Determining QoS Thresholds

unclerico · ‎10-03-2009

I'm back with another question regarding how one goes about determining the different thresholds used in a QoS policy. For WFQ how do I determine the CDT and number of dynamic queues to use?? For WRED how do I determine the low and high thresholds and the MPD?? If I were to use tail-drop, how would I defined the queue-limit?? Does it all have to do with the speed of the interface?? Is there a scientific forumla to use to calculate these numbers?? I have a very good grasp of the whole QoS technology, but I just feel like I'm really missing something when it comes to determining these numbers. Thanks ahead of time for your help.

Giuseppe Larosa · ‎10-04-2009

Hello Dennis,

this document can help

http://www.cisco.com/en/US/docs/solutions/Enterprise/WAN_and_MAN/QoS_SRND/WANQoS.html#wp96447

About your questions:

WFQ: when it is used it is usually used with default settings

dynamic queues can be up to 4096

CDT has to be a power of 2

the overall limit is given by

hold-queue N out

in interface config mode

WRED:

MPD low threshold and high treshold are chosen in an attempt to provide differentiated treatment to packets belonging to different classes.

making low thrshold = high threshold emulated tail drop.

More then the real values are the comparison of the different profiles that makes the difference.

the exponential weigthing factor of WRED is actually to be tuned based on the speed of the interface.

For this there is a formula but I'm missing it in this moment.

max size of a software queue is that of the packet memory outgoing of the interface or of the processor handling the port.

max size of physical queue is provided by

tx-ring-limit m

(exact command can vay with platform and type of interface).

Hope to help

Giuseppe

unclerico · ‎10-06-2009

Giuseppe,

That's a great document. Thank you very much for sharing it. I'll try to find something on the calculation of the weighting factor, but if you come across something please post back.

Joseph W. Doherty · ‎10-04-2009

Alas, don't believe there's a simple answer to these questions. Much depends on the nature of the traffic and what you're trying to accomplish. Even something as "minor" as packet size comes into play because queue depths and drop points generally count packets and packets can vary in size.

Generally for non-bandwidth probing traffic you want to queue bursts and avoid dropping packets up to the point where queuing delay is worst than the drop would be. (For such traffic, you also want to size available bandwidth such that it's not oversubscribed for such traffic and with expected average queuing delay that's acceptable for such traffic. I.e. QoS techniques can't make up for insufficient bandwidth for such traffic.)

For bandwidth probing traffic, you want to queue to maximize "goodput", which attempts to allow such traffic flows to discover and utilize all available bandwidth, but provide feedback (often via drops) to such flows when they oversubscribe available bandwidth.

Since we often have different types of traffic crossing an interface, we often need multiple settings that best handle each type of traffic.