PaceA
PaceK
SST
SST
0
100
1,000
Latency threshold (ms)
10,000
figure 6. end-to-end message latency based on message importance.
TCP Paceline
1,800
1,600
1,400
message Latency (ms)
1,200
1,000
800
600
400
200
0
0
0.2
message importance
(a) median
0.4 0.6
0.8
1
TCP Paceline
1,800
1,600
1,400
message Latency (ms)
1,200
1,000
800
600
400
200
0
0
0.2
(b) 99.9th Percentile
message importance
0.4 0.6
0.8
The low-level transport performance
compares the transport latency, utilization, and fairness of Paceline with
that of TCP. Paceline reduces the median, the 99.9%, and worst-case end-to-end latency by a factor of three to four
times. On the other hand, Paceline has
similar bandwidth fairness to TCP,
high network utilization, and reasonable wire overhead. Paceline is incrementally deployable on the Internet
since it shares bandwidth fairly with
TCP flows while retaining all latency
improvements. (For detailed transport
performance, see Erbad et al. 5)
application-Level Performance:
Does adaptation Work?
Here, we evaluate the performance
of adaptive applications in terms of
application-level quality metrics. The
evaluation sheds light on the trade-off between average quality and interactivity, and then shows the message
latency in Paceline with respect to assigned importance.
Quality and interactivity trade-off.
One of the main issues to consider is
the nature of the trade-off between
overall multimedia quality and interactivity—better interactivity (lower latency) generally comes at the expense
of video quality (for example, spatial
detail). The following experiments fix
the number of flows to eight videos
(4Mbps each, extremely congested
link) but vary the level of interactivity
using a configuration parameter of the
latency threshold (the amount of time
each ADU is given before it expires and
gets canceled by the sender) on outgoing messages.
To quantify video performance
close to the level of user experience,
we measure the frame rate in fps
(frames per second). Adaptation in
the videoconferencing test application prioritizes ADUs according to
two dimensions of video quality: temporal quality (frame rate) and spatial
quality, measured with PSNR (peak
signal-to-noise ratio) of frames. The
video format is scalable so each video
frame consists of eight ADUs with
one base spatial-layer ADU and seven
(progressive) enhancement spatial-layer ADUs. The default adaptation
policy is biased toward temporal quality. That is, as the bit rate of a video
stream drops, spatial enhancement
