scispace - formally typeset
Search or ask a question
Proceedings ArticleDOI

A server's perspective of Internet streaming delivery to mobile devices

Yao Liu1, Fei Li1, Lei Guo2, Bo Shen, Songqing Chen1 
25 Mar 2012-pp 1332-1340
TL;DR: Considering the huge demand of CPU cycles for online transcoding, server-side caching is examined in order to reduce CPU cycle demand and it is shown that a policy considering different versions of a video altogether outperforms other intuitive ones when the cache size is limited.
Abstract: Receiving Internet streaming services on various mobile devices is getting more and more popular. To understand and better support Internet streaming delivery to mobile devices, a number of studies have been conducted. However, existing studies have mainly focused on the client side resource consumption and streaming quality. So far, little is known about the server side, which is the key for providing successful mobile streaming services. In this work, we set to investigate the Internet mobile streaming service at the server side. For this purpose, we have collected a one-month server log (with 212 TB delivered video traffic) from a top Internet mobile streaming service provider serving worldwide mobile users. Through trace analysis, we find that (1) a major challenge for providing Internet mobile streaming services is rooted from the mobile device hardware and software heterogeneity. In this workload, we find over 2800 different hardware models with about 100 different screen resolutions running 14 different mobile OS and 3 audio codecs and 4 video codecs. (2) To deal with the device heterogeneity, transcoding is used to customize the video to the appropriate versions at runtime for different devices. A video clip could be transcoded into more than 40 different versions in order to serve requests from different devices. (3) Compared to videos in traditional Internet streaming, mobile streaming videos are typically of much smaller size (a median of 1.68 MBytes) and shorter duration (a median of 2.7 minutes). Furthermore, the daily mobile user accesses are more skewed following a Zipf-like distribution but users' interests also quickly shift. Considering the huge demand of CPU cycles for online transcoding, we further examine server-side caching in order to reduce CPU cycle demand. We show that a policy considering different versions of a video altogether outperforms other intuitive ones when the cache size is limited.

Summary (4 min read)

Introduction

  • The authors set to investigate the Internet mobile streaming service at the server side.
  • For this purpose, the authors have collected a one-month server log (with 212 TB delivered video traffic) from a top Internet mobile streaming service provider serving worldwide mobile users.
  • A video clip could be transcoded into more than 40 different versions in order to serve requests from different devices.
  • Recently, mobile devices are getting increasing popularity.
  • To understand the key challenges of Internet mobile streaming and the difference from traditional Internet streaming, a number of studies have been performed.

II. BACKGROUND AND WORKLOAD OVERVIEW

  • To investigate how current Internet streaming services are delivered to mobile devices, the authors have collected a 30-day server log from one of the largest Internet mobile streaming service providers, Vuclip [9].
  • Vuclip allows any mobile user to search for interested video available on the Internet, and transcodes them on-demand and on-the-fly based on the type of the mobile device.
  • According to their analysis, each video was accessed in more than 2 versions on average (as shown in Table I), and the most popular video was accessed in 41 different transcoded versions .
  • Examining the received requests, the authors find most of them come from UserAgent strings representing mobile users: more than 94% (181 million out of 192 million) requests are from mobile devices.
  • Figure 3 further depicts the hourly pattern from Nov. 8th to Nov. 15th (a week).

A. Mobile System Heterogeneity

  • To provide Internet streaming services to all kinds of mobile devices like Vuclip, a unique challenge is the heterogeneity among mobile devices.
  • Different from the pre-coding approach that was taken by many other service providers to serve specific types of mobile devices, transcoding has to be used to customize the video into a proper format for the requesting mobile device.
  • Among the 84,281 User-Agents that represent mobile devices, the authors are able to get the brand and model information from more than 74,708 (88.64%) distinct UserAgent strings.
  • As shown in Table II, accesses to Vuclip in these 30 days came from 2864 different device models.
  • Delving into this, the authors find that these devices have 92 different resolutions (width and height combinations), ranging from 84× 48 to 1600× 1200.

B. Audio/Video Codec Heterogeneity

  • To play video on a mobile device, both audio and video codecs are required.
  • Such heterogeneity would further increase the load for the server if the server conducts transcoding for the mobile device.
  • The authors find that typically there are 3 audio codecs being used, namely AAC, AMR, and WMA, and there are 4 video codecs being used, namely H.263, H.264, MPEG-4, WMV.
  • As the authors can observe from the table that a larger resolution video does not necessarily come with a high encoding rate.
  • Consider that Vuclip transcodes the video content on-demand, it is not surprising that 87% of video contents have at least one version encoded with High Quality.

