Network adaptive media transport CHAPTER 3.5 length of 20 frames. The base and enhancement layer bit rates and PSNRs were similar to those of the sequence in Section 3.5.3. In Figure 3.5-7, PSNR-versus-rate curves compare the multiple-deadline schemes and the single-deadline scheme, as well as a heuristic scheme. The heuristic scheme uses prioritized, deadline-limited ARQ in which base layer retransmissions had highest priority, followed by base layer transmissions, enhancement layer trans- mission, and enhancement layer retransmissions. Retransmissions were triggered when packets were not acknowledged within the 0.90 point of the cumulative distribution function of the round-trip time. Figure 3.5-7 shows that the multiple-deadline formulation yields up to a 3-dB improvement over a single deadline. The single- deadline scheme does not recognize the utility of late packets and often misses opportunities to schedule valuable data units close to, or after, their original dead- lines. The R­D optimizing schemes outperform the heuristic schemes regardless of whether the heuristic schemes are used with ARD-enabled clients. 3.5.5 Dependent packet delays the dependence among delays, however. In the Internet, successive packets usually travel along the same path, might experience a similar backlog while waiting in the same queues, and rarely arrive out of order. This results in strongly dependent delays of successive packets. In streaming simulations that employ measured In- ternet delay traces, we have observed that the iid model can lead to suboptimal scheduling performance. For ex- ample, Figure 3.5-8 shows simulation results when packets were delayed according to a delay trace mea- sured over a 14-hop Internet path with a cable modem last hop, as described in [10]. At transmission rates above 80 kbps, the multidead-line R­D optimizing formulation described in Section 3.5.4 is outperformed by the simple heuristic ARQ scheme (also described in Section 3.5.4). The suboptimal performance at high rates is due to the iid delay model assumed by the R­D optimization algo- rithm. With the iid model, policies that specify repeated transmission of a data unit at successive opportunities yield lower calculated error probabilities for errors due to late loss. The algorithm mistakenly believes that if the data unit is delayed the first time it is transmitted, sub- sequent transmissions may arrive earlier and on time. Thus at higher rates, the algorithm sends packets mul-