linux packet forwarding rate

linux packet forwarding rate

[Philip Ho]

Dear all

I want to forward 1Gbps at line rate. Has anyone successfully used linux
to forward 1Gbps (assuming 1500 byte packet length, that is equivalent
to 80kpps) on gigabit ethernet?

If yes, how to configure the kernel? remove all the unnecessary modules?
Is FAST_SWITCHING useful?

If no, why? what is the upper bound?

Thank you very much.

Regards,
philip
Hong Kong IP Multicast Initiative (HKIPMI)
Department of Information Engineering
The Chinese University of Hong Kong
Phone : 2603 5240
Fax : 2603 5032

Re: linux packet forwarding rate

[Patrick Schaaf]

On Thu, Mar 06, 2003 at 02:53:56PM +0800, Philip Ho wrote:
> Dear all
> 
> I want to forward 1Gbps at line rate. Has anyone successfully used linux
> to forward 1Gbps (assuming 1500 byte packet length, that is equivalent
> to 80kpps) on gigabit ethernet?

Please ask again on a more suitable mailing list. The question has
nothing to do with further development of iptables / netfilter.
A suitable list may be:

http://www.uwsg.iu.edu/hypermail/linux/net/

Nevertheless, I'm sure the S/390, PowerPC, Sparc, Alpha, and HP/UX ports
managed that for some years, now. Also high end four- and eight-processor
x86 systems, may have made it.

BTW, with line rate, one usually means the absolute worst case line
rate with SMALL packets. Also, Ethernet is full duplex nowadays,
so the line rate of a single GE interface, at 2gbit/s, with 64 byte
packets, would be roughly 4000000 (4 million) packets per second.

On a (dual) Pentium-III 800Mhz system, on a 64/66 Mhz bus, I was able, about
two years ago, to push slightly more than 700mbit/s (synthetic UDP)
with jumbo frames, and I vaguely remember about 50000 packets per
second a processing ceiling.

Finally, the ONLY way to REALLY find out about these things, is to
get some hardware, test software, UNDERSTAND WHAT IS TESTED, and
the TEST!

Oh, and on 99% of real life linux servers, it just doesn't matter at all.

best regards
  Patrick

Re: linux packet forwarding rate

[Harald Welte]

[-- Attachment #1: Type: text/plain, Size: 1111 bytes --]

On Thu, Mar 06, 2003 at 08:13:36AM +0100, Patrick Schaaf wrote:
> Finally, the ONLY way to REALLY find out about these things, is to
> get some hardware, test software, UNDERSTAND WHAT IS TESTED, and
> the TEST!

Yet to be added:

- understand exactly what hardware can even do gbit linerate (not all
  NIC's can do this
- understand a lot about the linux network stack and know where you can
  tune it
- understand that without NAPI you don't have a very high chance

So unless you are a very experienced linux networking guy [or you hire
one], I think you will not be able to do a realistic test.  You will
just find out that in your configuration and the default setup it will
not work.

> best regards
>   Patrick

-- 
- Harald Welte <laforge@netfilter.org>             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

[-- Attachment #2: Type: application/pgp-signature, Size: 232 bytes --]