| From bouyer@antioche.lip6.fr Fri Jun 19 03:32:04 1998
|
| On Jun 18, Ken Wellsch wrote
| > I have the following scenario that has me frustrated.
| > 
| > I have a server running on my i386/NetBSD-1.3.2 system that accepts a
| > stream connection from a client process on a Mac (running MacOS 8.0).
| > The client can only process the incoming data from the server very slowly
| > (it is a Java applet under IE4) while the server has no problem generating
| > plenty of data for that connection.
| > 
| > Within what seems like a few seconds I've consumed all my mbuf space,
| > crippling my server system.  netstat claims I've got the expected 16Kb
| > in the queue.
| > 
| > The same server to another system, e.g. W95, W/NT, Linux does not have
| > this problem - the NetBSD seems to be properly blocking my writes.
| > 
| > The same server running under SunOS, Solaris and Redhat 5.0 does not
| > exhibit this behavior at all.
| > 
| > So can someone tell me what I need to do to fix this bad interaction
| > problem I'm having between my NetBSD box and MacOS 8.0?
| 
| Could you try to increase the number of MBUF clusters ? Just do this by
| recompiling a kernel with, e.g 
| options NMBCLUSTERS=1024
| in your kernel config file. The default is 256, or 512 if you have
| 'options GATEWAY' defined, but it's always too low for a server.

I've been using this

	options		NMBCLUSTERS=2048

for about a year now.  When the system started to get more use with
multiple folks and web serving and such, I had to do this to deal
with the routine variety of "mb_map full" messages... sigh.

I have watched via "netstat -m" as the percent used number grows
rapidly when I'm running my application.  I mean rapidly too.  I
ran "netstat" as quickly as I could and the %-used increased something
like 6% to start, 23%, 56%, 73% and that is when I killed the process.

It only occurs between a Mac and our i386 based server.

-- Ken