Hello Harald & All, Thanks for your reply. Eventhough I don't use NAT now, I will have to use it soon. Connection setup really sucks after 'iptable_nat' module is inserted, my Pentium-4 RH9.0 box hangs with 7Mbps traffic using 20,000 concurrent UDP streams. Are there any tweaks that I can do in NAT code to increase performance of NAT, like: increase NAT hash-table size(I have got 1G RAM) or change hash algorithm etc?. For e.g: based on a recommendation from netfilter FAQ, I passed an odd-number(not a power of 2) as 'hashsize' parameter value to 'ip_conntrack' module, session-setup performance improved a lot after that in my tests when UDP packet streams with incrementing source IP address are sent. But this 'hashsize' tweak did not improve session-setup performance when UDP streams with incrementing destination IP address are sent. Is connection setup slow because of hash collisions i.e due to lot of connections getting hashed to same hash-bucket? or it due to lot of memcopies?. Can you please give some specific reasons for poor performance of NAT, it will help me in optimizing/tuning, if possible, that specifc part of netfilter nat functionality. Thanks, Kishore -----Original Message----- From: Harald Welte To: Kishore Dharmavaram Cc: 'netfilter@xxxxxxxxxxxxxxxxxxx'; 'netfilter-devel@xxxxxxxxxxxxxxxxxxx' Sent: 12/11/2003 11:13 PM Subject: Re: iptable_nat module slows/hoses my Redhat 9.0 box On Thu, Dec 11, 2003 at 09:47:25PM -0800, Kishore Dharmavaram wrote: > Hi All, > > I find that after inserting "iptable_nat" module on my RH9.0(2.4.20-19.9) > box, session-setup(ip_conntrack) performance gets very bad. yes, that's true. I think we cannot state often enogh: DO NOT load iptable_nat if you don't need nat. > I am wondering why "iptable_nat" is slowing doing the box so much when I > don't have any NAT rules defined. NAT is inherently complex. and in order to support any kind of nat, every connection (even the non-nat'ed ones) need to be placed in two additional hash tables. If you just use conntrack, a single hash table is sufficient.. > Do you guys know of any netfilter patches made to address/fix this issue?. There is no way to 'address' this issue other than not using NAT. > Thanks, > Kishore -- - Harald Welte <laforge@xxxxxxxxxxxxx> 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
