[c-nsp] WCCPv2 Cisco 7600 + mask assignment problems

[Lincoln Dale (ltd)]

> > - using L2 rewrite rather than GRE encap to the cache scales
> >   more.
> > - L2 + XOR traffic-distribution-hash uses MLS cache entries
> >   for partial h/w acceleration (1st flow in s/w sets up MLS
> >   entry) for intercepted flows.
> > - L2 + "hash-mask" traffic-distribution means _all_ traffic
> >   stays in h/w CEF path
> >
> > the former will get you up to a few hundred Mbps, the latter
> > will scale to the maximum h/w forwarding of the platform.
> 
> Given that one doesn't configure GRE, how do you find out what
> mode they're running? Do they (the router and the cache) somehow
> 'negotiate' that, or do you configure it on the router?

all of the WCCP 'capabilities' (Forwarding Method - L2 or GRE,
Assignment Method - XOR Hash or Hash Mask) are negotiated between the
router & cache based on their capabilities.

generally speaking, it is up to policy on the cache to set the 'best'
possible method based on the router telling what it is capable of.

e.g. see
<http://www.cisco.com/en/US/products/sw/conntsw/ps491/products_configura
tion_guide_chapter09186a0080614f0c.html>

> When using GRE encap, why is performance only 'a few hundred
> Mbps'? On the Sup720, can't GRE be offloaded in hardware?

unfortunately, it isn't 100% GRE.  it is GRE + an additional 4 or 8 byte
header, which means it cannot be done in h/w on Sup720.

> > what does your cache vendor support?  they should support all
> > permutations & if you used the latter, it'll scale to the
> > limits of the h/w.
> 
> Does anyone other than Cisco support "hash mask"?

apparently.
the specification has been public since 2000.


> PS: Do you know how this works on the 4500 platform?

as far as I am aware, cat4500 doesn't support WCCP except in software.
although I may be 100% wrong on this one.


cheers,

lincoln.