On Capacity Of P2P Live Streaming System

P2P live streaming system is becoming an important internet application. For example, PPLive, a popular P2P streaming system, had 100 million global installed bases,2 million simultaneous online users, and accounted for 10% traffic in China's internet in 2009. Besides, the scalability of P2P live streaming system is also impressing. It is reported that PPLive supported 1,480,000 audiences viewing a live play at the same time with 1 PC server and 10M bandwidth in 2007.We study the chunk propagation capacity of P2P live streaming system and propose the improvement algorithms. We regard the chunk propagation procedure as a collision resolution process, and propose a model to represent it. This model captures essential aspects of the system, including the streaming server's upload capacity and strategy, peers'upload capacities, and piece request rates. We use this model to evaluate different chunk acceleration algorithms. The evaluation result is consistent with that of experiments. We next find a simple but interesting cache rejection algorithm used in PPLive, i.e. the cache rejection of peers is synchronized with the chunk upload of media server on chunk offset, and model its buffer as a fixed-duration buffer. Based on this property, we propose a new method to measure the chunk propagation procedure in real-world large-scale network. The measurement result verifies our model. We also propose the blind random piece scheduling algorithm and two new piece acceleration algorithms to simplify the system design and decrease the piece propagation delay by more than 50%.We study the capacity of P2P live streaming systems to handle flash crowds. We noticed that the system has some degree of capacity to survive flash crowds:If the size of flash crowd does not exceed this capacity, the system can recover from the initial degradation of quality and restore to a stable state. Otherwise, the system will collapse, meaning that all newly-arrived peers can not start local playbacks. We present a model that represents this procedure and report our initial results. Moreover, we propose a level-by-level admission control mechanism to exploit this capacity of system to handle flash crowd of large size. We final prove that a P2P streaming system with admission control has excellent capacity to handle flash crowd:It can recover from flash crowds of any size, and the recovery time scales logarithmically with the size of flash crowd. We study the capacity of P2P live streaming systems to handle the abrupt increase of streaming rate. We find the abrupt increase of streaming rate degrades the system, but it has some degree of capacity to recover, although it is limited. We present a model that represents this procedure. We also find the special strategy used by PPLive when the user can not recover timely.