IV. CHARACTERIZATION OF MOBILE STREAMING VIDEOS

  • The previous section has shown that mobile device heterogeneity is a great challenge to the service provider.
  • Such resource consumption can drain the limited battery power supply at a very high rate.
  • Note that, here in this figure, each video may have been accessed in several versions in different formats and file sizes.
  • This shows that videos accessed by mobile devices are mostly small in terms of bytes.
  • This provides a great opportunity for reducing the transcoding cost as the authors discuss later in section V.

B. Popularity of Mobile Videos

  • Figure 10 shows the popularity pattern of videos accessed site-wide on Nov. 1st.
  • This figure shows that, in log-log scale, the popularity distribution of videos accessed can be well fitted with a Zipflike distribution yi ∝ 1 iα , where i is the popularity rank of the video, yi is the number of requested sessions for the video, and α is the skewness parameter.
  • Furthermore, this also means caching at the server side is more effective than caching at the edge/client side, if caching at the server side is needed.
  • While Figure 10 shows short-term (one day) popularity distribution, Figure 11 shows the corresponding distribution in long-term, spanning over the entire 30 days of their workload.
  • It was reported that for media workloads with a median file size < 5MBytes, the stretch factor is ≤ 0.2.

C. Popularity of Different Video Versions

  • The authors have shown in Figure 10 that the daily video popularity follows a Zipf-like distribution.
  • As discussed before, each video may be accessed by very diverse mobile devices, resulting in multiple transcoded versions.
  • Figure 12 shows that when different versions are considered as different objects, the popularity cannot be well-fitted with the Zipf distribution.
  • This figure shows that although the daily version popularity cannot fit well with either Zipf or SE distributions, the monthly version popularity pattern can be well fitted with an SE distribution.
  • This means poorer caching performance if caching is needed for these versions.

D. Popularity Evolution

  • The authors have shown that mobile users’ daily accesses for mobile videos are highly concentrated, but monthly accesses patterns are flatter.
  • First, the authors examine the temporal locality at the server side.
  • The authors have shown in Table VI that about 18% new video clips are added into the video repository daily.
  • Previous studies on a video-on-demand (VoD) system report the daily rate of change in Top-100 videos is less than 15% [16].

E. Correlation Between Popularity and Video Length

  • The authors have shown in section IV-A that the mobile video files are often short and the video popularity in a short-term has a Zipf-like distribution.
  • To examine this, the authors group video files into 1 minute interval based on their lengths.
  • Figure 16 shows the total number of requested sessions decreases as the video length increases, and videos shorter than 5 minutes account for about 70% total requests.
  • The authors results show that the correlation coefficient is 0.006, which indicates the correlation is weak.
  • The authors have also conducted similar tests based on versions, and the results are similar.

V. TRADE-OFF BETWEEN CPU AND STORAGE

  • This challenges service providers both technically and economically with the growing popularity of Internet mobile streaming services and it is thus very desirable to reduce the huge demand of CPU cycles for such transcoding.
  • Thus, caching on the server side, sometimes called reverse caching, could be explored to temporarily cache some transcoded objects so that on-the-fly transcoding would not be necessary if the same type of mobile devices access the same video.
  • Intuitively, these different versions could be considered as separate objects in the cache.
  • On the contrary, if the authors consider that different versions of a video are related because along the diminishing popularity of a video, all of its versions may get fewer and fewer accesses, they can also consider a policy in which all different versions of a video are bundled together as one object in the cache.
  • With the accesses of Nov. 1st, Figure 17(a) shows that when the cache size is smaller than 27% of the total size of accessed objects of that day, a videopopularity based replacement policy can achieve the highest cache hit rate, and save roughly about 55% CPU cycles, while a version-popularity based strategy performs the worst.

VII. CONCLUSION

  • The wide adoption of mobile devices in practice has made pervasive Internet streaming possible.
  • While a number of studies have been conducted to examine the streaming services from the client’s perspective, in this work, the authors have studied the Internet mobile streaming services from the server side via one-month server log collected from one of the largest Internet mobile streaming service providers.
  • Through detailed analysis, the authors have shown the great hardware and software heterogeneity of mobile devices, different characteristics of mobile videos, and different user access patterns from those in traditional Internet streaming services.
  • As a great challenge that Vucliplike system faces is the huge demand of CPU resources for online transcoding to deal with heterogeneity, the authors show that caching at the server side with a proper replacement policy can effectively trade-off limited storage size for great savings on CPU cycles.

