Skip site navigation (1)Skip section navigation (2)
Date:      Fri, 13 Sep 1996 01:47:39 -0400
From:      "Louis A. Mamakos" <louie@TransSys.COM>
To:        Terry Lambert <terry@lambert.org>
Cc:        karl@Mcs.Net (Karl Denninger), avalon@coombs.anu.edu.au, koshy@india.hp.com, freebsd-hackers@FreeBSD.org
Subject:   Re: SYN Resisting (fwd) 
Message-ID:  <199609130547.BAA11521@whizzo.transsys.com>
In-Reply-To: Your message of "Thu, 12 Sep 1996 11:26:05 PDT." <199609121826.LAA07267@phaeton.artisoft.com> 
References:  <199609121826.LAA07267@phaeton.artisoft.com> 

next in thread | previous in thread | raw e-mail | index | archive | help
> > > I heard someone mention 25% packet loss on an Australia-USA line (packets
> > > were being dropped in the USA).
> > 
> > With a 25% packet loss rate you've got bigger problems.

Yeah, you should price trans-Pacific DS3 circuits.  It's quite an
astonishing number of dollars per month.  And there are 6 digits
in the number.  

> Like you are probably ignoring source quench.

The routers on those links (or pretty much any these days) don't generate
source quench.  It's death to the fast-forwarding path in the router, plus
the packet being dropped may not be the one which is inducing congestion.
Source quench is pretty much depricated these days since it really hasn't
been demonstrated to actually help.

With modern TCP's these days, your best approach is to not take up additional
network capacity, but just drop the packet.  TCP with VJ slow-start will
back-off and Do The Right Thing.  The lack of the packet (or it's ACK)
will be the notification of the congestion event.  Sort of like a
virtual "DEC-bit"/"congestion experienced" bit in CLNS and FECN/BECN
in Frame Relay.

The real problem are applications using transport protocols which don't 
respect congestion events in the network, and reduce their offered load.
Things like the some internet telephony packages running over UDP, that
just Blast and Lose.  

Soon you'll see Random Early Drop (RED) implemented in backbone-class
routers which will drop packets in router output queues at the onset
of congestions.  They get chosen at random which does the right thing - if
you're misbehaving, then you'll have more packets stacked up in the
output queue, and be more likely to get nailed.  Grief is conserved,
and infliction of pain is a closed loop - cause pain (congestion) and
you experience pain (and lose).

louie





Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199609130547.BAA11521>