CoDiO determines which packets to send, and when, by comparing different possible
schedules in terms of their expected impact on video distortion and network
congestion. Forming the expected Lagrangian of expected end-to-end delay (the
metric we choose for congestion) and distortion (i.e. mean squared error), is better
suited than the traditional rate-distortion metric proposed in the seminal work, when
evaluating the performance of a sender operating over a throughput-limited network.
In particular, end-to-end delay depends on the capacity of the network path, reflects
time varying network conditions, and increases without bound as the sending rate
approaches capacity.
To find an optimized schedule, CoDiO selects the most important packets in terms of
video distortion reduction, and transmits them in an order which minimizes the
congestion created on the network. For example, I or SI frames are sent in priority
whereas some B frames might not be transmitted at all. In addition, CoDiO avoids
sending packets in large bursts as this has the worse effect on queuing delay.
Fig. 1. Video quality performance of different schedulers under varying latency constraints.
Figure 1 illustrates the benefits of CoDiO scheduling compared to a traditional ARQ
scheduler which sends packet sequentially and re-transmits lost packets based on out
of order acknowledgement, as long as they are not past their play out deadline. The
performance of a scheduler denoted “CoDiO Light” is also represented. This is a
low complexity CoDiO scheduler which determines sequentially which frame should
be transmitted to minimize the expected Lagrangian cost of congestion and
distortion, and spaces transmissions based on the time needed for the last frame to
traverse the bottleneck link. In the experiment, the video sequence Mother and
Daughter, encoded with H.264 at 270 kb/s, is sent over a 400 kb/s capacity path;
the one-way propagation delay is 50ms and the packet loss rate is 2%. In this
simulation, network paths are simulated in the network simulator NS-2.
Video quality is measured as the average PSNR measured in decibel and the
performance is shown for different maximum tolerated latencies, that is the time
between which a video frame is available at the server and the time at which it is
played out at the receiver. As shown in Fig. 1, the video quality of the three
schedulers is comparable when the latency constraint is relaxed, however, for
tight delay constraints, both optimized schedulers outperform by a significant
margin the ARQ scheduler. The results also indicate that the low-complexity
CoDiO scheduler achieves most of the gain. Hence, in the following, it is this
type of scheduler which will be extended to the P2P scenario, to reduce
computational complexity at each peer.
schedules in terms of their expected impact on video distortion and network
congestion. Forming the expected Lagrangian of expected end-to-end delay (the
metric we choose for congestion) and distortion (i.e. mean squared error), is better
suited than the traditional rate-distortion metric proposed in the seminal work, when
evaluating the performance of a sender operating over a throughput-limited network.
In particular, end-to-end delay depends on the capacity of the network path, reflects
time varying network conditions, and increases without bound as the sending rate
approaches capacity.
To find an optimized schedule, CoDiO selects the most important packets in terms of
video distortion reduction, and transmits them in an order which minimizes the
congestion created on the network. For example, I or SI frames are sent in priority
whereas some B frames might not be transmitted at all. In addition, CoDiO avoids
sending packets in large bursts as this has the worse effect on queuing delay.
Fig. 1. Video quality performance of different schedulers under varying latency constraints.
Figure 1 illustrates the benefits of CoDiO scheduling compared to a traditional ARQ
scheduler which sends packet sequentially and re-transmits lost packets based on out
of order acknowledgement, as long as they are not past their play out deadline. The
performance of a scheduler denoted “CoDiO Light” is also represented. This is a
low complexity CoDiO scheduler which determines sequentially which frame should
be transmitted to minimize the expected Lagrangian cost of congestion and
distortion, and spaces transmissions based on the time needed for the last frame to
traverse the bottleneck link. In the experiment, the video sequence Mother and
Daughter, encoded with H.264 at 270 kb/s, is sent over a 400 kb/s capacity path;
the one-way propagation delay is 50ms and the packet loss rate is 2%. In this
simulation, network paths are simulated in the network simulator NS-2.
Video quality is measured as the average PSNR measured in decibel and the
performance is shown for different maximum tolerated latencies, that is the time
between which a video frame is available at the server and the time at which it is
played out at the receiver. As shown in Fig. 1, the video quality of the three
schedulers is comparable when the latency constraint is relaxed, however, for
tight delay constraints, both optimized schedulers outperform by a significant
margin the ARQ scheduler. The results also indicate that the low-complexity
CoDiO scheduler achieves most of the gain. Hence, in the following, it is this
type of scheduler which will be extended to the P2P scenario, to reduce
computational complexity at each peer.
No comments:
Post a Comment