Ethical Hacking

Learn to find vulnerabilities before the bad guys do! Gain real world hands on hacking experience in our state of the art hacking lab. Course designed and taught by expert instructors with years of penetration testing experience. 12 student maximum in every class. Certification attempt included in every package.
Computer Forensics Training at InfoSec Institute

Gain the in-demand skills of a certified computer examiner, learn to recover trace data left behind by fraud, theft, and cybercrime perpetrators. Discover the source of computer crime and abuse at your organization so that it never happens again. All of our class sizes are guaranteed to be 12 students or less to facilitate one-on-one interaction with one of our expert instructors.




Network Security Firewalls
[Top] [All Lists]

Re: Question about high perfomance iptables-based Firewall

Subject: Re: Question about high perfomance iptables-based Firewall
Date: Tue, 6 Sep 2005 21:16:22 +0200
Greetings!

On Tue, 06 Sep 2005 10:08:09 +0200
Javier Miguel Rodríguez <javier.miguel@talika.eii.us.es> wrote:

I have to build a FAST Linux firewall (12 gigabit ethernets), and I
need  your advice

Compaq Proliant DL380g4 (1 xeon 3.6 ghz, with hyperthreading, pci-x 
based, 1 gb ram)
2 broadcom gigabit ethernet cards (tg3 driver)
3 intel quad gigabit ethernet cards (e1000 driver)

...but with 64bit PCI-X @100 MHz for most of the interfaces you have an
internal bottleneck with a theoretical limit at ~6 Gbit/s for the bus
lines alone.

According to the proceedings of Netfilter Developer Workshop 2004
(09/2004):

        "Problems with Intel 4 port cards because of additional PCI bridge.
        Lose 30% of performance 
[...]
        PCI-X will have similar problems with multi-port cards E1000 PCI-x
        is broken with message signal interrupts. 
[...]
        building as 64bit kernel hurts performance, this is assumed because skb
        grows another cache line because of large pointers
[...]
        ip stack needs help to get better SMP scaleability
        with perfect affinity CPU's get only 20% gain with second Xeon"

Full test (nat, mangle, filter, ip_conntrack): down to 40% of raw
throughput.



My ruleset will be rather short: 500-600 lines, with SNAT/DNAT in
about  5% of these rules. Only ipv4 will be used in this firewall
setup.

According to HIPAC testing the throughput started caving in at ~500
rules, so make sure you have the right order (most used on top) and do
not grow larger.


I expect substained rates of 300-400 megabits on EACH gigabit
interface,  

400 Mbit/s * 11 interfaces = 4,4 Gbit/s
With ~6 Gbit/s as theoretical upper bus line limit alone that's quite a
close call. Especially if applying the reported 30% loss for the intel
quad cards...


So my guess is that the hardware won't support it. For one the bus limit
alone probably won't allow the throughout (I don't know the
interconnections between the busses - thus assumed the PCI-X as main
bus). I am not sure about the CPU power, but for my gut feeling that's
uncomfortably close, too.

Compared with CheckPoint firewall datasheets the HP DL-380 is put into
the 1-3Gbit/s category, so that's not increasing hope either.

So I'd suggest, to be on the safe side either do a gateway
reconfiguration so you can distribute the load onto several dedicated FW
systems (one Mail-FW, a Web-FW, one VPN-System, etc.) - or head for the
really big, proprietary irons, which usually are of the load-balanced
blade type (e.g. the big Crossbeams, Alteons). But with them you are at
the end of the expansion road. 

So the next network traffic increase will have to be countered with a
new network topology anyway. So I'd suggest to really re-think your
current network topology.

Good luck!

Volker

-- 

Volker Tanger    http://www.wyae.de/volker.tanger/
--------------------------------------------------
vtlists@wyae.de                    PGP Fingerprint
378A 7DA7 4F20 C2F3 5BCC  8340 7424 6122 BB83 B8CB

<Prev in Thread] Current Thread [Next in Thread>