You are on page 1of 2

QoE Performance of Adaptive Video Streaming in

Information Centric Networks


Koki Goto, Yusaku Hayamizu Masaki Bandai Miki Yamamoto
Graduate School of Science and Engineering, Faculty of Science and Technology, Faculty of Engineering Science,
Kansai University Sophia University Kansai University
Osaka 564–8680 Japan Tokyo 102–0081 Japan Osaka 564–8680 Japan
Email: {k067669, k911112}@kansai-u.ac.jp Email: bandai@sophia.ac.jp Email: yama-m@kansai-u.ac.jp

Abstract—In Adaptive Bit Rate (ABR) utilized in MPEG- higher average bitrate of downloaded segments brings the
DASH (MPEG Dynamic Adaptive Streaming over HTTP), there better QoE. However, it is desirable that quality change of
have been proposed three representative algorithms, rate-based, consecutive segments is low. In terms of delay, startup delay,
buffer-based and hybrid-based ones. In the information centric
networks, which have been widely studied as a network architec- i.e. waiting time from clicking(selecting) time of a video
ture for efficient content distribution including video streaming, content to starting time of playback, should be shorter as much
routers are generally equipped with cache storage. When some as possible, and rebuffering due to buffer exhaustion should
segments of a required content is occasionally downloaded from not occur. In order to take these streaming quality aspects into
cache storage in a router, content download from a router account, we use the following equation defined in [7] as our
might be faster than segment retrieval from the original content
server. This inaccurate throughput estimation might lead to evaluation metric.
inappropriate bitrate selection of ABR algorithm running on the
user’s application and significant degradation of QoE (Quality 
K 
K−1 
K
of Experience) performance. In this paper, from the viewpoint QoE = q(Rk ) − λ |q(Rk+1 ) − q(Rk )| − μ bk − μs D, (1)
of QoE, we evaluate the influence of in-network caches on three k=1 k=1 k=1
types of ABR algorithm in video streaming.
Index Terms—Information-Centric Networking, In-Network where one video content is divided into K segments. Rk
Cache, Streaming, Quality of Experience denotes bitrate of segment k, and q(Rk ) is utility function
when a user watches a segment whose bitrate is Rk . The first
I. I NTRODUCTION term represents the total bitrate downloaded by a user. The
MPEG-DASH (MPEG Dynamic Adaptive Streaming over second term represents the sum of quality change between
HTTP) [1] is widely used as adaptive streaming to select consecutive segments. The third term shows the sum of
appropriate bitrate that matches network bandwidth. ABR rebuffering time bk . The forth term D is startup delay. λ, μ
(Adaptive Bit Rate) algorithms used in MPEG-DASH are and μs are weighting factors.
classified into three types [2]: rate-based (RBA) [3] where In classical MPEG-DASH over IP, QoE depends only on
download throughput of each segment is used for bandwidth bandwidth variation of a bottleneck link from a user to the
estimation, buffer-based (BBA) [4] where playout buffer level original server. In ICN, however, a cache hit at an intermediate
of a client is used for bitrate selection, and hybrid-based router may affect throughput estimation of the user. If the
(AdapTech) [5] where both download throughput and playout cache hit occurs at the user’s side, i.e. request found the
buffer level are utilized. cached segment before it passes through the bottleneck link,
ICN (Information-Centric Networking) [6] has been widely throughput significantly increases. In this case, the user selects
studied for efficient content distribution including video a higher bitrate than the actual bandwidth with throughput-
streaming. In ICN, routers are equipped with cache storage and oriented approaches, i.e. rate-based and hybrid-based algo-
can provide cached video contents instead of original servers. rithms. When required bitrate is higher than the available
A user can quickly download video contents from the routers. bandwidth, segment download time increases and rebuffering
However, it is difficult for the user to accurately estimate which leads to QoE degradation may occur.
available network bandwidth because the users cannot specify
whether a downloaded segment is obtained from a content III. P ERFORMANCE E VALUATION
server or an intermediate router. Rate-based and hybrid-based To evaluate performance impact of in-network cache, we
algorithms which utilize download throughput for bandwidth conduct evaluation of three representative ABR algorithms
estimation may have negative effects for bitrate selection. from the QoE viewpoint. We modified ndnSIM2.1 for the
In this paper, we evaluate three representative algorithms in evaluation. Figure 1 shows evaluation model, and Table I
MPEG-DASH with occasional cache hit in in-network cache shows parameters for each ABR algorithm. In order to observe
over ICN from the viewpoint of QoE, which is a general user- influences of segment cache hit, segments are randomly cached
perceived metric of video streaming. on Router1 before the bottleneck link. Duration of the video
content is 60 [s] and each segment duration is 2 [s]. We
II. Q O E: Q UALITY OF E XPERIENCE prepare 10 bitrates levels: R={0.1, 0.2, 0.3, 0.5, 0.7, 1.2, 2.0,
QoE is one of the most important metrics that represent 3.0, 5.0, 8.0[Mbps]}. Playout buffer capacity is 30 seconds.
streaming quality perceived by users. In video streaming, In [7], q(Rk ) in Eq. 1 maps bitrate to quality perceived by
users prefer to watch high-quality(resolution) videos, thus the a user, and there are three user-types: QoElin , QoElog , and
 5%$

5%$  5%$
%%$ %%$  %%$
$GDS7HFK $GDS7HFK $GDS7HFK
 

 