Did you find this useful? Give us your feedback

Content maybe subject to copyright    Report

A Server’s Perspective of Internet S treaming
Delivery to Mobile Devices
Yao Liu
1
Fei Li
1
Lei Guo
2
Bo Shen
3
Songqing Chen
1
1
Dept. of Computer Science
2
Dept. of CSE
3
Vuclip
George Mason University Ohio State University Milpitas, CA, USA
{yliud, lifei, sqchen}@cs.gmu.edu lguo@cse.ohio-state.edu bshen@v uclip.com
Abstract—Receiving Internet streaming services on various
mobile devices is getting more and more popular. To understand
and better support Internet streaming delivery to mobile devices,
a number of studies have been conducted. However, existing stud-
ies have mainly focused on t he client side resource consumption
and streaming quality. So far, little is known about the server
side, which is the key for providing successful mobile streaming
services.
In this work, we set to investigate the Internet mobile stream-
ing service at the server side. For this purpose, we have collected
a one-month server log (with 212 TB delivered video traffic)
from a top Internet mobile streaming service provider serving
worldwide mobile users. Through trace analysis, we find that
(1) a major challenge for providing Internet mobile streaming
services is rooted from the mobile device hardware and software
heterogeneity. In this workload, we find over 2800 different
hardware models with about 100 different screen resolutions
running 14 different mobile OS and 3 audio codecs and 4 video
codecs. (2) To deal with the device heterogeneity, transcoding
is used to customize the video to the appropriate versions at
runtime for different devices. A video clip could be transcoded
into more than 40 different versions in order to serve requests
from different devices. (3) Compared to videos i n traditional
Internet streaming, mobile streaming videos are typically of much
smaller size (a median of 1.68 MBytes) and shorter duration
(a median of 2.7 minutes). Fu rthermore, th e daily mobile user
accesses are more skewed following a Zipf-li ke distribution but
users’ interests also quickly shift. Considering the h uge demand
of CPU cycles for online transcoding, we f urther examine server-
side caching in order to reduce CPU cycle demand. We show
that a policy considering different versions of a video altogether
outperforms other intuitive ones when the cache size is limited.
I. INTRODUCTION
Recently, mobile devices ar e getting increasing popularity.
For example, according to I nternational Data Corporation, the
total number o f smartphones sold worldwide in 2010 is 302.6
million [1], which is a 74.4% increase from the previous year
(173.5 million). By September 2010, over 58.7 million pe ople
in the US owned smartphones [2].
Besides general web surfing on the Internet, these days more
and more accesses from mobile devices are directed to all
kinds of Internet streaming services. For example, YouTube [3]
is th e amon g the earliest to p rovide streaming services to
mobile devices such as iPhone. Today bo th iOS and Andro id
have native support for YouTube. Other popular streaming
service providers, inc luding Netflix [4] and Hulu [5], also
provide streaming services to subscribed mobile users via
APPs built in various mobile operating systems. Placeshifting
services like Orb [6] and AirVideo [7] allow mobile users to
access media content stored on their home co mputers. Qik [8]
allows users to upload fro m mobile devices and the n broadcast
the video content to their friends. Different from the above
services, Vuclip [9] lets users search and play all kinds of
Internet videos on their mobile devices regardless of their
mobile device ty pes.
To understand the key challenges of Internet mobile stream-
ing and the difference from traditional Interne t streaming, a
number of studies have been performed. As today the majority
of Internet mobile streaming services are delivered in a client-
server architecture, many studies have focused on the resource
consumption and streaming quality received on the mobile de-
vice. For example, Xiao et al. [ 10] studied energy consumption
when watching YouTube on mobile devices. Huang et a l. [11]
investigated fetch ing policies of different m obile video players,
and Finamore et al. [12] examined the potential causes for
inferior stre aming qu ality of mobile YouTube accesses.
However, these studies mainly focus on the client side
by examining specific devices [10], [11] or via local ex-
periments [12]. As the key to the curren t Intern et mobile
streaming services, the server side plays a critical role in the
entire streaming delivery process. Unfortunately, so far, little
is known about the server side, possibly due to the limited
availability o f data from the server side.
To provide in-depth u nderstanding of the current Intern et
mobile streaming services, in this study, we set to investigate
the server side in stre aming delivery to mobile devices. For
this purpose, we have collected a 30-day server log from a
top Internet mobile streaming service provider. In 30 days,
about 105 million video sessions were served with about 212
TB video traffic delivered. Through our analysis, we have a
number of findings. While the details are presented later in
the paper, some highlights are as follows:
A un ique challenge for Internet streamin g delivery to
mobile devices is roo ted from the fact that mobile devices
are very heterogeneous. In this workload, we find over
2800 different har dware models with 92 different screen
resolutions running 14 different mobile OSes, using 3
audio c odecs and 4 video codecs. This greatly challenges
the traditiona l I nternet streaming delivery infrastructure
where the bottleneck often lies in the limited bandwidth.
To deal with the device heterogeneity, run time transcod-
ing is used to customize a vide o to the appropriate

