Disruption effect on the continuous playback of the patch next up previous
Next: Disruption effect on continuous Up: Failure Recovery - Providing Previous: Failure Recovery - Providing

Disruption effect on the continuous playback of the patch

Let us consider the effect of a disruption on the patch first. In P2Cast, a client receives the patch from its patch server and begins play back immediately. The departure of the patch server interrupts this patch playback. However, the length of the patch is equal to the time difference of the clients' arrival time and the session starting time, which is no larger than the threshold and presumably much shorter than the video length. For instance, if the video length is 100 mins and the threshold is 10 mins, the patch length averages 5 mins in length, given the arrival process is Poisson. The short duration of the patch makes it relatively immune to disruption since the likelihood that the disruption hits the patch is much smaller than that of the base stream. Fig. 16 plots the probability of the number of candidate clients contacted (i.e., the candidate patch servers contacted during the patch recovery, or the candidate base stream servers contacted during the base stream recovery) before a client can successfully recover from the failures. The left column of Fig. 16 is for the patch recovery; and the right column is for the base stream recovery. Since a client can be disrupted more than once, we collect the accumulated number of candidate clients contacted in recovery. The top two plots in Fig. 16 are for a threshold equal to 10% of the video length; and the bottom two plots are for a threshold equal to 30% of the video length. We assume that a client departs early with a probability of 0.1, and an early departing client is equally likely to depart at any point during the playback. Most clients (0.996 for threshold 10% and 0.983 for threshold 30) will not be disrupted at all during the playback of a patch stream in our example. Patch delivery is disrupted only if the patch server departs while serving the patch. Since the length of patch is usually short, the chance that a patch gets disrupted is small. Furthermore, the probability that a client has to contact more than 5 candidate clients during patch recovery is less than 0.0003 and 0.0019, respectively, for the threshold of 10% and 30%. Thus if we can delay the playback for a short period and let the buffer build up a bit, say by delaying the playback to the time to contact 5 clients, the continuous playback of the patch stream can be provided with high probability.

Figure 16: Probability of no. candidate clients contacted for recovery caused by client departure.
\begin{figure}
\centering
\epsfig{file=figures/TRANSIT/distr_nocontact.eps, height=1.9in, width=3in}
\end{figure}


next up previous
Next: Disruption effect on continuous Up: Failure Recovery - Providing Previous: Failure Recovery - Providing
Yang Guo 2003-03-27