2008-09-18 Readings - Quality of Service

Written post-facto due to network prelim ...

FUNDAMENTAL DESIGN ISSUES FOR THE FUTURE INTERNET

Scott Shenker

Summary: Really helpful paper that frames many issues concerning the whole QoS business. Offers a game-theoretic, utility-maximization view of the goal of network design. Argues that global utility will be increased by introducing different services in the internet. Acknowledges that the increase in global utility may come at the cost of decrease in individual utility. Suggests that applications could receive services implicitly assigned by the network, or they could explicitly request services from the network. Pros and cons exist for each, no clear resolution, need pricing mechanisms to align incentives. Uses application utility/bandwidth profiles to argue for admission control. Over-provisioning seen as less cost-effect alternative.

Background: None really, just need to appreciate the network basics.

Discussion & Criticism: The paper already identified some of the potential issues. The application utility-bandwidth profiles would be hard to characterize in detail, but the general classes outlined in the paper are sufficient. However, the game theoretic argument may fall apart due to incentive mis-alignments. The biggest concern is that the global increase in utility should come at the cost of decreased utility for some users. Without any mechanism for utility transfer, there would be no incentives for the disadvantaged users to cooperate and switch to the global optimal. It's hard to imagine what utility transfer mechanisms can be usefully put in place. In business/economy, utility transfer is generally facilitated by a transfer of money. The mirror mechanism does not exist in the Internet.

Perhaps a more fundamental solution is a new architecture with better accounting abilities. Recall from the original DARPA design that accounting was of low priority.



SUPPORTING REAL-TIME APPLICATIONS IN AN INTEGRATED SERVICES PACKET NETWORK: ARCHITECTURE AND MECHANISM

David D. Clark, Scott Shenker, Lixia Zhang

Summary: Describes the architecture and mechanism of what became the Intserv standards. Has nice taxonomy of internet applications vis-a-vis their delay and adaptation requirements. Outlines to token or leaky bucket mechanism to ensure provable delay and guaranteed service. It's actually token bucket + weighted fair queueing that does the trick. Scheduling employs ideas like spreading and transferring the jitter. Hinted at the need for pricing mechanisms to support. Outlines general issues associated with the need for admission control.

Background: Network basics.

Discussion & Criticism: The paper uses a totally over-elaborate and round about way to explain the token bucket mechanism. I guess second pass explanation would always be more succinct. But haven read a succinct explanation before, I really was quite struck by the complexity of their treatment. Mathematics quite elaborate but not a problem - they're presumably checked by Parekh and Galleger. The handshake for establishing the level of needed services was deferred to a late work. Scott said QoS has been a waste of time for him ... well what can we say ... Business + practicality > technology? Or maybe another warning sign that technology done without an eye to implementation practicality will inevitably fail? Scott is awesome for butchering his own work though.