Date: Mon, 20 Jul 2009 07:31:13 +0200 (CEST)
From: Mikael Abrahamsson
> With a little creativity, it can _almost_ be done for IPv4.
That's most likely a big _almost_.
Maybe. And maybe I'm using worst-case synthetic test sets in addition
to real routing sets.
When someone asks for "2600 class router" they probably also want
"2600-like platform"
And I'm unaware of Cisco 2600-class routers that handle anywhere close
to 10 Gbps.
WFQ/fairqueue/LLQ, L2TPv3, PPPoE and a heap of other things that
impede pps quite a lot on a CPU based platform.
Perhaps the OP can clarify whether his omission of these was accidental,
because such features were assumed, or because he does not need them.
If you can bring all (or most) of the IOS functionality into a modern Intel
Xeon/i7 platform with all that memory access speed etc and you use all the
cores efficiently, then you might be able to do a lot. I've heard a lot of
And minimize both task switching and packets' in-queue time. I'm aware
of the requirements.
claims before (Luleå Algorithm from Effnet for instance) but it never came
I was unaware of Lulea. I've [obviously] not implemented it, and can't
comment on performance with modern loads and CPUs. However, it's
encumbered -- although I question the patent-worthiness of what I see
described. Route updates appear painful, which obviously would be
problematic. (I went down the painful-updates fox hole half a dozen
years ago. Yes, it's a dealbreaker.)
Other algorithms exist in the literature.
The truly insane might even be able to "strike gold" with a little
creativity.
to much because functionality/stability is everything, if I want a stupid
We also could argue the stability of the routers that he has used, and
of COTS boxes. I seem to recall having to load an interim IOS release
(on 2600-series boxes even!) due to instability.
pps forwarding device I might as well get myself an L3 switch, it'll use
less power and have less parts that can break.
Perhaps the OP can clarify his requirements. I understood him to want
low cost and high PPS, with IPv6 being mandatory. A list of priorities
and non-priorities might be useful.
I interpretted the post as being keen on high processing power and low
cost.
On a semi-related note: Has anyone dealt with Cavium (or similar) NICs?
Eddy