On Wed, 2007-07-25 at 10:54 +1000, James Gregory wrote:
> On Tue, 2007-07-24 at 13:37 +1000, Gavin Carr wrote:
> > Hey sluggers,
> >
> > Anyone have any pointers to open source projects (or features of projects)
> > around WAN link optimisation? I'm specifically looking for a way of
> > duplicating traffic across multiple links to avoid resends on high latency
> > links, but I'm interested in the whole area.
> Don't have references handy, but I expect that you would be able to
> improve performance substantially by tinkering some values in proc.
So following the interest this thread has received, I dug up some
references. Some time ago, Evgeniy Polyakov posted a link to some neat
graphs detailing the effects of different retransmit ("resend" as you
call it) algorithms. These graphs:
http://developer.osdl.org/shemminger/tcp/net-2.6.22/
Are the most illuminating, and some of Stephen Hemminger's presentation
on the subject detail other network performance stuff. Take a look
around:
http://developer.osdl.org/shemminger/
Now, I realise that this does not directly address your interest in
duplicating TCP traffic to mitigate the effect of per-link packet-loss,
but it does demonstrate that the cost of that packet-loss can be
substantially reduced by small changes to the behaviour of TCP. Huge
windows will get you the biggest benefit on high-latency links, as, if
they are huge enough, you can keep the pipeline full enough to hide the
latency of receiving the duplicate ACKs you require to trigger a
retransmit.
I'm sure there's a bunch of commercial stuff you can use, like that
'Fast TCP' thing (which actually sounds pretty cool), but if you want
open-source, I suspect the biggest win you'll see is in just shuffling a
few parameters in your kernel's network stack.
HTH,
James.
--
James Gregory -- http://codelore.com -- [EMAIL PROTECTED]
--
SLUG - Sydney Linux User's Group Mailing List - http://slug.org.au/
Subscription info and FAQs: http://slug.org.au/faq/mailinglists.html
