On Tue, Jan 30, 2001 at 02:46:53PM -0800, Hal Murray wrote:
> 
> > A yes, you need to increase NMBCLUSTERS
> 
> Thanks.
> 
> I'm already running with 4K.  These machines have a lot of network 
> cards so that may not be enough. 
> 
> The out-of-buffers messages only happen when I run the UDP blast 
> test.  I assume the buffers are all piling up on some to-be-dispatched 
> input queue because there aren't enough CPU cycles to keep up with 
> the sending side. 
> 
> Do all the queues have limit checks?  If I make NMBCLUSTERS big enough 
> will the messages (and associated troubles) go away?

I think I have 8k or 16k myself.

> 
> If not, making NMBCLUSTERS even bigger just means more buffers get 
> on some queue before the messages start happening.
> 
> 
> I might try commenting out the didn't-get-a-buffer message in each 
> driver in hopes that the system will spend fewer cycles printing 
> messages and thus have more to spend on processing buffers.
> 
> 
> > I have a netgear in a DS20, it works pretty well.
> 
> What sort of throughput can you get?  How hard have you hit it?

around 90MB/s output with ttcp -u.
Input was limited to something like 50MB/s because the other end was a
CPU-bound (celeron400).

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