Reading notes by Yu Guan: VIA: Improving Internet Telephony Call Quality Using Predictive Relay Selection

2017-11-11

Posted by 关宇

Background:

  1. Telephony call becomes an important application in today’s network.
  2. In the current network, the performance of telephony call is poor. And there are limit work on how internet influence telephony call QoE.
  3. So authors analyze how internet influence telephony call QoE and then present a relay selection algorithm to improve the QoE of telephony call.

 

Contribution:

  1. Analyzing, at scale, the impact of network performance on audio call quality.
  2. Present VIA, a centralized control system for telephony call. Quantifying the potential benefits of a managed overlay network for improving audio call quality.
  3. Highlighting the challenges in achieving these benefits and presenting a practical relay selection algorithm that delivers close-to-optimal performance.

In the rest part of this reading note, I describe the work according to these 3 contributions.

 

Analyzing the impact of network performance on telephony call quality

This paper analyzes 3 network metrics: (i) round-trip time (RTT), (ii) loss rate, and (iii) jitter.

Conclusions:

  1. These 3 metrics are all important to QoE. Improving any one of them can improve QoE dramatically.
  2. Wide-area communication, such as international and inter-domain calls, are more prone to bad network performance, and have a large room of improvement.
  3. Poor QoE happens spatially and temporally.

 

These conclusions motivate the need for a network overlay (Conclusion 1) that provides better paths with a global footprint of overlay nodes (Conclusion 2), and the need to choose routes selectively and dynamically (Conclusion 3).

 

Present VIA, a centralized control system for telephony call

Optimal function:

Min \sum Q(c,Assign(c))

C: a pair of clients.

Assign(C): assignment for clients c. Which two relay nodes should be selected to connect these 2 clients. (of course, we can also choose only one, or let 2 client connet directly if they close to each other)

Q(c,Assign(c)): QoE of the assignment for client pair c.

 

Presenting a practical relay selection algorithm

In this paper, a traditional exploration and predict method is applied.

 

2 challenges when applying this method to this problem:

  1. Exploration challenge: huge decision space. (client * client * relay_node1 * relay_node2) Impossible to exploration all of them to find a good choice.
  2. Predict challenge: because of the huge decision space, most of the choices have no historical record. So how to know the predictive QoE of these choices?

 

Solutions of these 2 challenges:

For exploration challenge: Only explore the choice in the rank N.

For predict challenge: Network tomography is applied to predict the QoE of a choice which has no historical record.

 

Some thoughts:

  1. A really well-written paper. Need to learn the writing skill.
  2. The motivation of this paper has problem: Why “RTT, loss rate, jitter influence QoE —-à we need a control plane to do relay node selection”?
  3. In the experimental results, authors test each feature of their method and analyze the benefits can be obtained by each of them. Good experiments.