[c-nsp] Cisco's new 4500-X 10G Aggregation Switches

[Ian Cox]

On 2/16/12 1:52 AM, Saku Ytti wrote:
> On (2012-02-16 17:31 +0800), Lincoln Dale wrote:
>
>> Stride doesn't come into it as its not likely a mtrie lookup but
>> rather FIB is a TCAM.
> Yeah, I guess SUP7 isn't NPU+DRAM yet :>
>
>> Performance is therefore less because to do a 144 bit lookup rather
>> than 72 bit key (IPV4) is logically a double-wide lookup.
>>
>> Not all platforms have to do that but many do.
> Yes. And point remains, if you'd implement it in TCAM where atom is 144b,
> it would artificially halve your IPv4 rate and your IPv4 table size, with
> no apparent design wins.
>
> If there is way to make IPv4 and IPv6 same rate, without screwing up IPv4,
> I'd be curious to hear what it is.
You design the forwarding path to be able to do 2 reads to FIB to get
144b of information instead of 72b for a single packet lookup. This is
what is done on the N7K-F248XP-25 modules in the Nexus 7000, where IPv4
and IPv6 performance is the same.  It is a design tradeoff in the end of
pps compared to increasing internal datapaths to handle 4x larger
addresses. 

Ian