[clug] Emitting Source Quench When Queues Are Full?
Alex Satrapa
grail at goldweb.com.au
Wed Oct 1 20:56:42 EST 2003
On Wednesday, October 1, 2003, at 09:50 , Matthew Hawkins wrote:
> Why not configure your QoS stuff (with tc, etc) to shape your bandwidth
> the way you need it? ICMP source quench is a little pointless when some
> old legacy systems like Microsoft Windows ignores it and continues
> flooding you.
The QoS stuff ultimately ends up dropping packets in order to force the
remote end to slow down.
ECN breaks stuff (since many routers interpret ECN-enabled packets as
having invalid ToS fields). I wonder if I can selectively disable ECN
for certain connections... or even certain routes? Hmm.
I'm hoping that enough stuff out there will honour source quench, so
that I don't end up having to drop packets. My windows boxen will be
collecting most of their downloads through a proxy server, which is
running a real operating system (Linux for now, *BSD when I start
getting adventurous), so source quench *will* work, if I can figure out
how to get the SQ messages onto the wire automagically.
I'd really like to be able to delay ACK packets "enough" to force the
remote end to slow down too (ie: run out of xmit window) without
retransmitting packets.
I'd prefer to try source-quench when the buffers become, say,
half-full. Then start dropping packets when the buffers get to 75% -
most likely by reclassifying "bandwidth hogging" connections into a RED
queue. I love the euphemism "Random Early Detection". Dammit... the
packet's made it 3/4 of the way around the world, only to be dropped at
the upstream from my connection? Hello?!?!
On Wednesday, October 1, 2003, at 06:39 , Sam Couter wrote:
> Retransmission will also cause the sending TCP stack to back off. Is
> this not the goal?
Absolutely not! The goal is to not drop packets *or* have packets
retransmitted. Either way around will end up in wasted bandwidth.
I'd prefer to waste 10Mb of RAM in queues and buffers than drop a single
packet. Ideally, any packet-dropping would be done by my upstream,
before they try to shove packets down the skinny pipe downstream to my
network. I don't see the sense in shaping traffic over my link by
dropping packets once they've already been sent to this network.
Dropping packets at my end seems an awful waste of bandwidth for the
whole intermediary network.
I would also like to try shaping traffic before it leaves the remote
network - if the remote end starts off slow and stays slow, usage of my
local network link can be more fairly allocated.
Perhaps one day people will be using Fast TCP (or TCP Vegas, I think
it's called), and I can start playing silly buggers with delay queues
for ACK packets ;)
Alex
More information about the linux
mailing list