Where is an IP Anycast Nameserver system implemented?

Question

Ive been reading alot about nameserver the last days. For our websites we want to optimize the waiting time of the visitors that is caused by our namserver. I will have some questions about IP Anycast and the general function of the DNS. Let me start by explaining what I understood the DNS works from user side:

User X wants to visit www.example.com, the following steps happen to get the IP address:

1.Step: User X sends request to the Nameserver of his ISP or nameserver by choice.(recursive nameserver)

2.Step: If the adress is not found, the recursive nameserver will send a request to one of the 13 root nameserver to get the nameserver for the .com TLD

3.Step: Query the .com nameserver to get the auhorative nameserver

4.Step: Query the auhorative nameserver to get the ip-address for www.example.com

First I realized that as a owner of a website you can only optimize Step number 4 and all other steps are not in our hands.

I came across IP Anycast nameserver (what is also used for the 13root nameservers) and totally understand the concept of distributed machines. But what I dont understand is where the decision logic, to which of the distributed machines the user will be send, according to his "position",is implemented? I mean when i buy an anycast nameserver, the logic should be implemented on the .com nameserver (Step 3), so that this nameserver decides to which machine of my anycast nameserver the user will be send.

For me thats really hard to understand and im asking myself if it really works that way? I hope someone can help me with these understanding questions.

Beside of that i found out, that another small method to gain some speed for the user, is to only use A Records and no CName Records anymore.

Are there some more ways to optimize a nameserver?

Thanks in advance!

Answer 1

Short answer is: you are right. The NameServers is where you can optimize and all "IP Anycast" products I have seen is just a NameServer setup that has a lot of locations.

They use the same system as the "root servers of the internet" but this does not mean that they have the same function. The IP Anycast is simply a method for multiple servers in different locations to serve the same IP address.

From WIKIPEDIA (http://en.wikipedia.org/wiki/Anycast) On the Internet, anycast is usually implemented by using Border Gateway Protocol to simultaneously announce the same destination IP address range from many different places on the Internet. This results in packets addressed to destination addresses in this range being routed to the "nearest" point on the net announcing the given destination IP address.

If you are using a big ISP like ASCIO or someone using ULTRADNS you probably do not have to worry about this step too much, but if the NS is a local ISP it is worth considering. Make sure you have NS where your visitors are.

I assume this is where you came into contact with "IP Anycast" products. None that I have seen offers anything to attack step 1-2-3 but rather offers a large setup of NameServers allowing them to reduce resolving time due to closeness of networks.

Let me know if you are of the understanding that the offer is for a root NameServer setup, because I would like to see this.

Answer 2

Your question is not really related to programming, but more to operations, and is also a little too vague ("Are there some more ways to optimize a nameserver?").

But let us try to give you pointers.

User X wants to visit www.example.com, the following steps happen to get the IP address:

Your following description is then mostly correct. Note that at each step, by default until very recently, a recursive nameserver will send the whole name queried to each nameserver. Recently, QNAME Minimization appeared as a standard and now recursive nameservers can send to each authoritativ nameservers only the labels it neeeds to reply. This enhances privacy without changes to the protocol, but is not widespread today because some authoritative nameservers do not work correctly when queried that way.

As a domain name owner you can indeed only have an impact at the last step. But remember that recursive nameservers have caches, so the list of root nameservers as well as the list of .COM nameservers for example are so "hot" (so often needed) that they surely sit always in resolvers' caches, so basically step 1 and 2 happens do not happen often (at start when cache is empty typicaly).

I came across IP Anycast nameserver (what is also used for the 13root nameservers) and totally understand the concept of distributed machines. But what I dont understand is where the decision logic, to which of the distributed machines the user will be send, according to his "position",is implemented?

First things first, IP anycasting is not specific to the DNS, it is just hugely popular here because

1. it solves the load balancing/fail over problem that all big TLDs have
2. it works specially well with DNS over UDP which is a simple one query one reply protocol.

So any service can theoretically be anycasted. It means that a given IP address just appears at different locations in the world.

To summarize very broadly, Internet traffic between providers (AS numbers) is exchanged at peering points, where they interconnect and each provider says "I know about IP block 192.0.2.0/24, please send me all traffic for it", etc. for each blocks (again this is a summary. Blocks are allocated by RIRs, and yes by default this is not very much authenticated so BGP hijacks happen when another provider also says "give me this traffic" when it shouldn't - and it happens because of malicious goals or just simple human errors).

For a normal (technical term: "unicast") IP address, only a single provider (AS) will announce it somewhere (technically: announce its block not just a single IP) and everything will be configured in such a way that wherever the start of the exchange is, for this single IP as destination, it will arrive at the exact same box.

On the contrary, for an anycast IP address, either a single provider or multiple ones (that is multiple Autonomous Systems) will announce this IP at various locations (peering points) in the globe. At each peering point, traffic for this IP will get taken by the provider announcing it there and then it will route this traffic to a specific server "nearby". Announcements of the same IP at peering point A and peering point B, will drive corresponding traffic on one side in datacentre X and the other for datacentre Y.