versions on the fly for different devices. A video clip
could be transcoded into more than 40 different versions
in order to serve r equests from different d evices.
Compared to videos in traditional Internet streaming,
mobile streaming video clips are typically of much
smaller size (with a median of 1.68 M Bytes) and the
video duration is shorter a s well ( w ith a median of 2.7
minutes). Furthermore, the daily mobile user accesses
are more skewed following a Zipf-like distribution but
users’ intere sts also sh ift quickly, resulting in a stretched -
exponential distribution in the long term.
To reduc e the huge CPU cycles demanded for transcoding
on the fly, we further explore caching at the server side by
trading off storage for CPU cycles. Our study shows that a
policy that considers different versions of a video altog ether
outperforms other intuitive ones ( e .g., a file based one) when
the cache size is limited. As far as we know, we are the
first to provide a server-side ana lysis on a Vuclip-like Internet
mobile streaming service. Our findings provide new insights
and lay some foundations to improve the curr ent Internet
mobile streaming delivery.
The rest of the paper is organized as follows. We describe
some background and the workload overview in section II
and study the device har dware and software heterogeneity in
section III. We examine various mobile video properties in sec-
tion IV and further explore the trade-off between the storage
and the CPU at the server side in section V. Some related work
is described in section VI and we make concluding remarks
in section VII.
II. BACKGROUND AND WORKLOAD O VERVIEW
To investigate how current Internet streaming services are
delivered to mobile devices, we have collected a 30-day server
log from one of the largest Internet mobile streaming service
providers, Vuclip [9]. Vuclip provides mobile users with the
search-and-delivery ser vices. It allows users to search for and
watch any videos on any video- e nabled mobile phones and
devices.
Different from many existing services that only provide
streaming services to spe c ific mobile devices, Vuclip can
serve any type of mobile devices that are capable of stream-
ing playback. Vuclip allows any mo bile user to search for
interested vid e o available on the Interne t, and transcodes
them on-de mand and on-the-fly based on the type of the
mobile device. To serve different ty pes of mobile devices,
Vuclip employs on-dema nd transcoding at the server side.
Transcoding is a process to convert the requested vid e o clip
to the appropriate codecs, format, and size at r untime upon
a request so that the video can be proper ly rendered and
played on the requesting mobile device. Vuclip transcodes a
video into different versions by choosing the best audio/video
codecs, frame size, f rame rate, and quality level com bination
for the mobile device. Acc ording to our analysis, each v ideo
was accessed in more than 2 versions on average (as shown
in Table I), and the most popular video was accessed in 41
different transcoded versions (as shown later in Figure 7).
To deliver video content, Vuclip uses the tra ditional
client/server (C/S) architecture. T he video file is delivered via
pseudo streaming over HTTP. That is, when the requested
content is available on the server, the client would issue an
HTTP GET request to download the content. A video may
be downloaded via several HTTP GET requests w ith different
partial ranges specified (i.e., range requests). To differentiate
video requests from HTTP requests, we define a request as a
single HTTP transfer between the clien t a nd the server, and a
session as the set of requests that are involved in downloading
an entire vid eo clip.
TABLE I
SUMMARY OF WORKLOAD
Workload Length 30 Days
# of Sessions 105,389,370
# of Requests 192,255,173
# of Requests from Mobile Devices 181,556,344
# of Unique Videos Accessed 4,052,740
AVG. # of Formats Per Video 2.31
Total Traffic Volume 212 TB
The data we collected is from Nov. 1st to Nov. 30th, 2010.
In this o ne-month server log, there are about 105 million
sessions watching more than 4 million different video s. T here
are a total of about 192 million HTTP requests. The total
traffic delivered from the server in these 30 days is about
212 TB. Table I gives a summa ry of this workload. Note
that among all these requests, some are from desktop/laptop
computers instead of mobile devices. In order to focus on
the r e quests from mobile users, we differentiate them in the
server log through the User-Agent strings specified in each
HTTP request. By analyzing the User-Agent, we find there is
a total of 150,072 unique U ser-Agent strings. Among them,
84,281 (56%) rep resent mobile devices. However, examining
the received requests, we find most of them come from User-
Agent strings representing mobile users: more than 94% (181
million out of 192 million) requests are from mobile devices.
With the exclusion of desktop/laptop traffic, Figure 1 gives
an overview of the server side traffic in 30 days. Note that
the left y-axis rep resents the total number of r equests per day,
while the right y-axis represents th e total traffic volume per
day. During this one-month period, despite a small decrease in
the middle, the number of the requests and the delivered traffic
amount kept increasing, indicating the popularity of Vuclip.
Figure 2 shows the hourly mobile streaming access patterns
in a day. The figure indicates that hourly accesses peak around
17:00 GMT. Furthermor e, the total number of requests and the
traffic volume ser ved durin g peak ho urs almost double these
in non-peak hours. Figure 3 further depicts the hourly pattern
from Nov. 8th to Nov. 15th (a week). The figure shows clear
peak and off-peak hourly patterns for each day. The figu re
shows some drop after Nov. 12th. It is likely due to the fact
that Nov. 13th was a Saturda y and Nov. 14th was a Sunday.
We can observe the in crease of accesses again on M onday.

