4 P ROACTIVE C ONTENT C ACHING FOR HIGH - RELIABLE VIDEO
5.1 H IGHLY -R ELIABLE B UFFER S TRATEGY
In this chapter, we introduce a highly-reliable buffer strategy with the efficient wireless resource usage. Although our primal goal is to avoid video playback interruption because of the network quality degradation, in particular on subway passengers, similar to [62], we also aim to reduce wasted mobile traffic that is generated by quitting video playback.
In order to achieve these goals, we provide two main approaches called
“long-term throughput prediction” and “guaranteed buffer filling mechanism.” Because a recent mobile terminal has an enough large memory capacity, we assume that the large amount of playout buffer can be temporally allocated on the client terminal.
5.1.1 Long-Term Throughput Prediction
We have proposed history-based long-term throughput prediction in [81].
Unlike the previous prediction methods [2, 56, 57], our proposed method considers the
Long-Term Throughput Prediction Scheme and Its Application for
Video Streaming Services
Chapter 5: Long-term Throughput Prediction Scheme and Its Application for Video Streaming Services
104
user mobility. We assume that a moving route of users is fixed, corresponding that the user rides on a train, although our approach can be extended to free route cases such as in a car or walking likely in Chapter 3. Because the mobile communication quality varies depending on a distance to a base station and/or crowd density around the user, we simply construct several throughput prediction models depends on time of day and location information. In particular, while moving a long distance, the user has to hand-over several base stations, including Wi-Fi spots, in order to keep connect with the Internet. In such case, the future throughput can be predicted by switching these prediction models, which is provided on each base station, in conjunction with the user location and the hand-over. In order to construct these models, we utilize cloud computing platforms and collect wireless network information, including location and throughputs. An overview of our proposed scheme is showed in Figure 5.1.
Figure 5.1: Overview of our proposed scheme for long-term throughput prediction [81].
In order to validate our scheme, we evaluate the prediction accuracy over actual Long Term Evolution (LTE) networks between Hirai and Asakusabashi stations located in Tokyo, Japan, at 9 a.m. and 9.p.m with Galaxy SC-03E. We acquire hourly throughput data in different regions in advance and build the throughput prediction models for each region (i.e., each base station.) Note that we adopt Yoshida method [2]
to construct each the model.
The prediction accuracy of our proposal and conventional method is shown in Figure 5.2. In order to calculate the prediction accuracy, we adopt the same calculation as described in [2] (calculating Root-Means Square Error). We also adopt Yoshida
Chapter 5: Long-term Throughput Prediction Scheme and Its Application for Video Streaming Services
105
method [2] as a conventional method and this method uses a single throughput prediction model. As shown in Figure 5.2, we can successfully predict next 500 sec throughputs with high accuracy.
Figure 5.2: Prediction accuracy of our proposal and conventional method [81]. (The horizontal axis represents next moving time of the user, and the conventional method indicates [2].)
5.1.2 Guaranteed Playout Buffer Filling Mechanism
In order to avoid the video playback interruption because of the network quality degradation, it should be necessary to complete downloading enough video segments until the network quality gets low. Our proposed approach called “guaranteed buffer filling mechanism” can calculate an optimal playout buffer size and schedules video download timing in a theoretical manner.
Our buffer strategy model and parameter definition are shown in Figure 5.3 and Table 5.1, respectively. We assume that the network quality degradation can be detected preliminarily using the long-term throughput prediction scheme as discussed in the previous sub-section. This indicates that the client can know the time when the network quality degradation happens (toff) and also its duration (Δtoff). We also assume that, in our approach, no video segments are requested during the network quality is low. This is
Chapter 5: Long-term Throughput Prediction Scheme and Its Application for Video Streaming Services
106
because we expect that this behavior contributes an energy saving for the mobile device as shown in [8, 82]. We aim to calculate download time for buffer filling (Δton) and schedule its timing (toff Δton) in a theoretical manner.
Because no video segments are downloaded during the network quality degradation, maximum playout buffer size is defined as follows:
t
offB
B
max
min
(5.1)where Bmax is maximum playout buffer size, Bmin is minimum playout buffer size and toff
Δ is duration of network quality degradation.
Note that some video segments are decoded during downloading video segments. Thus, the amount of additional video segments that should be downloaded is computed as follows:
on
t
t
B
B
max
( )
(5.2)where B(t) is current playout buffer size and Δton is duration of additional buffer filling.
Thus, download time for these additional video segments (Δton) is calculated as follows:
high on t
on
C
R t
B
t B
(
max ( ))
(5.3)
where R is video bitrate and Chigh is the average predicted throughput of high-speed networks.
Hence, by substituting equation (5.1) and (5.2), (5.3) becomes as follows:
) (
* ) 1 (
1
) (
min off t
high
on
B t B
R
t C
(5.4)Therefore, download time for buffer filling (Δton) heavily depends on the average predicted throughput and the video bitrate.
Chapter 5: Long-term Throughput Prediction Scheme and Its Application for Video Streaming Services
107
The guaranteed buffer filling mechanism should be combined with other playout buffer strategies. As shown in Figure 5.3, once the network quality degradation is detected, the amount of additional video segments (Bmax – B(t)) and their downloading time (Δton) are calculated by equation (5.2) and (5.4), respectively. Then, the buffer strategy should be switched to our proposal method when inequality (5.5) is fulfilled.
) ( B
mint t
t
on off
(5.5)
where α is a margin for switching buffer strategy.
Otherwise, other basic buffer strategy should be selected.
Figure 5.3: Our proposed buffer strategy model [15].
Chapter 5: Long-term Throughput Prediction Scheme and Its Application for Video Streaming Services
108
Table 5.1: Parameter definitions [15]
Paramter Denition
t
0 Start timet
Current timet
off Time when network quality degradation happens ton [sec] Download time for buffer filling
toff
[sec] Duration of network quality degradation
)