I think that is .2% - .3%, no? On Tue, Apr 2, 2013 at 5:41 PM, Joe Abley <jabley@hopcount.ca> wrote:
On 2013-04-02, at 18:18, John Kristoff <jtk@cymru.com> wrote:
I would expect from stubs this will be close enough to zero to be effectively zero. At least I would hope so.
This (below) is one of four resolvers, together providing service for two recursive DNS servers used by residential DSL and cable internet users at a medium-sized ISP in Canada. These resolvers are running unbound, which will never choose a source port of 53 on an outbound query; hence anything we see here with src port = dst port = 53 is one of those effective zeros.
[dns1-p1:~]% sudo tcpdump -i em0 -n -c 10000 -w sample.pcap udp port 53 tcpdump: listening on em0, link-type EN10MB (Ethernet), capture size 96 bytes 10000 packets captured 10267 packets received by filter 0 packets dropped by kernel [dns1-p1:~]% tcpdump -r sample.pcap -q udp src port 53 and udp dst port 53 | wc -l reading from file sample.pcap, link-type EN10MB (Ethernet) 26 [dns1-p1:~]%
26/1000 is more than zero but still quite small. Subsequent samples with bigger sizes give 332/100000, 3017/1000000.
No science here, but 2% - 3% is what it looks like, which is big enough to be a noticeable support cost for a medium-scale provider if the customer damage is not robo-mediated in some way (e.g. whitelist known offenders to avoid the support phone glowing red when you first turn it on).
Joe