For the client, when everything works, it does not change anything, as long as all the replying servers react the same way to the same query. The client does not (and sometimes can not) even know the IP is anycasted or that it want to location X when another client doing the same thing will instead hit location Y.

So in short nameservers "decide" nothing in this regard. At each point of the DNS resolution, when they need to contact nameserver NS1 they know its IP address is IP1 and they just open an UDP (or sometimes TCP) connection to this IP, absolutely normally. It is the underlying IP and BGP protocols that will, if anycast is in action, make the response come from the appropriate "close" server.

Note that anycasting in this way, for DNS, achieve both:

1. fail-over : if one server dies, with appropriate monitoring, its provider withdraws its IP announcement, that is this local copy kind of disappear and the traffic will automatically (in order of seconds) shift to any other instance where the same IP is announced

2. load-balancing : rougly speaking, if you anycast one IP on 2 locations, each should receive 50% of traffic. It is not true in practice, and is very complicated (read: impossible) to predict or even monitor, because it all depends on the peering points, the agreements between the providers and various other policies (simple example: if you peer at two points where on first there is only one provider sending you trafic, and at other point you have 100 providers with whom you exchange traffic, then you may get more connections going to the second instance... except of course if single provider at first peering point is an ISP with millions of clients, where the other 100 providers are single small organizations...)

So, some nameservers are anycasted. Nowadays all the root ones are (but this was not true 16 months ago, see https://b.root-servers.org/news/2017/04/17/anycast.html as b.root-servers.org was the last one to board the anycast wagon) as well as all big TLDs, sometimes even with more that one "Anycast DNS providers".

For any domain name, you can get some providers that will give you a DNS service for it, based on a "cloud" of anycasted nameservers.

See for example:

• https://www.pch.net/services/dns_anycast
• https://www.netnod.se/dns/netnod-dns-services
• https://dyn.com/dns/network-map/
• http://www.cdns.net/anycast.html
• https://www.rcodezero.at/en/home/
• https://aws.amazon.com/route53/
• https://cloud.google.com/dns/
• and many others.

Now following on a totally different topic:

Beside of that i found out, that another small method to gain some speed for the user, is to only use A Records and no CName Records anymore.

This is not really something you gain things with, and CNAME records are useful in many other cases.

Again, you need to remember that there are caches. So even if your configuration is:

www.mywebsite.example CNAME www.mywebsite.example.somegreatCDN.example
www.mywebsite.example.somegreatCDN.example A 192.0.2.128

it is true that this means on paper two DNS requests to finally be able to do an HTTP query, but in practice things will be cached (even more so today with big public open resolvers such as 1.1.1.1 or 8.8.8.8 or 9.9.9.9, that are anycasted too in fact), so the difference will be negligible (and only impacts the first time, never again until it is in cache) ... especially in the case of HTTP and everything that happens later that is opening, frequently, dozens of connections to download javscript source codes, CSS files, fonts, etc. that may be hosted elsewhere.

A lot of websites use CNAME records without negative impact. See www.amazon.com for example, right now:

;; ANSWER SECTION:
www.amazon.com.     730 IN CNAME www.cdn.amazon.com.
www.cdn.amazon.com. 11 IN CNAME d3ag4hukkh62yn.cloudfront.net.
d3ag4hukkh62yn.cloudfront.net. 11 IN A 54.239.172.122

You may however argue that some names will be more popular than others and hence stay longer in cache, which is certainly the case.

And finally:

Are there some more ways to optimize a nameserver?

Based on what? We touched various subjects above, all are compromises, you sacrifice something (it may be just "money") to gain something else (redundancy, etc.). There is no generic rule to declare when this compromise makes sense or not, it will depend a lot on your situation and what you are trying to do.

You are right, and should be congratulated about that, that you should invest some time around DNS setup, both for security and performance reasons. While a lot of money is often invested in huge HTTP setup to sustain various problems or spikes of activity (but even the best fail sometimes, see the recent Amazon Prime Day opening that was a gigantic failure), but often people forget about the DNS because it is on the infrastructure level so not well known nor understood (using UDP makes it already stand out from all other known protocols, as this is rare).

For example there is another completely different thing (it is orthogonal to anycasting, so it can work with or without it, the goals are different) that is related: "geo-DNS" means when a nameserver will reply differently based from where the client asks. This is meant to give, for example, a different IP for a webserver, one that is closer to client (so in that case the webserver itself is probably not anycasted). This is done by just looking at the source IP from the DNS packet, but it is often not good enough because the authoritative nameservers only see as source IP the one from the recursive nameserver and not the real end client one and nowadays with big open public recursive nameservers the location should be far off, so you also have a specific DNS option called EDNS Client Subnet that can be passed between recursive and authoritative nameservers so that they get the end client real IP address (in fact a block not a single IP for privacy reasons) and can act upon it.