[dns-wg] DNS hardware

On Aug 11, 2005, at 13:07, Bobby Ölander wrote:

> I'm working at a company called Com Hem AB in Sweden, we are at the  
> moment the biggest cabelnetwork company in Sweden.
>
> I have a question that concerns the hardware in the server that  
> should be a primary DNS (even the hardware on the secondary DNS  
> could be nice to know about).
>
> What I wonder most about is, should the primary DNS be build with a  
> RAID-system? Or is that not importent of the fact there is a  
> secondary DNS?

Bobby, it is not possible to answer your questions in detail. It's  
not clear what your company's requirements are or what sort of DNS  
service is needed: number of queries, authoritative or caching (or  
both!) name servers, frequency of zone file & configuration file  
changes, number of zones & RRs to serve, SLAs, peak and steady-state  
throughput, etc, etc. Without this information, all you can expect  
here is educated guesswork. Perhaps someone on the WG could work with  
you to document your requirements and devise a suitable DNS  
infrastructure. After that comes the hardware specification...

Unless you will be managing huge datasets -- eg hundreds of thousands  
of zones and/or millions of resource records -- a RAID solution would  
be overkiil. For DNS, the most significant advantage of RAID would be  
the facility to hot-swap broken disk drives. That would reduce the  
window where customers couldn't update their zone files. Most name  
servers store everything in RAM and will generally only load zone  
files from disk when they start. So having the DNS data "backed up"  
by a RAID system is unlikely to be worth it.  Unless of course fixing  
a broken drive and restoring from backup takes so long that it annoys  
your customers . Or embarrasses your management.

The DNS is designed to be resilient and can cope with server outages.  
So if one of the name servers for a zone dies, the others will  
seamlessly pick up the load. The only exception concerns the zone's  
master (primary) server. This is the *only* place where the contents  
of a zone can be changed. So if the master server dies, the zone  
can't be updated until it returns. The other servers for the zone  
will of course continue to answer queries for that zone. They'll  
automatically resynchronise with the master once it's back on-line too.

Memory provision is usually more important for a name server than the  
CPU or disk subsystem because name servers almost always store  
everything in RAM. The amount of RAM depends on the number of zones  
and resource records the name server loads. [A good rule of thumb for  
BIND is allow 1K of RAM for each zone and 100 bytes for each RR.] The  
number of queries the server gets can also be a factor. Big Iron is  
rarely if ever justified for DNS. Commodity hardware is often used  
though some people prefer low-end Sun/IBM/HP boxes because they  
neatly fill a 1U or 2U slot in a 19" rack.