Tuesday, October 28, 2014

Session 3, Paper 4: Using Video-Based Measurements to Generate a Real-Time Network Traffic Map (Sun, Jiang, Sekar, Zhang, Lin, Wang)

(the paper can be found here: http://conferences.sigcomm.org/hotnets/2014/papers/hotnets-XIII-final81.pdf)

Vyas Sekar is the presenter and he opens the presentation with an interesting analogy: In the real-world, real-time traffic information is shown on google map. How come we don't have real-time Internet traffic information shown somewhere?

Why is this problem hard? The authors provide three reasons:
1. It needs millions of vantage points
2. Bandwidth measurement incurs non-trivial cost
3. Real-time view requires continuous updates

However, the authors argue, the rise of the Internet video can be an opportunity. The reason being:
1. Internet video accounts for 30-50% of the traffic. Netflix alone has 50M users (i.e., 50M vantage points). 
2. The cost is low, since now we can do passive throughput measurement during each video session. 
3. Continuous updates are already collected by various service providers. 

The authors envision an Internet Traffic Map Service, which takes the input of measurement results from various providers and network topology, using their proposed inference algorithm, and provides insights on link capacity and link utilization of the Internet. 

Vyas briefly talked through how the inference algorithm works and showed preliminary results on accuracy.

Q&A:

Q: (Anja Feldmann from UT Berlin) Video traffic are rate limited, then how do you use rate-limited traffic to derive the link capacity?
A: That's a fair point, one thing you can do is to look at each burst. But that is definitely something we need to take into account into the inference algorithm. 

Q: We can look at each chunk, but the throughput will be limited by last mile. As a result, you only have visibility at the bottleneck link. Also, how can a service provider, such as Akamai, monetize such a service?
A: How they monetize this is a separate problem, I am not going to answer that, but I want to emphasize that all these service providers already have these measurements today. 

Q: A quick follow-up — Do you think Google and Netflix has a good coverage of the Internet?
A: Yes, we believe so. 

Q: (Nina from Google) What’s the incentive for service providers to open up these information? how do you address privacy issue?
A: I am not these service providers, but again, these is what they have been doing and they have all these data. 

Q: (Jeff Mogul from Google) Any ISP have these data, but the problem is they will never want to reveal it! (comments from the audience on network neutrality concerns.)
A: There is a paper in SIGCOMM reveals that we can infer network neutrality even without ISP providing data. 

Q: (Dave Oran from Cisco) - How confidence are you that you are not measuring a middle box?
A: It’s very possible, especially if they are the bottleneck.  

Q: (An audience from USC) Video traffic is served from CDN, how much of the measurement is going through the core network?
A: That is a fair point, we might have a biased data set towards edge networks. 

Q: (An audience from Google) Regards to the analogy in the beginning. The reason why we can have traffic information is because we still drive our own car. But today, users don’t drive their own traffic. Do you think it's possible that this kind of information will ever be public?
A: I don't know the answer. (TY Huang from Netflix) Netflix already publishes ISP speed index, and we are fairly open as long as there is a need. Our motivation is really to share the knowledge with our users on how the Internet works. (Dave Oran from Cisco) Google also had a SIGCOMM demo sharing these information this year.