Re: [Linux-decnet-user] FYI: linux-decnet performance

[Steven W. <st...@gw...>]

Hi,

> 
> On Fri, May 23, 2003 at 11:55:38AM -0700, Larry Baker wrote:
> >    I ran some crude timing tests on our network to compare the speed of
> >    OpenVMS Backup using several combinations, including linux-decnet.
> >     
> 
> Those are good numbers to see :-) Are you using Phase V or Phase IV DECnet onthe
> VMS side? ISTR there is some extra flow-control going on in the Phase V (in
> Phase IV emulation mode) that makes it faster than "real" phase IV. You can
> tweak with by putting window sizes into
> 
> /proc/sys/net/decnet/no_fc_max_cwnd
> 
> so it might go even faster...
> -- 
> 
> patrick
>
Well DECnet only takes notice of this setting in the case that there is
no flow control in use, otherwise the flow control sets the maximum
congestion window.

Currently Linux DECnet only supports flow control on the outgoing side of
connections (as required by the spec) and not for the incoming side
(since most Linux DECnet boxes are much faster than the machines they
are talking to) and since thats optional.

It certainly would be interesting to see what makes the performance so good
in this case. DECnet doesn't do nagle (yet) for example but its on the list
of things to do. Also I very much suspect that as soon as there is any
packet loss on the link, DECnet will perform much worse than IP since the
IP stack has a whole bunch of tricks up its sleeve to keep data moving in
cases of mild packet loss.

Next on my list is to sort out dn_recvmsg() and dn_sendmsg() to add support
for nagle and zerocopy. This will require a change in the way that packets
are queued for transmission. Also on the receive side, I need to make sure
that I've understood the requirements for MSG_PEEK correctly - its not as
simple as it looks at first glance,

Steve.