0 5 10 15 20 25 30
5.5
6
6.5
7
7.5
8
x 10
6
Total # of Requests Per Day
Day
5.5
6
6.5
7
7.5
8
Total Traffic Volume Per Day (TB)
Total Requests
Total Traffic
Fig. 1. Daily Accesses in Nov. 2010
0 3 6 9 12 15 18 21 24
1
2
3
4
x 10
5
Total # of Requests Per Hour
Hour
200
250
300
350
Total Traffic Volume Per Hour (GB)
Total Requests
Total Traffic
Fig. 2. Hourly Accesses On Nov. 1st 2010
8 9 10 11 12 13 14 15
1
2
3
4
x 10
5
Total # of Requests Per Hour
Day
200
250
300
350
Total Traffic Volume Per Hour (GB)
Total Requests
Total Traffic
Fig. 3. Weekly Accesses from Nov. 8 to Nov.
15 2010
III. CHARACTERIZATION OF MOBILE DEVICE
HETEROGENEITY
A. Mobile System Heterogeneity
To provide Internet streaming services to all kinds of mobile
devices like Vuclip, a unique challenge is the heterogeneity
among mobile devices. Different from the pre-coding ap proach
that was taken by many other service providers to serve
specific types of mobile devices, transco ding has to be used
to customize the video into a proper format for the requesting
mobile device. Although transcoding is very flexible and
desirable to serve hetero geneous mobile devices, transcoding
demand s huge CPU cycles on the fly.
To get a realistic picture of the mobile device h e te rogeneity,
we retrieve detailed device information from WURFL [13]
based on the User-Agents information we have extracted from
the server log. Among th e 84,281 User-Agents that represent
mobile devices, we ar e able to get the brand and model
informa tion from more than 74,708 (88.64%) distinct User-
Agent strings. The rest only have browser information.
TABLE II
SYSTEM HETEROGENEI TY OF MOBILE DEVICES
Models 2864
Resolution 92
Mobile OSes 14
As shown in Table II, accesses to Vu clip in these 30
days ca me from 2864 different device models. These devices
have different screen sizes that can support video playb ack
with different resolution rates. Delving into this, we find that
these devices have 92 different resolutions (width a nd height
combinations), ranging from 84 × 48 to 1600 × 1200. Figure
4 shows the most popular resolutions, including 320 × 240,
480 × 360, and 480 × 320. They also r un on 14 different
mobile operating systems.
B. Audio/Video Codec Heterogeneity
To play video on a mobile device, both audio and video
codecs are required. On different devices, the supported codec s
may be different as well. Such heterogeneity would further in-
crease the load for the server if the server conducts tran scoding
for the mobile device. Note that if such transcoding is done
at the client side, it would lead to excessive battery power
consumption.
To examine the codec hete rogeneity, we further look into the
supported aud io/video codecs on th e se 2864 hardware models.
We find that typically there are 3 audio codecs being used,
namely AAC, AM R, and WMA, and there are 4 video codecs
being used, namely H.263, H.264, MPEG-4, WMV. Figures 5
and 6 show the popularity of these codecs. As shown in these
figures, AMR is the most popular audio codec, as more than
59% devices support it, and H.263 and MPEG-4 are the most
popular video codecs.
TABLE III
VIDEO CODECS
Type Video Audio #Videos Sessions
ASF WMV WMA 293,025 2,031,161
3GP H.264 AAC 692,004 12,636,639
3GP H.263 AMR 2,805,494 46,790,565
3GP MPEG-4 AMR 138,022 1,213,319
3GP MPEG-4 AAC 1,762,132 36,552,760
3GP in Total 3,746,548 97,193,283
With 3 audio code cs and 4 video codecs, we expect a total of
12 combination s of different audio/video codecs. In practice,
however, not all these combination of the audio and video
codec are used. In the workload, we only find 5 combina tions.
Table III shows the 5 video+audio encoding scheme s used .
While more than 4 million (4,052,740) unique videos were
accessed, we are able to extract about 3.7 million (3,789 ,229)
that are accessed as videos. The rest were only accessed as
audio. Table III shows that H.263+AMR and MPEG4+AAC
are the most popular encodin g schemes, accountin g for 84%
of total viewing sessions. This is not surprising as H.263 and
MPEG4 are the most widely supported video codecs on the
2864 models of mo bile devices.
In addition to different co decs, vid eo files are also encoded
into two different formats, i.e, two types of container s, 3GP
and ASF. 3GP is the 3GPP file form at, which is a multimedia
container format defined by the Third Ge neration Partnership
Project (3GPP) f or 3G UMTS multimedia services. 3GP is
often used on 3G mobile ph ones. On the other hand, ASF

