The Holy Grail of computer networking is to design a network that has the flexibility and low cost of the Internet, yet offers the end-to-end quality-of-service guarantees of the telephone network. - S. Keshav
What QoS Is Not It is not a synonym for good performance It is not about local rationing E.g. Packeteer, Vonage ATA And, in this talk, is not taken to include non-elevated services E.g. ABE, QBone Scavenger
What QoS Is Differentiated network service to provide better-than-default (BE) service WLOG, assume hereafter a Van Jacobson, virtual leased line, Premium Service Note that QoS is about removing only one factor that can cause a networked transaction to fail
Some Problems with QoS Theoretic Practical Economic
GigaPoP A Campus A Campus C Campus D Backbone Campus B GigaPoP B Theoretic Problems How do edge-to-edge virtual trunks concatenate to form an e2e service? What exactly are the policers and shapers at inter-domain boundaries?
Practical Problems Requires all-or-nothing network upgrades (e.g. all access interfaces must police) Dramatic changes to network operations, peering arrangements, and business models
Practical Problems (cont.) In a well-provisioned network Premium is indistinguishable from BE How can a user (or even a provider) verify service? What happens to Premium service in the face of a determined adversary?
Some Alternatives to QoS Overprovisioning Cheapest way to provide fabulous service to important apps, is to provide it to all apps Pricing Congestion pricing Nice theoretic properties But not practical Usage-based pricing Would help a lot Business access is increasingly metered Could provide differentiated services (e.g. Paris Metro Pricing)
A History of Non-Deployment QoS Wasnt Needed ( ) QoS Isnt Needed ( ) QoS Shouldnt Be Needed (2007-)
QoS Wasnt Needed ( ) Ambitious QoS program (QBone) Many hard-won lessons Negative outcome not at all a foregone conclusion Naïve codecs ported from ISDN world wouldnt tolerate packet loss Few users of real-time applications anyway
QoS Isnt Needed ( ) Adaptive, loss-tolerant codecs Many users of real-time applications (Vonage, Skype, Internet2 videoconf) Generous provisioning ensures that real- time apps just work
Hang On a Second! ~10 4 hosts with nothing slower than switched 100Mbps Ethernet between them ~25 of these could congest the 2.4 Gbps Abilene backbone (or 100 the 10 Gbps) 90% of traffic is TCP TCP is designed to congest Yet, the backbone is lightly loaded Whats going on?!
The Terrible Truth (overprovisioning works, because TCP doesnt)
QoS Shouldnt Be Needed (2007-) Either TCP will stay broken or be replaced New transport protocols (e.g. XCP, MaxNet, PCP) dont build huge queues Even better packet loss concealment through improved codecs
Summary QoS is interesting QoS is expensive Scarcity should become scarcer QoS has not been needed thus far QoS should not be needed for the foreseeable future