You know, with CPUs as blindingly fast as they are today, it wouldn't surprise me if you couldn't sustain 10,000 + ICP lookups a second with the right amount of coding work. DNS servers are able to push half a million DNS request/replies on current hardware. I'm pretty certain cache lookups can be distributed and handled pretty effectively. Of course the ICP dropped messages problem also occurs, but its certainly not out of the realm of possibility to do something like batching key lookups over TCP, like with HTCP? Adrian On Tue, Feb 05, 2008, Chris Woodfield wrote: > Yes, we would us using cache digests to keep the ICP/multicast traffic > from becoming a bottleneck. Another benefit of cache digests is that > there will be a certain level of content duplication - if server A > caches an object, and server B gets a request for the same object > before if refreshes A's digest, the object will live on both servers, > which helps with the flash-crowd issue. > > Then the question becomes this - if a server gets a request for an > object and has multiple cache digests that contain it, will the > decision use normal peer selection metrics? > > -C > > On Feb 5, 2008, at 6:57 AM, Amos Jeffries wrote: > > >Chris Woodfield wrote: > >>Hi all, > >>I'm facing an issue where we'd like to implement cache peering on > >>our squid farms, primarily to leverage the combined disk capacity > >>across all our boxes into a larger cache. I would presume that this > >>requires the use of the proxy-only directive to avoid content > >>duplication. However, this has raised the issue of server overload > >>- there's a very real possibility that a single hot object, if it > >>only lives on a single server, could overload that server with > >>requests from all the other peers in the event of a flash crowd. > >>I'd like to find some middle ground. > >>I'm wondering if there's a way to configure squid such that content > >>retrieved from peers is cached locally, but for a much shorter > >>period of time than content the cache retrieves directly. Is this > >>possible within squid, or should this be a feature request? > > > >Squid contains a few methods for reduction of ICP requests. > >Primarily the use of Cache Digests and to a lesser degree multicast- > >ICP. > > > >http://wiki.squid-cache.org/SquidFaq/CacheDigests > > > >I think you possibly may be able to hack some ACLs that prevent disk > >storage but leave items in the hot-memory cache. Though I have never > >looked at that myself. > > > >Amos > >-- > >Please use Squid 2.6STABLE17+ or 3.0STABLE1+ > >There are serious security advisories out on all earlier releases. > > -- - Xenion - http://www.xenion.com.au/ - VPS Hosting - Commercial Squid Support - - $25/pm entry-level VPSes w/ capped bandwidth charges available in WA -
