Many Many table rules

Michael Eck michael.eck at gmail.com
Wed Sep 15 14:21:24 CEST 2004


I've been considering both of your suggestions.   I'm already doing
something that accomplishes your second suggestion.  My last rule and
first two rules are for CONNMARK saving and restoring (I'm marking not
for acceptence but for htb classification).

/sbin/iptables -t mangle -A INBOUND-SHAPER -p tcp -j CONNMARK --restore-mark
/sbin/iptables -t mangle -A INBOUND-SHAPER -p tcp -m mark ! --mark 0 -j ACCEPT
...
/sbin/iptables -t mangle -A INBOUND-SHAPER -p tcp -m mark ! --mark 0
-j CONNMARK --save-mark
/sbin/iptables -t mangle -A INBOUND-SHAPER -p tcp -m mark ! --mark 0 -j ACCEPT

Your first suggestion would, in my case, work better by first matching
by IP.  How much performance gain would I really achieve?  Is there a
way to quantify the impact that a given number of rules would have? 
In other words, is the difference between 200 and 1000 rules dramatic?

Thanks for your help.


On Tue, 14 Sep 2004 14:24:05 -0500, Aleksandar Milivojevic
<amilivojevic at pbl.ca> wrote:
> Michael Eck wrote:
> > I have a bridging traffic shaper that uses htb and sfq.  My iptables
> > and kernel are patched with ipp2p and layer-7 filtering to mark p2p
> > traffic.  Currently, this unit is at the head end of a broadband
> > network.  I'm dividing up my users into htb classes based on their
> > location on the network.  This amounts to 5 rules per IP address, 1
> > for generic traffic, and 4 for ipp2p and a few layer7 rules.  That
> > puts me at about 820 rules for the inside interface portion of my
> > iptables.
> 
> It might be a good idea to optimize things using user defined chains.
> That way, packet that matches your 820th rule, wouldn't need to go
> through 820 rules.
> 
> For example, simple solution would be to have rules that match only by
> protocol, and than from them jump into the chain where you match by IP
> address, something like:
> 
>    -A FORWARD -p tcp --dport aaa -j mychain
>    -A FORWARD -p tcp --dport bbb -j mychain
>    -A FORWARD -p udp --dport ccc -j mychain
>    -A mychain -s a1.b1.c1.d1 -d e1.f1.g1.h1 -j ACCEPT
>    -A mychain -s a2.b2.c2.d2 -d e2.f2.g2.h2 -j ACCEPT
>        ... and so on ...
>    -A mychain -j DROP
> 
> Or you could do it the other way around (first by IP addresses, and than
> by protocol)...  Even with simple setup, if you have 200 clients and you
> are allowing 5 protocols per client, in worst case packet would go
> through 205 rules instead of 1000 rules.
> 
> Another idea that might work for you is using statefull matching, and
> putting rule that matches ESTABLISHED packets as the first rule of
> INPUT, OUTPUT and FORWARD chains, for example:
> 
>    -A FORWARD -m state --state ESTABLISHED -j ACCEPT
> 
> If this is the very first rule in forward chain, 99.99% of your network
> traffic will be matched by it (so as performance goes, it's like you
> have only one rule instead of 820).
> 
> Only the very first packet of each connection would have to go through
> subsequent rules (say 400 rules on average?).  And if you presorted
> packets to user defined chains first, than this number will be even lower.
> 
> --
> Aleksandar Milivojevic <amilivojevic at pbl.ca>    Pollard Banknote Limited
> Systems Administrator                           1499 Buffalo Place
> Tel: (204) 474-2323 ext 276                     Winnipeg, MB  R3T 1L7
> 
>



More information about the netfilter mailing list