Re: why does netfilter make upload very slow? (was: Re: e1000 -> 82540EM on linux 2.6.0-test[45] very slow in one direction)

Hi Florian!

I'm Cc'ing all the mailinglists in order to keep them posted about the
question you've raised there.  All further discussion will move to
netfilter-devel, so for those interested: Please continue there.

On Wed, Oct 15, 2003 at 10:28:50AM +0200, Florian Zwoch wrote:

quoted

a) CONFIG_NETFILTER disabled.  you won't even have the netfilter hooks
  in the network stack (so certainly no netfilter-using modules loaded)

no problem

quoted

b) CONFIG_NETFILTER enabled, but _no_ modules (iptable_filter,
  ip_conntrack, ...) attached to the netfilter hook

no problem

quoted

c) CONFIG_NETFILTER enabled and iptable_filter.o (which pulls ip_tables.o)
  loaded, NO RULES in the table

no problem

quoted

d) CONFIG_NETFILTER enabled and iptable_filter.o (which pulls ip_tables.o)
  loaded, RULES in the table

no problem (as long as i dont load any rules that require ip_conntrack)

quoted

e) CONFIG_NETFILTER enabled and ip_conntrack.o loaded, iptable_filter
  loaded or not, rules or not

*boink*

So It's clearly the connection tracking subsystem.  This is on one hand
good (because it means it's neither netfilter nor iptables).  

whenever i try to load ip_conntrack the nic performance drops from 5mb/s 
to 200k/s.

On the other hand, this is definitely way worse than you would expect.
Can you please tell me more information about:

- number of connections you have? (cat /proc/net/ip_conntrack | wc -l)
- number of buckets and ip_conntrack_max (printed at ip_conntrack
  loadtime
- your traffic pattern.  Are you spraying udp packets with random
  src/dst? What kind of connections (protocol, application) are you
  testing with?
- what about the hardware (cpu, memory, smp?)

Even the worst tests we've had so far (random UDP packets) 'only'
reduced the througput by about 50%.   Maybe we can do better than 50%
worst case behaviour, but you will always observe a visible impact as
soon as you start connection tracking for every single packet (which is
what 'insmod ip_conntrack' implies).

still using 2.6.0-test6.

Have you observed this behaviour with other kernel versions?  Was there
a performance change between 2.4 and 2.6?  Or did you always observe
this grave performance loss?

regards,
Florian

-- 
- Harald Welte [off-list ref]             http://www.netfilter.org/
============================================================================
  "Fragmentation is like classful addressing -- an interesting early
   architectural error that shows how much experimentation was going
   on while IP was being designed."                    -- Paul Vixie