disc0:  _at_64/1205kpps
  IP fastforwarding bge0->bge1:   _at_64/565kpps/192us/575kpps(60%Loss)/1090us
  IP fastforwarding bge0->bge1:   _at_1500/81kpps/730us
  IP fastforwarding bge0->disc0:  _at_64/1160kpps
  net.isr.direct=1  bge0->bge1:   _at_64/476kpps/211us/487kpps(68%Loss)/1284us
  net.isr.direct=1  bge0->bge1:   _at_1500/81kpps/760us
  net.isr.direct=1  bge0->disc0:  _at_64/1250kpps
  polling (*) bge0->bge1:        
_at_64/420kpps(9%Loss)/1385us/416kpps(72%Loss)/1600us
  polling (*) bge0->bge1:         _at_1500/71kpps(9%Loss)/850us
  polling (*) bge0->disc0:        _at_64/697kpps

 Comments: Under full load the normal processing breaks completely down
  while with IP fastforwarding it levels off but continues to forward.
  Strangely with polling it has 9% loss at all loads (even at 1% wirespeed).
  May be related to HZ=1000.

 em:
  normal forwarding em0->em1:     _at_64/372kpps/112us/396kpps(11%Loss)/131us
  normal forwarding em0->em1:     _at_1500/81kpps/170us
  normal forwarding em0->disc0:   _at_64/1130kpps
  IP fastforwarding em0->em1:     _at_64/565kpps/45us/585kpps(4%Loss)/1600us
  IP fastforwarding em0->em1:     _at_1500/81kpps/135us
  IP fastforwarding em0->disc0:   _at_64/1116kpps
  net.isr.direct=1  em0->em1:     later
  net.isr.direct=1  em0->disc0:   later
  polling (*) em0->em1:           later
  polling (*) em0->disc0:         later


(*) max_burst=1000, user_frac=0, each_burst=30

Sponsored by:	TCP/IP Optimization Fundraise 2005

-- 
Andre