Re: caching at windmills (i mean exchange points)

[Duane Wessels <wessels () nlanr net>]

Ton Verschuren writes:

> I know. Still, for a cache to fill it has to fetch docs and thus act as 
> a parent. If the NLANR cache acts as a parent for all MW connected 
> ISP's than *all* misses from those ISP's get routed through the NLANR 
> cache and for the transatlantic traffic the lines of NLANR's ISP will 
> fill up. No?

In order for the cache to fill, yes some MW customers will need to call
the cache a parent.  Some could call the cache a sibling.

However, it isn't so black-and-white as that.  Squid can be configured
to measure RTTs to origin servers.  Squid's ICP replies will include
the RTT estimations, and the child cache can compare these values to
its own measurements, or those made by other (parent) caches.

Additionally, for the MW cache, we have modified Squid to return an
ICP_MISS_NOFETCH reply for requests origin servers which appear to be
unreachable (presumably due to incomplete routing info).  Upon
receiving this reply, the child cache will not forward the request to
us.

Thus, a child cache can use our MW cache as a parent only for origin
servers which are reachable by us, and which our cache is "closer" to.
If the child cache is closer, then it will forward the request
directly.

A potential problem is that the RTT measurement feature is not compiled
in by default.  It requires building Squid with a macro defined, and
installing the ICMP sending/receiving program with root privs.

A few more details can be found at
http://squid.nlanr.net/Squid/FAQ/FAQ-7.html#ss7.6

Duane W.
--
wessels () nlanr net                           Think Globally, Cache Locally.