[OpenAFS-devel] fileserver parameters

[Troy Benjegerdes]

On Sun, Jun 19, 2005 at 11:45:36AM -0400, Ken Hornstein wrote:
> >All of this is rather theoretical, and I only have a little bit of
> >code thus far.  I hope to find the time to make some of these changes
> >happen, but these are ambitious suggestions.  I can't promise
> >anything, unless other people with lots of time volunteer... ;) 
> >Comments?  Criticisms?  Volunteers?
> 
> My boss has tasked me with improving AFS performance as well, but I've
> taken a completely different direction.  Specifically, I've been working
> on a TCP transport for Rx.  We're specifically interested in using AFS
> on long-haul networks, and my thinking was that instead of investing all
> of the time in fixing up the Rx protocol, why not leverage all of the
> work that's been done on TCP performance in the last decade?
> 
> So far I have a rough implementation working.  It's performance is
> good, but it's not ready to use by AFS applications just yet.  Clearly
> working out all of the thread interactions is WAY more complicated
> than I thought it would be, and you've clearly thought more about
> this than I have.  It sounds to me like we should try to coordinate
> our work, and have both people's work proceed simultaneously.

Do you wind up serializing multiple calls into 1 TCP stream? How hard
would adapting this work to SCTP be?

Based on some tests I did on 802.11 wirless awhile back, TCP had better
throughput with zero packet loss, but SCTP did a lot better with 5-10%
packet loss rates. And when you get to TCP flows of a gigabit or more over
long-haul links, packet loss of even something like 0.01% can really
kill performance.