QoS Question

Monkerz

I am green when it comes to QoS, I haven't reached detailed QoS in my studies yet so everything I have grasped so far as been from a co-worker. So pardon me if I wording something wrong or seem eTarded.

Suppose I have a branch with a full T1 that is prioritizing voice traffic over the WAN with the policy map below:

policy-map llq  
  class VoIP
   priority 768
   set ip dscp ef
  class MNGMNT
   set ip dscp af31
   priority 8
  class class-default
   set ip dscp default

Class VoIP is matching any UDP packet eq 5004 and marking it as dscp 46 or ef. So with 768k allotted for this queue, and 8k allotted for "management"; does this only leave 760k for class-default or is class-default allowed to use as much as it needs until the interface starts seeing VoIP matches?

Just trying to grasp how these queues work.

networker050184

The QoS only kicks in in times of congestion. Packets are placed into the queues if the interface output queue is full. So, if nothing is congested the queuing doesn't kick in.

Any extra bandwidth that is not allocated by the policy map is shared proportionately between the queues. However, any priority class can not go over it's priority amount due to a built in policer designed to prevent the priority class from starving other traffic.

So in your case, if there is congestion on the interface the class-default will receive all bandwidth that is not allocated to the priority queues. If there is no other traffic than the class-default can use all bandwidth.

Monkerz

If QoS is only used in times of congestion, does that mean packets are FIFOed out the WAN interface until congestion occurs?

networker050184

Monkerz wrote: »

If QoS is only used in times of congestion, does that mean packets are FIFOed out the WAN interface until congestion occurs?

Yes there is a single hardware output queue for the interface that packets are emptied into from the software queues.

Also keep in mind that different interface types have different default queuing types. E1 and below uses WFQ and everything else uses FIFO. That is changed when you apply your policy map though. These are software queues, but everything has a single hardware queue obviously for packets leaving the interface. You can not change the order of packets once they reach this point.

Monkerz

Ok, so that makes sense. Is there a threshold at which the interface is known as congested or does it just go by 100% load on the interface?

networker050184

Its not necessarily going to be when you use 100% of the bandwidth while looking at a show interface. There are small bursts that can fill the queue before the interface ever reports being at 100% of its bandwidth due to the interval in which data is collected. Basically once the hardware queue is full packets start being placed into software queues. I'm not sure of an exact % where it reaches this point though.

Monkerz

And when the software queues fill, you see taildrop from the specific queue?

I really appreciate your help, Networker. My mind is wrapping around this now.

networker050184

Correct. When the software queue fills you get your tail drops. You have WRED that prematurely drops traffic to prevent tail drop though. This helps prevent higher priority traffic from dropping on the tail end.

shodown

One thing I want to mention is to just not do a map just for voice traffic. I would look at all the critical data that flows from there you can possibly starve out other traffic. Next is this just a T1 PTP or is this a MPLS T1 which you will have markings that the provider is expecting?

Monkerz

This is an MPLS T1, but also this is not the only class of traffic we mark. I was just throwing a simple example out there hoping for a response from someone.