320*240 480*360 480*320 220*176 Others
0
5
10
15
20
25
30
35
40
% of Accesses
Fig. 4. Most Popular Resolutions
AMR AAC WMA
0
10
20
30
40
50
60
% of Supported Models
Fig. 5. Audio Codecs
H.263 MEPG−4 H.264 WMV
0
10
20
30
40
50
60
% of Supported Models
Fig. 6. Video Codecs
(Advanced Systems/Streaming Format) belongs to Microsoft
Media framework and it is a proprietary digital audio/digital
video con tainer format. Apparently, 3GP is much more widely
used in practice th an ASF for mobile videos.
TABLE IV
VIDEO RES OLUTI ON AND ENCODING RATE
Quality Frame Width Encoding Rate (Kbps)
Low 176 51 - 55
Low 320, 360 71 - 187
High 176 81 - 147
High 320, 360 172 - 335
WiFi 320, 360 358 - 423
Besides the above hard ware and software heterogeneity,
mobile devices may have different network speed, d ue to
various reasons, such as accessing throu gh cellular network
or WiFi. To support different mobile Internet access speed,
Vuclip also tran scodes video clips into 3 different quality
levels: Low Quality, High Quality and WiFi Quality. Ta ble IV
shows the corre sponding r a nge of object encoding rate for
different quality levels. Con sider the variety of re solutions,
videos are also customized into 3 different frame widths: 176,
320, and 360. As we can observe from the table that a larger
resolution (width) vid eo does no t necessarily come with a high
encodin g rate. On the other hand, a video with a high encod ing
rate typically comes with a larger resolution.
TABLE V
VIDEO QUALI TY
Quality # of Videos # of Sessions
Low 1,694,108 26,365,900
High 3,323,211 72,392,729
WiFi 91,761 465,815
Table V further shows the number of videos that are of
Low, High, and WiFi Quality as well as the number of their
requested sessions. The v ideos in High Qu ality are mostly
requested: 73% viewing sessions are for videos encoded
with High Quality. Consider that Vuclip transcodes the video
content on-demand, it is not surprising that 87% of vid e o
contents have at least one versio n encoded with High Quality.
WiFi Quality, however, is the least reque sted quality level. This
is likely due to the relatively slow mobile accessing speed and
tiered data plan billing mod e l today.
Since Vuclip transcodes the original video to acc ommodate
mobile devices with different codec s, fr ame width, and quality
level, for the ease of presentation, we use versions to refer to
different transcoded video files for each video in the rest of
the paper. On the other hand, we use videos to refer to a set
of vide o clips that correspond to the same content. Figure 7
shows the CDF of number of versions each video has. As
shown in the figure, in this workload, about 59% videos have
only one version, and about 3% videos are accessed in 10 or
more versions. The largest version number is 41.
IV. CHARACTERIZATION OF MOBILE STREAMING VIDEOS
The previous section has shown that mobile device hetero-
geneity is a great challenge to the service provider. With such
a level of heterogeneity, what kind of video clips are being
served is of our great interest. In this section, we further
analyze the mobile video clips that we have collected from the
server lo g in order to reveal the com mons with and differences
from the traditional Internet streaming content.
A. Video Playback Duration and File Size
Figure 8 depicts the d istribution of video playback duration
in seconds. In this figure, v ideos a re sorted in decreasing order
of the playback duration, and the y-axis is in log scale. As
shown in the figure, video clips acce ssed by mob ile users are
mostly short in terms of playback duration : more than 97%
videos are less than 10 minutes long, and the median playb ack
duration is 162 seconds (less than 3 minutes). Compared to the
longer duration of traditional Internet stre a ming video clips,
such a shor te r duration m akes it more f easible for mobile
devices because vid eo streaming c onsumes a lot of limited
resources on mobile devices, including the network for data
receiving, the CPU for decoding, and the display for rendering.
Such resource consumption can drain the limited battery power
supply at a very high rate.
Correspon dingly, Figure 9 shows the file size (bytes) distri-
bution. Again, we sort the vide o files (versions) based on their
sizes in decreasing order. As shown in Figure 9, the video
file distribution is similar to that of the duration as shown
in Figure 8. Note that, here in this figure, each video may
have been acc e ssed in several versions in different for mats
and file sizes. As we can see, most video files accessed by
mobile devices are smaller than 8 MBytes, with a mean file
size of 2.78 MBytes and the median file size 1.68 MBytes.
This shows that videos accessed by mobile devices are mostly
small in terms of bytes. This can r e duce the total network

