On Mon, Jan 29, 2001 at 10:49:49PM -0800, Hal Murray wrote:
> 
> I'm running my collection of network testing/bashing scritps.  Performance 
> is "strange" but I can't put my finger on it yet. 
> 
> I'm running 1.5 on Miatas (600au) and XP1000s.
> 
> Are there are any known problems/quirks?  Or parameters I should 
> tweak?  Anybody managed to get good results on this gear?
> 
> Any good/bad results with jumbo packets?
> 
> 
> The symptoms are that sometimes things are reasonable and then, an 
> hour later when I run the same script again, it doesn't work very 
> well.
> 
> For example, a TCP request-response test with short data blocks takes 
> either 200 microseconds or 10 miliseconds.  It works within reason 
> for blocks over 20K bytes - throughput is 230 megabits. 
> 
> I'm guessing there is some internal mode/state that I'm getting into 
> that's "bad".  Obviously, I don't know what it is yet.
> 
> 
> 
> One of my tests is a simple UDP-blast test that often overloads the 
> receive side.  I see things like this in the /var/log/messages.
> 
>   Jan 29 18:16:16 ruby /netbsd: WARNING: mclpool limit reached; increase NMBCLUSTERS
>   Jan 29 18:16:16 ruby /netbsd: ti0: cluster allocation failed -- packet dropped!
>   Jan 29 18:16:36 ruby last message repeated 867 times
>   Jan 29 18:16:36 ruby /netbsd: fpa0: can't alloc cluster
>   Jan 29 18:16:36 ruby /netbsd: ti0: cluster allocation failed -- packet dropped!
>   Jan 29 18:16:49 ruby last message repeated 800 times
>   Jan 29 18:17:09 ruby last message repeated 412 times
>   Jan 29 18:17:21 ruby /netbsd: WARNING: mclpool limit reached; increase NMBCLUSTERS

A yes, you need to increase NMBCLUSTERS
I think on an alpha you can increase it dynamically with
sysctl -w kern.mbuf.nmbclusters=x
You may want to try 4096 or 8192 for x.

I have a netgear in a DS20, it works pretty well.

--
Manuel Bouyer <bouyer@antioche.eu.org>
--