full socket buffers

[Miroslav Lauš]

On Dec 19, 2010, at 8:32 AM, Miroslav Lauš wrote:

> On Dec 18, 2010, at 10:53 PM, sthaug at nethelp.no wrote:
> 
>>>> We're also running 4.1.1p1, on FreeBSD 7.3-STABLE. Server hardware is
>>>> Dell PE 1850, with two SCSI disks in a RAID-1 configuration. We have
>>>> around 100k leases on two servers in a failover pair, with most of the
>>>> pools on 24 hour leasetime. We don't see the problem you're seeing.
>>> 
>>> We're running the dhcpd service on two IBM x3550 servers in failover,
>>> both have two 74GB 15k disks in RAID-1 so very similar to yours.
>> 
>> How many DISCOVERs and REQUESTs do you handle per second? Do you have
>> battery backed cache for your RAID-1 config? Have you tried gstat, and
>> looked at the %busy column?
>> 
>> An additional point which might be relevant - we run our failover
>> configuration with a "delayed-ack 28" setting. This cuts down on the
>> failover traffic (including disk I/O). Yes, we know there have been
>> some warnings about delayed-ack - however it has never given us any
>> problems. Note that it needs to be explicitly compiled in.
> 
> See DISCOVER/REQUEST ratio in attached RRD graph. I've also attached
> the udpInErrors graph which is currently in normal (very few dropped udp
> packets due to full socket buffer) and I'm now waiting for it to increase.
> 
> I'm not sure about the battery cache of the RAID, but the gstat never goes
> higher than 25% busy. If we were disk i/o bound wouldn't the issue persist
> all the time? I will check the delayed-ack as we have anyway selfcompiled
> the dhcpd package with DHCPRELEASE mechanism disabled as it caused us a lot
> of trouble in conjuction with dhcpsnooping mechanism on L2 switches.
> 
> Would putting the leasefile on RAM disk help a bit?
> 
> -ml

I've tried with ramdisk but have a small problem with it - the ramdisk is 1GB in
size but after about six hours the dhcpd started to complain there's no space
to commit the leases and crashed a few moments later. How big should the ramdisk
be in size to hold the leasefile? Is the dhcpd process doing some leasefile
compaction in time or just the leasefile grows boundless?

-ml