[ On Wednesday, April 11, 2001 at 10:17:59 (+0900), itojun@itojun.org wrote: ]
> Subject: kern/12606: NMBCLUSTERS too small for nowadays use (netbsd-1-5)
>
> 	it seems to me that NMBCLUSTERS (256, or 512 if GATEWAY)
> 	is too small for nowadays applications.

indeed....

> 	my mail/web server choked this morning by web crawler with mbuf
> 	cluster shortage.  my upstream is just 128K!

My gateway box (a wee pentium 150MHz machine, but with 16MB RAM) chokes
sometimes too.  I've got a ~2mbit cable modem and a 3.0mbit aDSL modem
on 10mbit ethernet interfaces, and a 10mbit LAN interface too of course.

(my gateway's running 1.5F)

The symptoms are very bad -- nothing happens on any network interface.
and even worse nothing is printed on the console either.  I suspect one
interface at a time goes dead but it's hard to see for sure when your
real-time view (eg. tkined/scotty) is only through one of the interfaces.

Just as with your scenario a "netstat -m" doesn't show any more than 25%
mbufs in use.

If I login on the console and try to ping out I get a "No buffer space
available" error.  Still nothing from the kernel though.

If I have a telnet or rlogin session open to the machine it seems to be
receiving data OK as if I hit return a number of times the send queue
builds up for the session (i.e. on the frozen machine) and when I
recover (see next paragraph) the pending prompts are displayed.

I can usually cure it temporarily by logging in on the physical console,
downing all the interfaces, waiting a few minutes, bring the LAN
interface back up and wait another few minutes, and finally bring the
WAN interfaces back up.

However once it's been in this state once I can pretty much reproduce it
at will by trying to do a download from a reasonably fast site (eg.
where I get more than 50KB/s)

-- 
							Greg A. Woods

+1 416 218-0098      VE3TCP      <gwoods@acm.org>     <woods@robohack.ca>
Planix, Inc. <woods@planix.com>;   Secrets of the Weird <woods@weird.com>