In the experiments, we transmit two MDC coded streams from the source host to the destination host. The first path we use is the direct path, the route of which is determined by the default routing algorithm. For the second path, we send the stream to a designated relay server in a different location, and let the relay server forward the packets to the destination. For each stream, we send 30ms UDP packets with payload size of 180 bytes, reflecting 8-bit PCM for finer quantization and 4-bit ADPCM for coarser quantization, at 8KHz sampling rate. Packet sequence numbers and delays are collected and recorded at the receiving host for both streams. Each experiment runs for 180 seconds for both steams. The clocks of the source and destination hosts are synchronized using the Network Time Protocol [11] before the measurements start.
Figure 3: Experiment setup. Source and destination hosts are shown as white
circles, and relay servers as gray circles, all with their IP addresses
labeled. Intermediate service providers are represented by boxes, and
numbers in parentheses show rough delay for the packets to traverse
corresponding providers or other interconnected networks.
In Experiment 1, the source host is located at Netergy Networks in Santa Clara, California, and the destination host is at MIT. The first stream follows the direct path. For the alternative path, we explicitly direct the flow to the relay server at Harvard first, and the packet received by the relay server is then forwarded to the MIT host immediately. The route from the source to the relay, and that from the relay to the destination are determined by routing algorithms without our intervention. The setup of Experiment 1 is shown in Fig (a), with the intermediate service providers shown. Surprisingly enough, although Harvard and MIT are close neighbors, the flow sent to them from the same source follows very different paths, which provides us with two independent channels. The only link shared by the two paths are owned by Exodus Communications, the major service provider of Netergy Networks. The shared link constitutes only a very small part of the total routes, and does not contribute to any violently varying behavior of the channels. This can be observed from the correlation coefficient of the delay of the two streams listed in Table 2. Other statistical quantities for the two streams also listed in Table 2 include the delay median, link loss rate, and delay standard deviation, which characterizes the delay jitter. It should be noted that stream 1 in the first experiment has very high variation in delay since occasional outage periods of up to 2000-3000 ms are observed in the trace and a few packets have experienced delays of up to 2000-3000 ms as a result.
Experiment 2 is performed in a similar way. The source host is at Rutgers State University, New Jersey, and the destination is in Erlangen, Germany. We use the same relay at Harvard to forward the packets of the second stream. In this experiment a long cross-Atlantic link is shared by the two paths. However the delay correlation between streams is still quite low, meaning that the channels are largely independent in delay statistics. This is because, although the cross-Atlantic link contributes most to the total end-to-end delay, the delay on this link is nearly constant due to the high bandwidth and good quality of the fiber connection. The 40ms delay shown in Fig. 3 (b) is mainly the propagation delay from US east coast across the Atlantic to Europe, which is a constant component. On the contrary, although the delay introduced by domestic service providers is only a very small part of the total delay, it presents large variations due to limited bandwidth and network congestion. However these variations are largely independent for the two paths.
Table 2: Delay and Loss Statistics of the Two Streams in the Experiments.