0 10 20 30 40 50
0
0.2
0.4
0.6
0.8
1
Cumulative Fraction
Fig. 7. # of Versions Per Video (CDF)
0 1 2 3 4
x 10
6
10
0
10
1
10
2
10
3
10
4
10
5
Duration (seconds) (log scale)
Fig. 8. Video Playback Duration Distribution
0 2 4 6 8 10
x 10
6
10
2
10
3
10
4
10
5
10
6
10
7
10
8
10
9
File Size (bytes) (log scale)
Fig. 9. Video File Size Distribution
10
0
10
1
10
2
10
3
10
4
10
5
10
6
10
7
10
0
10
1
10
2
10
3
10
4
10
5
10
6
10
7
# of Sessions (log scale)
alpha = 0.955
R
2
= 0.998711
data in log−log scale
Zipf model fit
Fig. 10. Daily Video Popularity Distribution
Fig. 11. Monthly Video Popularity Distribution
transmission for downloadin g the video file. Note the network
interface car d could consume 30% to 40% of the total ba ttery
power consumed during a streaming session to a mo bile
device [14], [15].
Moreover, compared to the size of the traditional Internet
video files [16], [17], the size distribution we find in this server
log is much smaller. This provides a great opportunity for
reducing the transcoding co st as we discuss later in section V.
B. Popularity of Mobile Videos
Figure 10 shows the popularity pattern of videos accessed
site-wide on Nov. 1st. In this figure, the x-axis represents
videos ranked by the number of requested sessions in dec reas-
ing order, plotted in log-scale, while the y-axis represents the
number of viewing sessions of this video, also plotted in log
scale. This figure shows that, in log-log scale, the popularity
distribution of videos accessed can be well fitted with a Zipf-
like distribution
y
i
1
i
α
,
where i is the popularity rank of the video , y
i
is the number
of requested sessions for the video, an d α is the skewness
parameter. Moreover, we find α = 0.955 fits our data very well
with the goodness of fit value R
2
very close to 1, indicatin g
the popularity distribution is not only Zipf-like, but also very
close to the Zipfs law where α = 1. Similar patterns have
been found for other day s in the workload.
The Zipf-like distribution is known to be efficient in mod-
eling web traffic, and is the premise for efficient web c aching.
Specifically, α is an in dicator of request concentration, and
proxy caching can be more efficient with a larger α value.
For example, it was reported in [18] that α varies b e tween
0.64 and 0.83 for web traffic, while it tends to be smaller
for media traffic (for examp le , work [ 17] reports 0.56 for
YouTube traffic). Different from previous me a surement studies
where data were collected at edge networks, the mobile video
accesses a re highly con centrated at the server side. Such
discrepancy is reasonable as collecting traffic at edge networks
can on ly reflect the local users’ accesses, while the server
logs c a n provide a complete view of the video popularity.
Furthermore, this also means caching at the server side is more
effective than caching a t the edge/client side, if caching at the
server side is needed. Note for content delivery, caching at
the server side is typically not for reducing network traffic as
caching at the client side.
While Figure 10 shows short-term (one day) popularity
distribution, Fig ure 11 shows the corresponding distribution in
long-ter m, spanning over the entire 30 days of our workload.
In this figure, the left y-axis is in powered scale wh ile the
right y-ax is is in lo g scale. T he x-axis is in log scale as well.
As shown in the figure, the monthly vid e o popularity deviates
from a stra ight line in log-log scale, meaning not a Zipf-like