4R(


4R(
4R(

 

 


 
 
                    
7KH1XPEHURI&DFKHG6HJPHQWV 7KH1XPEHURI&DFKHG6HJPHQWV 7KH1XPEHURI&DFKHG6HJPHQWV

(a) QoElin (b) QoElog (c) QoEhd


Fig. 2. QoE performance for each user-type

QoEhd . The exact values of q(Rk ) are provided in Table require low-quality segments after the cache hits has high QoE
II. As shown in this table, with QoElin and QoElog , q(Rk ) performance. However, this immediate action causes drastic
gradually increases. In contract, with QoEhd , QoE increases in change of streaming quality between consecutive segments.
a stepwise manner. In other words, quality depends on whether With QoEhd , this quality change can cause significant QoE
a downloaded segment is high-quality or not. We change the reduction, thus QoE performance of RBA is lower than
number of cached segments from 0 to 30 and evaluate QoE AdapTech. AdapTech whose rate reduction is dull, i.e. de-
characteristics for cache hit. crease its bitarte just one representation, has high QoE perfor-
8VHU 5RXWHU 6HUYHU mance. AdapTech does not decrease its bitrate compared to
5RXWHU
RBA and continues to require high-quality segments. In other
0ESV 0ESV 0ESV words, AdapTech prevents unnecessary decrease of bitrate and
PV PV PV conducts moderate bitrate selection. From these results, we
show that appropriate ABR algorithm significantly depends on
Fig. 1. Evaluation model how the user perceives quality of downloaded video segments.
TABLE I We need to select appropriate ABR algorithm suitable for
PARAMETERS OF ABR ALGORITHMS streaming user’s preference in ICN.
RBA [3] β 0.67
BBA [4] (Bmin , Bm ) (10, 26) IV. C ONCLUSIONS
AdapTech [5] (θ1 , θ2 , T, δ, ζ) (7, 14, 10, 0.8, 0.8) In this paper, we evaluate QoE performance of ABR al-
TABLE II gorithms in ICN environment with in-network caches. We
PARAMETERS OF QOE showed that rebuffering is caused by cache hits in throughput-
Name q(Rk ) λ μ, μs oriented algorithms as we expected. Also, we focused on
QoElin R 1 8 how streaming users perceive quality for downloaded video
QoElog log(R/Rmin ) 1 4.3 segments and revealed that appropriate ABR algorithm for a
0.1 → 0.6, 0.2 → 0.8, 0.3 → 1, 0.5 → 1.4,
QoEhd 0.7 → 1.9, 1.2 → 3, 2.0 → 12, 3.0 → 16, 1 8
streaming user significantly depends on the user’s preference.
5.0 → 22, 8.0 → 33 In the future work, we plan to newly develop an ABR algo-
rithm which adapt to various situations and users’ preference.
In the evaluation, we confirmed that rebuffering oc-
ACKNOWLEDGMENT
curs in throughput-oriented approaches due to cache hits.
The maximum rebuffering time of RBA(rate-based) and This work was partly supported by JSPS KAKENHI Grant
AdapTech(hybrid-based) are 3.1 [s] and 8.5 [s], respectively. Number 17H01740.
RBA decreases its bitrate immediately after throughput mis- R EFERENCES
understanding, thus its QoE performance is basically better
than AdapTech. BBA(buffer-based) has no rebuffering due to [1] ISO/IEC 23009-1: 2014, “Information technology – Dynamic adaptive
streaming over HTTP (DASH) – Part 1: Media presentation description
its conservative policy, however, its downloaded bitrate is very and segment formats,” May 2014, https://www.iso.org/standard/65274.
low and it has low QoE performance. html
Figure 2 shows the average QoE performance for each user- [2] J. Kua, et al., “A Survey of Rate Adaptation Techniques for Dynamic
Adaptive Streaming Over HTTP,” IEEE Communications Surveys &
type. Each plot is the average value for 100 different random Tutorials, Vol. 19, no. 3, pp. 1842-1862, 2017.
cache placements. In the case of low cache hit ratio, the [3] Y. Sun, et al., “Rate adaptation for adaptive HTTP streaming,” in Proc.
best is RBA, the next is AdapTech, and the worst is BBA MMSys’11, Feb. 2011, pp. 169-174.
[4] T. Huang, et al., “A Buffer-Based Approach to Rate Adaptation:
in all user types because throughput-oriented approach can Evidence from a Large Video Streaming Service,” in Proc. ACM
stably estimate network bandwidth. However, when cache hit SIGCOMM’14, Aug. 2014, pp. 187-198.
ratio increases, the performance ranking changes according to [5] S. Akhshabi, et al., “An experimental evaluation of rate-adaptive video
players over HTTP,” Signal Processing: Image Communication 27, pp.
user-type. With QoElin and QoElog , RBA has the best QoE 271-287, 2012.
performance and the hybrid-based (AdapTech) has the worst [6] V. Jacobson, et al., “Networking Named Content,” in Proc. ACM
one. In contrast, with QoEhd , AdapTech is the best. CoNEXT’09, Dec. 2009, pp. 1-12.
[7] H. Mao, et al., “Neural Adaptive Video Streaming with Pensieve,” in
With QoElin and QoElog , RBA which requires high-bitrate Proc. ACM SIGCOMM’17, Aug. 2017, pp. 197-210.
during cache hits and immediately decreases its bitrate to

You might also like