Workshop on Adaptive Video Streaming over IP Networks
I attended the Workshop on Adaptive Video Streaming over IP Networks held at Cisco Systems in Boxborough, MA, USA, on 12-13 October 2010, to present an update on our adaptive IPTV work.
As usual for these workshops, the presentations were an interesting mix of academic research and industry pragmatism. The first day starting with a talk from Sameer Akshabi and Constantine Dovrolis (Georgia Tech) on some experimental evaluation of commercial HTTP streaming video systems, to understand their rate adaptation behaviour. The chunk-based HTTP pull model seems widely accepted, but there is enormous variation in the rate at which chunks are fetched, the size of the playout buffer built-up, and the way in which the applications adapt to changes in the available bandwidth. Some of the systems tested seem to work reasonably well, although they're slow to make use of increases in available bandwidth (possibly intentionally, since variable quality gives a poor user experience); others show various signs of instability. It's clear that this is a rapidly evolving area, with much still to learn.
Zhi Li and Bernd Girod from Stanford outlined some ideas on how to make a more realistic HTTP adaptive streaming simulator using ns-2. Following on from the previous presentation, they clearly described how the two control loops—TCP congestion control operating on the RTT timeframe and application-level adaptation operating over several seconds—interact, highlighting the cause of the stability problems. They also noted an interaction between TCP and client request patterns, which can cause a client that pauses for more than the TCP RTO between chunk request to go into TCP slow start, negating the advantage of keeping the connection open.
The morning concluded with a keynote from David D. Clark (MIT) on “After the triple play: a ten year plan” exploring the economics of the ISP market in the US. Interesting, but it's clear that the ISP marketplace in the US is very different from that in many other parts of the world.
In the afternoon, Ashish Khisti from the University of Toronto presented some theoretical models for adaptive streaming over multihop wireless networks. Then, Dina Katabi (MIT) described their SoftCast system for joint source-channel coding of wireless video for error resilience. This is a very interesting system, leveraging the best ideas from analogue video transmission and modulation to enhance digital video transmission.
The end of the first day, and the morning of the second were filled with discussion sessions and an Overview of Cisco's Video Streaming and CDN Products, and a review of Current Developments in Adaptive Streaming Standards (focussing on IETF, MPEG, and 3GPP work for HTTP streaming).
In the afternoon, we presented our measurements of UDP streaming performance of various DSL and cable access links in the UK and in Finland, including wide area distribution networks to show that loss patterns include bit error-induced losses as well as losses due to congestion and (early) data rate caps. Using a trace-driven simulation of RTP/SSM with feedback and different FEC mechanisms we then showed that, for the observed links and media transmission rates, simple parity FEC mechanisms are capable of reducing the loss rate to acceptable proportions. We concluded with some early thoughts on the applicability of our measurements to TCP-based streaming.
The final presentation, from Mostafa Ammar of Georgia Tech, focussed on hybrid CDN/P2P adaptive live streaming. All-in-all, it was an extremely interesting workshop, covering a lot of ground. Thanks to Dave Oran and Ali Begen for organising!