I was doing some benchmarking with ab and noticed these in my syslog:

Jan 14 16:20:42 vps1 TCP: time wait bucket table overflow
Jan 14 16:20:42 vps1 TCP: time wait bucket table overflow
Jan 14 16:20:42 vps1 TCP: time wait bucket table overflow
Jan 14 16:20:42 vps1 TCP: time wait bucket table overflow
Jan 14 16:20:43 vps1 printk: 1060 messages suppressed.

Google turned up some info on this:

http://bugzilla.openvz.org/show_bug.cgi?id=460


But I've already got that parameter turned up: 

# cat /proc/sys/net/ipv4/tcp_max_tw_buckets_ub 
32768

Is there somewhere else I should look? 

Note that the benchmarking I was doing was actually in the host OS, not
a VE.

Regards,
Cliff

On Wed, 2008-01-16 at 09:35 +0300, vaverin wrote:
> 
> Hi Cliff,
> 
> I think network stress testing can create much more time wait buckets.
> It is information message and not an error therefore IMHO it is not a
> problem for you.

I was concerned because I was seeing odd results during testing (the ab
test was running 500-1000 concurrent connections).  Most of the
connections finished in <2ms but a few would take almost as long as the
entire test (15s).  This may be an ab bug however.
> 
> Do you want to allow it? Try to increase 
> /proc/sys/net/ipv4/tcp_max_tw_buckets (global limit)
> and
> /proc/sys/net/ipv4/tcp_max_tw_buckets_ub (per-VE limit, but host
> system itself accounts like ususal VE)
> values.
> 
Yes, I tried this.  Apparently I did not go high enough.  

Thanks for the help.

Regards,
Cliff