Slashdot Mirror
DNS Root Servers Attacked
liquidat and others wrote in with the news that the DNS Root Servers were attacked overnight. It looks like the F, I, and M servers felt the attack and recovered, whereas G (US Department of Defense) and L (ICANN) did less well. Some new botnet flexing its muscle perhaps? AP coverage is here.
the root servers are setup in such a way that *2/3* of them can fail, and noone would notice.
[RFC2870]
2.3 At any time, each server MUST be able to handle a load of
requests for root data which is three times the measured peak of
such requests on the most loaded server in then current normal
conditions. This is usually expressed in requests per second.
This is intended to ensure continued operation of root services
should two thirds of the servers be taken out of operation,
whether by intent, accident, or malice.
http://en.wikipedia.org/wiki/DNS_root_zone
The root DNS servers are essential to the function of the Internet, as so many protocols use DNS, either directly or indirectly. They are potential points of failure for the entire Internet. For this reason, there are 13 named root servers worldwide. There are no more root servers because a single DNS reply can only be 512 bytes long; while it is possible to fit 15 root servers in a datagram of this size, the variable size of DNS packets makes it prudent to only have 13 root servers.
It's not like they haven't figured out the whole failover/fault tolerance thing.
That's kind of the point here, actually. Several of the root servers do not have any redundancy. You can see the list at http://www.root-servers.org/. In particular, the A, B, D, E, G, H, and L servers have only a single location a piece.
F, I, J, K, and M, on the other hand, are heavily redundant and have multiple geographic locations, routed via Anycast, so a single client only "sees" the server nearest to them. This makes them difficult to DDoS, because a zombie in S. Korea pinging the J server would be sending packets to the server in Seoul, while one in California would get the one in Mountain View.
What's odd, looking at the list, is that anyone operating something as critical to the internet infrastructure, wouldn't develop some geographic and systems redundancy; unfortunately, I suspect that the government agencies in particular tasked with these responsibilities probably don't keep it at the very top of their priority lists when allocating resources and funding.
"Ladies and gentlemen, my killbot features Lotus Notes and a machine gun. It is the finest available."
You can see the list of sites for F here:
http://www.isc.org/index.pl?/ops/f-root/sites.php
That's about 40 locations. Now, each of which has a couple of servers, a management box, and a couple of routers, so yeah something like 200 machines total.
Even nukes can't stop it! Or at least they shouldn't, since the internet was originally designed to run as a communications network in the event of a nuclear attack.
And the primary design feature that enabled that was removed during the rise of the ISPs.
The early internet was a NET. Redundant links everywhere. Routers all potentially knew the whole topology and could find a connection if it existed.
As the net went commercial that caused a table explosion in the routers. So BGP replaced RIP and things became less robust. Usable routes became a subset of all possible routes. Within the backbone there was still a lot of redundancy - but it wasn't quite up to the former "find a path if it exists" level.
Meanwhile, the typical host went from being something ad-hock connected to sever neighbors to being something connected solely to a single ISP - typically by a single link. The big guys might have redundant paths into their ISP's Network Operations Center. But if something took out the NOC (and often there was only one - or only one of some critical component) you were hosed. Ditto if something corrupted their databases. Even with redundant links there would only be a few, perhaps going through several single-points-of-failure - and if fully redundant still allowing a double-failure to take you down. The little guys would typically have one line (say DSL) to one box. Cut the line or crash the box - or the typically two links from it to the NOC - and you're hosed.
(Perhaps you have a dialup-backup for your DSL. Did YOU configure it to come up automagically if your main link goes down? Is it on the same phone line with the DSL? If not, does it take a different path to the central office? Or is it right up the same cable bundle on the same poles next to the same road full of the same drunk drivers or in the same underground cable running past the same backhoe...)
So the internet evolved from a nuclear-strike-survivable net to a less-robust net rooting a bunch of trees. Oops!
(And that's just for routing the packets once you've GOT the IP number. Translating names to IP numbers is a whole separate can of worms: It's what the root servers are about - which is why there are so many of them, most of them are clusters, and some are clusters that are geographically diverse. You only need to hit ONE operational root server to get started on your translation - if your answer isn't cached somewhere between you and the root, and the list is small enough to keep handy on every machine that wants to do its own nameservice.)
Bantam Dominique roosters crow a four-note song. Once you've heard it as "Happy BIRTHday" you can't NOT hear it that way