FEC versus PRT - Scalable Continuous Media Streaming Systems - page 247

85 %

80 %

75 %

70 %

65 %

60 %

200 1000 1200 1600 1800

0 400 600 800 1400 2000

Number of servers

MTTF =1E+18 hrs MTTF =1E+20 hrs MTTF =1E+22 hrs

MTTF =1E+24 hrs MTTF =1E+26 hrs MTTF =1E+28 hrs

Figure 13.8 Reduction in bandwidth overhead versus number of servers with six different MTTF
requirements

13.5.4 Scalability
Finally, we investigate the scalability of the PRT protocol. Specifically, we want to know if
PRT can maintain the reduction in bandwidth overhead when we scale up the system to more
servers. We plot in Figure 13.8 the bandwidth overhead reduction achieved by PRT versus the
number of servers in the system for six different MTTF requirements. The results clearly show
that the bandwidth overhead reduction achieved by PRT in fact increases with the system scale,
suggesting that the PRT protocol is indeed scalable.

13.6 Summary
In this chapter, we modeled and compared the reliability of two redundant data transmission
protocols, namely Forward Erasure Correction (FEC) and Progressive Redundancy Transmis-
sion (PRT). Surprisingly, the PRT protocol can achieve over 50% reduction in transmission
bandwidth overhead while maintaining the same or better system reliability. The key is to
encode the data with more redundant data (compared to FEC) but only transmit some of them
initially. As the failure-detection time is far shorter than the server MTTF, the chance of ex-
periencing multiple server failures within a short time, which could lead to system failure in
PRT, is very small. The only trade-off in PRT is increased storage requirement, which is rela-
tively small (e.g., storage overhead ≤ 20% for K ≥ 5) and thus can easily be accommodated