Mailing List Archive

You didn't explain how you were doing the QoS policy on the line.

What you are describing is what is SUPPOSE to happen when your network
interface is congested. You can't magically make the interface do more
work than what is allowed by creating a IP connection track ruleset.

Is there a linux outer in between these two transmission and receiving
points? Whats the bandwidth we are dealing with?

If it is a linux router, make sure it can actually handle the traffic
input your throwing at it.

With that said, I would like to point out not ALL ethernet cards are
created equal.

For example, if I use a RTL 8139 card on a 100MB network, I have noticed
it makes my router work very very hard, much harder than it normally has
to, and this can cause problems with lower throughput and dropped packets.

At the time though, my primary backbone router died and I needed
something quick and dirty. Thats exactly what I got for about 4 hours
till I got the old router hardware backup and running. Things were
slower, but at least things were moving.

Now, if I throw a Intel EtherPRO 100 on the same interface, packet loss
magically disappears and my router is magically able to do all my IP
connection tracking and processing and QoS without dropping packets.

But not because it magically made more bandwidth, bandwidth is the same,
its how the time is spent at the driver level for the cards is what matters.

Maybe, you might want to take a look at putting decent network cards in
your router if this is the case and try again.

-gc

Justin Schoeman wrote:
> I have posted this before under another thread, but did not get many
> replies. So I thought I would post it under a more appropriate subject.
>
> OK, so we have a link that has a fair bandwidth, and a high latency.
> This means that TCP windows get nice and big.
>
> Now I have a problem with ingress shaping, because the current
> implementation just drops packets. This means that we have to wait
> for the sender to notice the packet drop (OK, or for the receiver to
> notice at out of order inbound backet). But either of these can take
> quite a while, during which the sender is still sending data at a rate
> higher than what you want to throttle it to.
>
> What I was considering was, instead of just dropping the packet, send
> out an ACK packet (to the sender of the packet we are dropping),
> repeating the last ack sequence, as recorded in the conntrack table.
>
> This should be the second ack the sender receives, which should
> immediately start a 'slow start' procedure, and get the sender to back
> off.
>
> This is still as wastefull as just dropping the packet, but should
> have a more immediate effect.
>
> The problem is, how will the sender and receiver respond? They may now
> receive a number of packets in completely unexpected order.
>
> Is this practical? Will it work? Will it help?
>
> Thanks!
> Justin

Justin:

TCP window scaling is an inherent behaviour of the tcp protocol and the
parameter can be tunned.

Because you didn't references the devices involved in the problematic link
is very hard give you an opinion about the cause of the problem.

You mentioned the problem with the ingress shapping. Maybe is a problem with
different mtu crossing the frontier (e.g.: optical fiber to ethernet media)
or maybe is a problem in the traffic shapping itself (e.g.:bad
configuration).

In any event, I think that is not a good option "hacks" the protocol
inyecting packets to the tcp sessions.

Having a more information (like a traffic trace when the problem is
happening) will give us more elements to answer your question.

Jorge Dávila.

On Thu, 30 Aug 2007 12:45:43 +0200
Justin Schoeman <justin@expertron.co.za> wrote:
> I have posted this before under another thread, but did not get many
>replies. So I thought I would post it under a more appropriate subject.
>
> OK, so we have a link that has a fair bandwidth, and a high latency. This
>means that TCP windows get nice and big.
>
> Now I have a problem with ingress shaping, because the current
>implementation just drops packets. This means that we have to wait for the
>sender to notice the packet drop (OK, or for the receiver to notice at out
>of order inbound backet). But either of these can take quite a while,
>during which the sender is still sending data at a rate higher than what
>you want to throttle it to.
>
> What I was considering was, instead of just dropping the packet, send out
>an ACK packet (to the sender of the packet we are dropping), repeating the
>last ack sequence, as recorded in the conntrack table.
>
> This should be the second ack the sender receives, which should
>immediately start a 'slow start' procedure, and get the sender to back off.
>
> This is still as wastefull as just dropping the packet, but should have a
>more immediate effect.
>
> The problem is, how will the sender and receiver respond? They may now
>receive a number of packets in completely unexpected order.
>
> Is this practical? Will it work? Will it help?
>
> Thanks!
> Justin
>
>

Jorge Isaac Davila Lopez
Nicaragua Open Source
+505 430 5462
davila@nicaraguaopensource.com