Random thoughts:
	- if you leave a ping running during this do you notice any
	  change in latency during the problem
	- Does 'netstat -m' show any patterns during the problem
	- Do you have another network card you can test temporarily?

                David/absolute
			       -- www.netbsd.org: A pmap for every occasion --


On Mon, 4 Sep 2000, Marc wrote:

> 
> I've built two machines so far with the 1.5 BETA code and I've noticed
> a strange problem that keeps cropping up in my 1.5/RAIDFrame system.
> (Celeron, Asus + onboard i82557, 256 MB RAM, Two identical 10 GB
> fujitsus.)
> 
> Specifically, under certain loads (75 concurrent connections totalling
> 300K/sec on 100 Mbit line) the transfers become very "bursty". High
> troughput for a moment, then a pause, then repeat. I've examined the
> traffic and there is no packet loss.
> 
> I've been combing through the mailing list archives--and I've tried
> tuning various sysctl variables such as tcp.recvblah and
> tcp.sendblah to as much as 65K and as low as 16K, and there is no
> difference.
> 
> I've tried to look for problems with this type of network
> card--nothing so far except that the card is difficult to program for.
> 
> I've increased NMBCLUSTERS to 8196 via gdb --write /netbsd, etc.
> 
> I've modified my ftp daemon (proftpd in my case) to use large, then
> small receive and send buffers.
> 
> I've turned on, then off the Nagle algorithms.
> 
> I've turned on, then HURRIEDLY turned off net.inet.tcp.cwm.
> 
> I've turned on, then off newreno.
> 
> I've tried various combinations..
> 
> When I first built the machine it was running at a good clip--well
> over 8 MB/sec downloads to a single client, and a speedy, handy
> response time.
> 
> Now that it's under a light load, though, performance is just a
> little.. off? CPU is sitting at 98% idle.. lots of free memory..
> 
> Can you think of any suggestions for me? Shall I convert back to a
> primary drive that's just mirrored every night to a second drive in
> case of failure?
> 
> Thanks,
> 
> Marc
> 
>