Reading notes by Yu Guan: CS2P

CS2P: Improving Video Bitrate Selection and Adaptation with Data-Driven Throughput Prediction

This paper applies data-driven method to predict the throughput. The method used in this paper is exactly the same as《VIA: Improving Internet Telephony Call Quality Using Predictive Relay Selection》，《Shedding Light on the Structure of Internet Video Quality Problems in the Wild》，《CFA: A Practical Prediction System for Video QoE Optimization》and many other (almost all) papers written by Junchen Jiang.

Motivation:

1. Accurate throughput prediction can improve QoE.
   1. Initial bitrate selection: Throughput prediction can help select a suitable initial bitrate when a video session starts.
   2. Midstream bitrate adaptation: throughput-aware bitrate adaptation can deliver a better QoE than pure buffer-occupancy based approaches.
2. Current throughput prediction methods are not accurate enough.

So authors apply data-driven methods to make accurate throughput prediction.

Dataset

Dataset coms from IQiYi.

X: [ClientIP, ISP, AS, Province, City, Server]

Y: [Throughput(0s-6s), Throughput(6s-12s), Throughput(12s-18s), … ]

Key observations:

1. There is substantial throughput variability within a given session and a range of simple prediction models using previous observations in the same session do not provide high accuracy.
2. Many sessions exhibit stateful characteristics in the evo- lution of the throughput.
3. Sessions sharing similar critical characteristics tend to exhibit similar throughput patterns.
4. The nature of the relationships between session features and throughput are quite complex and simple last-mile predictors are inaccurate.

Data-Driven Throughput Prediction

1. Client starts a session s.
2. Clustering similar sessions.
   1. Picking a critical feature M_s (eg. Province=Shanxi, City=Taiyuan)
   2. Aggregating all previous sessions of feature M_s (eg. All sessions in Shanxi, Taiyuan). We call these sessions Agg(M_s).
3. Predict the throughput based on previous similar sessions Agg(M_s)

Hidden Markov Model (HMM) is applied to do the prediction.

Implementation

CS2P can be used both in the server side of client side.

Server side

Advantage: it requires little updates or modifications on the clients.

Disadvantage: the centralized server needs to collect throughput measurements from all clients and compute bitrates for each video session, making it a potential bottleneck.

Client side

Advantage: client is often in the best position to quickly detect performance issues and respond to dynamics.

Disadvantage: it requires client to maintain its own HMM.

Authors choose the client side implementation. Video player is implmented by Dash.js, Server is implemented by Node.js.

Evaluation

Because CS2P is only a throughput prediction method, it need to deploy on a bit-rate adaptation method. Authors choose FastMPC (it is a Sigcomm paper) to build CS2P+MPC strategy.

Baseline method (to be compared with proposed CS2P+MPC):

1. History-based predictors: LS (Last Sample), HM (Harmonic Mean) and AR (Auto Regression).
2. Machine-learning predictors: SVR (Support Vector Regression) and GBR (Gradient Boosting Regression trees).

Two comparisons are considered:

1. Improvement in Prediction Accuracy.
2. Improvement in video QoE.

Of course, CS2P+MPC performs better in authors evaluation.

Some thoughts:

1. Hidden Markov Model is the only thing which interests me in this paper. Other part of this paper is almost the same as the paper I’ve read in these few weeks.
2. Well-written paper.