[c-nsp] to shape or not to shape

[Arie Vayner (avayner)]

Roger,

This is a sub rate link, as you use a physical rate of 10Mbps with a
downstream service of 5Mbps - This means that somewhere down the link
(on the SP network) they would be dropping anything above 5Mbps.

Your router does not have any way (except shaping) to know that there is
a limit for 5Mbps, and it would be empting anything queued in its egress
buffers at a rate of 10Mbps, which would be then dropped by the
downstream rate limit at 5Mbps.

If you want to be able to control which traffic is dropped or
prioritized in case you have bursts or an overload (for example someone
uploading a file, while someone else running a VOIP call) you must
perform the shaping so that your router would know that it must release
packets to the line at a rate not higher than 5Mbps. The result would be
that any drops (if any) would be done by your router, which is aware of
the child QOS policy, and not on the SP side, which has a flat drop
policy for anything above 5Mbps.

You need to remember that IP traffic is very bursty, and even if on
average (usually 5 minute average, or even 30 seconds average) you see
only <5Mbps of upstream traffic, it down not mean that on a sub-second
level you do not have full line rate bursts (you have them for sure...)

Arie

-----Original Message-----
From: cisco-nsp-bounces at puck.nether.net
[mailto:cisco-nsp-bounces at puck.nether.net] On Behalf Of Roger Wiklund
Sent: Saturday, October 09, 2010 17:49
To: Cisco-nsp
Subject: [c-nsp] to shape or not to shape

I have a question I have been thinking about.

Let's say we purchased a 5Mbit Ethernet Link. The physical speed of
the link is 10Mbit, so we shape outbound traffic to 5Mbit, like such:

class-map ef
match ip dscp ef
class-map af4
match ip dscp af41, af42, af43
class-map af3
match ip dscp af31, af32, af33
class-map af2
match ip dscp af21, af22, af23
class-map af1
match ip dscp af11, af12, af13
class-map be
match ip dscp be

policy-map qos
class ef
priority 1024
class af4
bandwidth remaining percent 40
random detect dscp-based
class af3
bandwith remaining percent 30
random detect dscp-based
class af2
bandwith remaining percent 20
random detect dscp-based
class af1
bandwith remaining percent 9
randon detect dscp-based
class be
bandwith remanining percent 1


service-policy shape
class class-default
shape avarage 5000000
policy-map qos

interface wan
service-policy output shape

So, as we shape, as long as we have buffers, we will never see any
tail drops, as we will just delay the packets until we send it,
correct?

Now imagine we have a framed e1.

interface wan
bandwith 1984

As we have the full bandwith, no need to shape, so I will just apply
the qos service policy for outbound traffic.

If this e1 is 100% utilized, we will get tail drops when the buffers are
full.

So my question now, what if the shape the e1 to 1984, we will still
have the full speed, but we shape, and thus avoid tail drop, and just
delay the packets instead. I'm thinking we avoid TCP restarts etc etc.
pros/cons, or am I wrong about the whole thing? :)

Appreache any comments,

Thanks!

/Roger
_______________________________________________
cisco-nsp mailing list  cisco-nsp at puck.nether.net
https://puck.nether.net/mailman/listinfo/cisco-nsp
archive at http://puck.nether.net/pipermail/cisco-nsp/