Subject: Re: NFS/RPC and server clusters
To: None <tech-net@netbsd.org>
From: William Allen Simpson <wsimpson@greendragon.com>
List: tech-net
Date: 10/16/2003 07:44:50
David Laight wrote:
> 
> > > If a system sends a large TCP message it will go out as back to back
> > > ethernet frames.
> >
> > To a max of 4 segments, assuming following RFC recommendations.
> 
> Why? a full window of TCP data will be sent (at least once the window
> has opened).  The ethernet card will just shovel them out.
> 
The algorithms have been floating around and well documented for years.
I'd assumed NetBSD was up-to-date.  Most recently, RFC-2581 and RFC-3390 
(just checked).

And for RPC/UDP, folks should have read RFC-3448.


> Ethernet cards do use more complicated backoff algorithms than when I
> first saw that traffic pattern on a LAN, but I don't know if they go as
> far as deliberately generating collisions because another system seems
> to be hogging the network.  Some store and forward switches/router
> do so when there buffer space is getting full.
> 
Of course, that's a good idea, too -- but I've only seen it on high-end 
giga-switches.  However, not what I was talking about, since that's 
beyond the scope of NetBSD itself.
-- 
William Allen Simpson
    Key fingerprint =  17 40 5E 67 15 6F 31 26  DD 0D B9 9B 6A 15 2C 32