tuning for maximum dhcp performance

[Brad Dameron]

Speaking of performance. We have some very busy DHCP servers and have
come to find a couple tuning items that have seem to double our
performance.

The biggest one being adjusting your logging facility to not write each
line individually but to bulk write. In the syslog.conf file you should
enter a - in front of the log file path. This tells it to cache and
write several lines at a time. Which helps alleviate disk thrashing. 

Example:

Local6.*					-/var/log/dhcpd.log

Or if you are using syslog-ng set the sync option to 20. 

The next option we adjust is a setting in the code that requires a
re-compile. It appears the hashing algorithm isn't very well tuned for
larger instances that have to be restarted. We went from a 45 second
restart down to under 5 seconds with this little change.

In the includes/dhcpd.h file find LEASE_HASH_SIZE. Change it from 100003
to 250007. 

Hope this helps.

Brad

-----Original Message-----
From: dhcp-users-bounce at isc.org [mailto:dhcp-users-bounce at isc.org] On
Behalf Of Gordon A. Lang
Sent: Tuesday, May 06, 2008 9:13 AM
To: dhcp-users at isc.org
Subject: Re: tuning for maximum dhcp performance

Does anyone have performance data using ZFS on Solaris?

--
Gordon A. Lang

----- Original Message ----- 
From: "Dan" <dan at telcohero.com>
To: <dhcp-users at isc.org>
Sent: Thursday, May 01, 2008 4:02 PM
Subject: Re: tuning for maximum dhcp performance


>
> Blake,
>
> Write-back was on, I confirmed.  Even so, I was still unable to get 
> numbers approaching the ramdisk until I started trying different 
> filesystems.
>
> It seems that when starting dhcp with a blank dhcpd.leases file, using

> reiserfs filesystem, and dhcperf, you get poor performance.  JFS, XFS,
and 
> even EXT3 (in descending order) all resulted in over 300 5way 
> clients/second.
>
> All filesystems were mounted with noatime, nodiratime, nodev, nosuid,
and 
> noexec.
>
> I was a bit surprised that the difference was this drastic.
>
>
> On Fri, 25 Apr 2008, Blake Hudson wrote:
>
[...omitted...]