Presenter: Junchen Jiang
Co-authors: Rajdeep Das, Ganesh Ananthanarayanan, Philip A. Chou, Venkata N. Padmanabhan, Vyas Sekar, Esbjorn Dominique, Marcin Goliszewski, Dalibor Kukoleca, Renat Vafin, Hui Zhang3
This work introduces VIA, a predictive relay selection approach for improving internet telephony quality.
This work is important because Internet telephony is a big part of everyday internet usage of users and, as the speaker pointed out, internet telephony is increasing rapidly. The authors have analyzed 430 million calls from Skype and found out that 20% of these calls have poor quality.
The speaker started by showing the correlation of network performance and call quality, and showed clear trends that high RTT (>320ms), loss rate (>1.2%), and jitter (>12ms) and pointed out that there is room for improvement using overlay networks. While overlay networks are not a new concept, the authors use well managed relay networks where nodes are in well connected datacenters and the key is in picking the best relay nodes.
The speaker then went through two strawman solutions for picking optimum nodes which look at previous data or explore all options and showed the issues are 1)predicting tomorrow with last week’s data is prune to inaccuracy and 2)there might not be enough calls to explore all options in a short time.
The solution presented by the authors is VIA which combines prediction and exploration by:
Step1: prediction-based pruning to pick top options (k-top) based on confidence intervals (instead of the best option)
Step2: Explore top-k candidates
The speaker compared VIA with a oracle-based overlay and showed it does a better job in improving network performance across all metrics (RTT, loss, and jitter). They also showed that not all ASs have the same level of improvement and some have limited room for improvement.
In some cases via cannot use the full benefit compared to oracle approach due to sparsity of data. The speaker pointed this out as potential future work.
Q and A
Q: You mentioned 1/5 calls have problems and you can solve half of that, which is 10% of calls. But you didn’t talk about the bigger problem, where for example you can’t establish a call in the first place.
A: This work focuses on bad calls due to network performance, but improving Skype as a whole is definitely an important question, but not the goal of this work
Q: Do you have enough data to figure out where to put the super nodes?
A: That’s a really good question but I don’t think we can solve that with this work because we need more information, but perhaps with our data you can pin point which ASs are problematic and maybe for example do peering with them
Q: Did you look at the performance if your algorithm in different managed overlay configs? For example the size of the manage overlay?
A: That’s a very generic question. There are many configs that can affect VIA performance. We haven’t looked into it, but we looked info for example if you have a budget, and you change that budget, there are tweak in VIA which you can work with to cope with the budget