[OpenAFS-devel] performance study

[Jeffrey Hutzelman]

On Tuesday, February 15, 2005 07:28:03 PM -0600 Troy Benjegerdes 
<hozer@hozed.org> wrote:

> On Mon, Feb 14, 2005 at 07:33:20PM -0500, Ken Hornstein wrote:
>> > I was just commenting to a colleague that it would be nice to have
>> > something like Tim Shepard's TCP packet trace plots for RX.  (See
>> > xplot.org.)
>>
>> I thought of writing something like that, but since my ultimate goal is
>> to write a TCP transport for RX, I decided that I didn't see much point.
>> But hey, if someone else wants to write one, go for it :-)
>>
>
> What about an SCTP transport for RX?
>
> It seems a lot better suited for file transfer, since it's a reliable
> message-oriented transport, and you can have multiple streams controlled
> by a single congestion control algorithm.
>
> It also seems to have significantly better behavior under heavy
> packet-loss conditions (see a paper I did for a class at
> http://www.scl.ameslab.gov/Personnel/Publications/Troy/wireless-sctp/ )

I have long believed that SCTP is probably a pretty good match for Rx. 
However, doing it right will be harder than for TCP, and thus will likely 
take longer.  The decision we made at the hackathon in December was to work 
on TCP first, because it was expected it would take less time, and because 
the people actually doing any work were interested in TCP.  I believe the 
protocol design we came up with is sound, and I know some folks are working 
on implementing it.  Of couse, the fileserver will end up with lots of TCP 
connections, but it will have the option of throwing idle ones away, just 
as it would for Rx-over-UDP connections.

I would be quite interested in seeing and commenting on an Rx-over-SCTP 
design, especially if someone is willing to actually do an implementation.

-- Jeffrey T. Hutzelman (N3NHS) <jhutz+@cmu.edu>
   Sr. Research Systems Programmer
   School of Computer Science - Research Computing Facility
   Carnegie Mellon University - Pittsburgh, PA