request drops in BIND?

[Barry Margolin]

In article <c82ogd$7bf$1 at sf1.isc.org>,
 KyoungSoo Park <kyoungso at cs.princeton.edu> wrote:

> One question. Does BIND use the same udp socket to handle remote server 
> queries?

Only if you configure "query-source * port 53".  Without that, it uses a 
different source port for recursive queries, which requires using a 
different socket.

Someone mentioned the fact that the root servers handle thousands of 
queries per second.  It should of course be noted that root servers 
don't have to do recursion, so it's not totally fair to compare them 
with typical caching nameservers.

When I was at Genuity/Level(3) we occasionally had some severe 
performance problems on our caching servers, which were Sun workstations 
with lots of RAM.  The evidence always pointed to bottlenecks in the UDP 
stack, not BIND.  We noticed that queries to the publicized address was 
timing out, but queries to another address on the same physical 
interface responded quickly, so it was obviously a per-socket or 
per-logical-interface buffer that was overflowing.  We alleviated the 
problem by assigning another virtual address to the server, and 
configuring the DHCP server that serves the DSL customers to use that.  
This way, the queries from DSL customers would not interfere with 
queries from the higher-paying T1/T3 customers.

-- 
Barry Margolin, barmar at alum.mit.edu
Arlington, MA
*** PLEASE post questions in newsgroups, not directly to me ***