On Mon, Apr 07, 2003 at 12:22:42AM -0500, Pat Wendorf wrote:
> Hey all,
> 
> I know this is a terribly slow reply, but thinking like a typical user, 
> "problem solved" means no need to reply.  I appologize for that.
> 
> Thanks for the help, the machine has been working flawlessly since I 
> recompiled my kernel with the nmbclusters increased.
> 
> However, I still find this a little strange:
> 
> I had kernel messages of buffering problems coming from the nic hardware (both 
> nic's actually).  This cleared up with the nmbclusters fix.  This makes 
> little sense to me.  How does increasing the nmbclusters value affect the 
> network hardware buffering? I thought this was only to increase some magic 
> kernel buffer that was used for buffering packets before they're sent to the 
> interface or some such.  

I don't remember the exact problem, but I don't know what you mean with
"hardware buffering". Most NICs DMA packets to host memory without CPU
intervention, and it is really host buffers, not hardware buffers.
On BSD systems, this memory is managed by mbufs, and usually is mbuf
clusters.

> 
> Additionally, if my machine was doing such a light duty (PPPoE Nat box for my 
> network), why would I get these errors?  Is 1024 too small to do this task? 
> Would any machine require at least 4096 nmbclusters to function as a simple 
> server?  If so, can this be increased for the default kernel, or would this 

It's not a "simple server", it's a router. This makes a difference, especially
if, as I guess it, you have a fast link on one side and slower one on the
other: the router may need to buffer a lot of data. If would probably
need less if both sides were running at the same speed.

> break many things?  How about dynamic allocation of nmbclusters using the 
> kernel memory allocate/deallocate features?

It would be less efficient.

-- 
Manuel Bouyer <bouyer@antioche.eu.org>
     NetBSD: 24 ans d'experience feront toujours la difference
--