bind-9.3.2 / CPU issue.

[Stefan Schmidt]

On Mon, Aug 21, 2006 at 09:14:39PM +0900, JINMEI Tatuya / ?$B?@L at C#:H wrote:
> > I have four of these boxes running as caching nameservers for my customers. I get about 600 queries/second on each machine. After x amount of time my cpu usage peeks to 100% 
> > resulting in _real_ slow responses, the problem doesn't go away from itself, I need to restart named to get this fixed. All four machines have this issue. I tried several upgrades,
> > tuning config options like max-cache-size, max-ncache-ttl ... still no result. I do use an ACL to restrict users based on ip, tried to disable that to for a while, but still with no result. 
> 
> > Is there someone with the same issue ? Or some good tips ? :)

I switched to PowerDNS recursor [1][2] lately and didn't regret it.

...
> Also, there are some common known optimizations to work around this
> problem:
> 
> 1. build named with its internal memory allocator:
>    ./configure STD_CDEFINES='-DISC_MEM_USE_INTERNAL_MALLOC=1'

It was about a year ago that Tatuya shared that information with me on
the bind-workers ML. He also told me that disabling threads might help.
I did that, ran two processes per machine and it helped me a great deal.
Back then i also made up a short webpage [3] on the issue but never showed
it to a broad audience - if you like graphs as much as me have a look. ;)

My 9.3.2 instances were coping with about double the queries/s that you
have each without problems on 2.8GHz machines (HP DL360 G3) with just
1GB RAM by the way. No slow-responses on my smokeping at least.

> 2. upgrade to BIND 9.4.  This version now enables the internal
>    allocator by default, and automatically adjusts an amount of
>    cleaning task that blocks incoming query processing:

Worked great for me when singlethreaded, just don't try to ulimit its
memory usage - backfires with SERVFAIL then. (RT #16333)

	Stefan

[1] http://doc.powerdns.com/built-in-recursor.html
[2] http://svn.powerdns.com/snapshots/pdns-recursor-3.1.3-pre2.tar.bz2
[3] http://zaphods.net/~zaphodb/high-performance-bind9.html
-- 
A component selected at random from a group having 99% reliability, will be a
member of the 1% group.