On Fri, 16 Dec 2005, Min Qiu wrote:
Hi Chris,
hey :)
-----Original Message----- From: owner-nanog@merit.edu on behalf of Christopher L. Morrow Sent: Thu 12/15/2005 10:29 PM To: John Kristoff Cc: nanog@merit.edu Subject: Re: The Qos PipeDream [Was: RE: Two Tiered Internet]
snip...
Speaking to MCI's offering on the public network it's (not sold much) just qos on the end link to the customer... It's supposed to help VOIP or other jitter prone things behave 'better'. I'm not sure that we do much in the way of qos towards the customer aside from respecting the bits on the packets that arrive (no remarking as I recall). So, what does this get you aside from 'feeling better' ?
Not 100% true. Through I agree QoS has little impact in the core that has OCxx non-congested backbone (more comments below). In the edge, it does has its place, as Stephen Sprunk and Mikael Abrahamsson explained/described. I recalled we were busy at one time to find out why one of our _most_ important T1 customer's poor VoIP performance. It turned out his T1 was peaked in those peroid.
yup, for t1 customers (or dsl or dial) qos matters only if your like is full when you want to do something with stringent delay/jitter/loss requirements (voip). Possibly a better solution for both parties in the above case would have been MLFR ... possibly. (someone would have to run the numbers, I'm not sure how much the 'qos' service costs in real $$ not sales marked-down-for-fire-sale $$)
snip...
most large networks (as was said a few times I think) don't really need it in their cores. I think I've seen a nice presentation regarding the queuing delay induced on 'large pipe' networks, basically showing that qos is pointless if your links are +ds3 and not 100% full. Someone might have a pointer handy for that?
There is a little problem here. Most of the studies assume packet arrive rate governed by poision rule. Those data collections were from application sessions-->normal distribution when number of sessions-->infinity. However, this can only apply to core, specially two tiered core where packet arrive rate are smoomthed/aggregated. I did experied long delay on a DC3 backbone when the utilization reach to 75%~80%. Packet would drop crazy when the link util reach to ~90% (not 100% tied to quenueing, I guessed). That said,
i think this is where WRED is used... avoid the sawtooth effect of tcp sessions, random drop some packets and force random flows to backoff and behave. I think I recall WRED allowing (with significant number of flows) usage to reach 95+% or so smoothly on a ds3... though that is from some cisco marketting slides)
it only move the threahold in the core from DC3 to OC12 or OC48 (see Ferit and Erik's paper "Network Characterization Using Constraint-Based Definitions of Capacity, Utilization, and Efficiency" (http://www.comsoc.org/ci1/Public/2005/sep/current.html I don't have the access). I'm not sure the study can applied to customer access edge where traffic tend to be burst and the link capacity is smaller in general.
Maybe part of the discussion problem here is the overbroad use of 'QOS in the network!' ? Perhaps saying, which I think people have, that QOS applied to select speed edge interfaces is perhaps reasonable, I'd bet it still depends on the cost to the operator and increased cost to the end-user. it may be cheaper to get a second T1 than it is to do QOS, and more effective. Alternately, customers could use other methods aside from QOS to do the shaping, assuming 'QOS' is defined as tos bit setting and DSCP-like functions, not rate-shaping on protocol or port or source/dest pairs.