BIND 9.10 IPv6 performance

Carsten Strotmann cas at
Sun Mar 8 16:09:32 UTC 2015


I'm doing some performance tests on some modern Haswell CPU machines (20
cores) using Ubuntu Linux 14.04 (Kernel 3.13.0-46-generic) using BIND
9.10.1-P2 compiled with "--with-tuning=large".

With using 8 worker threads I get near 400K QPS via IPv4 UDP (from a hot
cache without resolving), which is a good. CPU utilization as seen by
"top" is near 800%, as expected (8 cores fully used).

When I switch BIND 9 to listen on IPv6 only, the performance drops to
less than 60K QPS.

When I run the tests using Unbound (same machine, same OS, 8 threads), I
do not see a significant difference between the IPv4 and IPv6
performance, which should rule out an issue in the kernel or with the
DNS load generation.

Testing with 9.9.6-P2 shows a similar pattern.

The configuration is simple:

options {
 directory "/var/named";
 listen-on { none; };
 listen-on-v6 { any; };
 recursive-clients 10000;
 tcp-clients 1000;
 allow-recursion { 2001:db8::/48; };

zone "." {
 type hint;
 file "root.hint";

Has anyone seen such an performance drop on IPv6?


More information about the bind-users mailing list