A Model, Analysis and Protocol Framework for Soft State-based Communication

Summary. The authors define and use a model for soft-state based communication to analyze the consistency behavior and bandwidth consumption behavior given different data arrival rates, loss rates and death rates. Via analysis, the authors conclude that ss can achieve 95% consistency under 0-20% loss rates but can waste 90% of their bandwidth at death rates of 10%.

To improve bandwidth cons. without negatively affecting consistency behavior, the authors explore:

1. A data aging scheme combined with dual-queue scheduling: one queue for already broadcasted data and one queue for fresh data. By allocating bandwidth to the new queue that is slightly more than the new data arrival rate, can improves consistency metric by 15-30% (I couldn't figure out where these percentages came from).
2. Receiver feedback. Question was how to allocate bw between feedback and data. Found that allocating 20% to feedback improved consistency by about 30%.

Introduced a protocol framework that centered around a link scheduler that accepted input from a congestion manager (to derive available bw) and RTP-like receiver reports in order to schedule bandwidth and, in turn, gave the application feedback (to derive loss rate) so that the app could input data into the queue appropriately.

Challenge in soft-state is to maximize consistency while minimizing redundant transmissions.

Data model. They use key/value pairs in a table. Insertion rate of lambda, deletion rate (death rate) of u. Consistency metric is defined and the average system consistency metric is defined to be the average of the instantaneous system consistency over the lifetime of the system (lim t->8)(integral of the instaneous system consistency/T).

The authors use a class-based queueing model in conjunction with Jackson's theorem to find a solution for eventual consistency in terms of lambda, death rate and loss rate. Graphing for various values, they find that a workload with a 15% death rate is 95% consistent for error rates up to 10%. Solving for redundant transmissions and graphing for various values, they find that at loss rates up to 20%, about 90% of the total available bw is wasted.

Based on their analysis, the authors introduce two improvements:

1. Multiple transmission queues. They divide the data into hot (new) and cold (previously transmitted) data. The challenge is to determine how much of the total bw to allocate to each queue. They found consistency is maximized when u[hot] just over lambda.
2. Receiver feedback. Again, the challenge is to allocate bw between data and feedback. With no feedback, consistency about 80%. With u[feedback] is 20-60%, consistency about 99%. u[feedback]>60% reduces consistency until is "collapses" due to insufficient bw for new data.

SSTP is summarized above. The scheduler must know the available bw and loss rates andbe able to determine how much bw to allocate for hot and cold data. Therefore, SSTP includes feedback from a CM and receiver loss reports. They mention SNAP in SSTP as useful for communicating what data a receiver has/has not received.