To:
Pekka Savola <pekkas@netcore.fi>
Cc:
DNS Operations <dnsop@cafax.se>
From:
Jim Reid <Jim.Reid@nominum.com>
Date:
Wed, 23 Oct 2002 08:07:24 -0700
In-Reply-To:
Message from Pekka Savola <pekkas@netcore.fi> of "Wed, 23 Oct 2002 15:03:12 +0300." <Pine.LNX.4.44.0210231501230.31655-100000@netcore.fi>
Sender:
owner-dnsop@cafax.se
Subject:
Re: anycast
>>>>> "Pekka" == Pekka Savola <pekkas@netcore.fi> writes: >> Unlikely. There are very few situations where DNS latency >> *really* matters. The occassional lookup to a root server -- >> which is all a well behaved name server should need to do -- is >> not one of them. You seem to be overlooking the fact that name >> servers cache. The first lookup of some name could take 200ms >> (or more) but subsequent lookups of that name will be answered >> from the server's cache for the RR's TTL value which is >> probably measured in hours or days. Against that background, >> the sort of nano-optimisation you appear to be advocating is >> pointless. Why care about shaving 100ms or so off *one* DNS >> lookup out of thousands or millions of lookups? What is the >> justification for engineering an "optimal" solution for just >> that one lookup? Pekka> I believe the discussion was also about those servers Pekka> possibly having ccTLD and gTLD data. News to me. But even so, what I said *still* holds. It applies equally well to the TLD servers and zones as the DNS root. Pekka> Imagine a situation where you could get every xxx.yyy name Pekka> without going outside of your AS? This is how things were done on the old ARPAnet before DNS was invented. The HOSTS.TXT file did not scale. Have you any evidence to show how such a scheme -- whether implemented with name servers or not -- would work better today? Would anyone notice, let alone care about, any performance improvement? [For one lookup out of many thousands remember.] Oh, and I'd like to see your solution for getting the ~2Gb of .com or ~1Gb of .de propagated to name servers on every AS in a reasonably timely manner. Hey, and why stop at xxx.yyy name lookups per AS? Why not have every stub resolver host these zones itself? :-) I'm not convinced that even if we discount DNS caching, there is much to be gained wrt DNS latency by having a name server for "important" zones -- for some definition of important -- inside every AS. [It may be better for other reasons, like compartmentalising the crap from broken clients or making DDoS attacks against the infrastructure for these zones a lot harder.] Most important zones are served or should be served by name servers at good internet exchanges. So all you'd maybe save is the overhead of sending a packet out from your AS, across the Gigabyte Ethernet or whatever at the IX to the important name server and back again. This might even be a shorter and quicker path than routing the packets to/from the hypothetical LAN segment at the same IX where my AS has its critical name servers for those important zones. #---------------------------------------------------------------------- # To unsubscripbe, send a message to <dnsop-request@cafax.se>.