Coda File System

Re: IP fragmentation

From: Jan Harkes <jaharkes_at_cs.cmu.edu>
Date: Wed, 3 Nov 1999 17:06:13 -0500
On Tue, Nov 02, 1999 at 12:13:00PM -0500, Greg Troxel wrote:
> I've been having problems with a coda client that runs something like
> mobile IP and is on a wavelan.  When the client lists a directory, the
> server sends a 2208 byte packet.  This is fragmented at the server.
> Both fragments travel to the home agent, and the first fragment is
> further fragmented to fit in the tunnel.  Then, with multi-hop wavelan
> forwarding, one (the second chunk) is almost always lost.  Adding
> IPSEC and PPP links with <1500 byte MTUs makes this worse.
> While my network should perhaps be better behaved, I'd like to avoid
> IP fragmentation entirely most of the time.
> 
> Is there some way to configure rpc2 to only send packets with 1K of
> data?  The "-se" part seems to break things up like this.
> 
> I'm running code from anoncvs from the last few weeks.  I've been
> impressed at how stable this is compared to the code from a year ago!
> 
>         Greg Troxel <gdt_at_ir.bbn.com>
> 

Hi Greg,

The -se part is the SFTP sideeffect, which does a windowed streaming
data transfer. Reducing (or enlarging) the packet size of that is easy.

Breaking up the rpc2 request/reply packets is a lot harder and it
probably won't be much of an improvement over IP level fragmentation.
I.e. we probably still have to retransmit the whole request or reply
message anyway, except when the reliable send/receive logic in rpc2 is
rewritten.

In which case it is probably almost as much work as implementing an
rpc2-over-tcp implementation, which would give additional benefits like
more easily 'punching' through (masquerading) firewalls, and leveraging
of tcp improvements which are resulting from experience and research on
web traffic and ad-hoc networking.

Jan
Received on 1999-11-03 17:09:47