Citations
More filters
Proceedings ArticleDOI
07 Jun 2012
TL;DR: This paper proposes and implements "Cloud Transcoder", which utilizes an intermediate cloud platform to bridge the "gap" between Internet videos and mobile devices via its special and practical designs.
Abstract: Despite the increasing popularity, Internet video streaming to mobile devices is still challenging. In particular, there has been a format and resolution "gap" between Internet videos and mobile devices, so mobile users have high demand on video transcoding to facilitate their specific devices. However, as a computation-intensive work, video transcoding is greatly challenged by the limited battery capacity of mobile devices. In this paper we propose and implement "Cloud Transcoder", which utilizes an intermediate cloud platform to bridge the "gap" via its special and practical designs. Specifically, Cloud Transcoder only requires the user to upload a video request rather than the video content. After getting the video request, Cloud Transcoder downloads the original video from the Internet, transcodes it on the user's demand, and transfers the transcoded video back to the user with a high data rate via the intra-cloud data transfer acceleration. Therefore, the mobile device only consumes energy in the last step - fast retrieving the transcoded video from the cloud. Running logs of our real-deployed system confirm the efficacy of Cloud Transcoder.

103 citations


Cites methods from "A server's perspective of Internet ..."

  • ...After getting the video request, Cloud Transcoder downloads the original video from the Internet, transcodes it on the user s demand, and transfers the transcoded video back to the user with a high data rate via the intra-cloud data transfer accel­eration....

    [...]

Journal ArticleDOI
TL;DR: This work designs an online algorithm which can achieve the optimal solution within provable upper bounds of Lyapunov optimization framework and Lagrangian relaxation and aims to minimize the long-term overall cost by determining whether a segment should be cached or transcoded online.
Abstract: Video transcoding in an adaptive bitrate streaming (ABR) system is demanded to support video streaming over heterogenous devices and varying networks However, it could incur a tremendous cost Meanwhile, most viewers terminate viewing sessions within 20% of their durations; only a small fraction of each video is consumed Built upon this user viewing pattern, we propose a Partial Transcoding Scheme for content management in media clouds Particularly, each content is encoded into different bitrates and split into segments Some of the segments are stored in cache, resulting in storage cost; others are transcoded online in the case of cache miss, resulting in computing cost We aim to minimize the long-term overall cost by determining whether a segment should be cached or transcoded online We formulate it as a constrained stochastic optimization problem Leveraging Lyapunov optimization framework and Lagrangian relaxation, we design an online algorithm which can achieve the optimal solution within provable upper bounds Experiments demonstrate that our proposed method can reduce 30% of operational cost, compared with the scheme of caching all the segments

83 citations


Cites background from "A server's perspective of Internet ..."

  • ...[5] investigated a top Internet mobile streaming service at the service side, and the measurements showed that a video content could be transcoded into more than 40 versions to deal with device heterogeneity....

    [...]

  • ...2) Video Popularity Distribution: Most of videos are not frequently requested by users, as the measurements in [7] and [5] show that only 10% of the most popular videos account for almost 80% user requests, and the remaining 90% of video contents account for only 20% of the user requests....

    [...]

  • ...It was demonstrated in [5] that a video content could be encoded into more than 40 versions to meet the requirements of different kinds of user devices and network conditions....

    [...]

Journal ArticleDOI
TL;DR: This paper leverages Media Cloud to deliver on-demand adaptive video streaming services, where those resources can be dynamically scheduled in an on- demand fashion and achieves significant cost savings compared with the existing methods used in content delivery networks.
Abstract: Nowadays, large-scale video distribution feeds a significant fraction of the global Internet traffic. However, existing content delivery networks may not be cost efficient enough to distribute adaptive video streaming, mainly due to the lack of orchestration on storage, computing, and bandwidth resources. In this paper, we leverage Media Cloud to deliver on-demand adaptive video streaming services, where those resources can be dynamically scheduled in an on-demand fashion. Our objective is to minimize the total operational cost by optimally orchestrating multiple resources. Specifically, we formulate an optimization problem, by examining a three-way tradeoff between the caching, transcoding, and bandwidth costs, at each edge server. Then, we adopt a two-step approach to analytically derive the closed-form solution of the optimal transcoding configuration and caching space allocation, respectively, for every edge server. Finally, we verify our solution throughout extensive simulations. The results indicate that our approach achieves significant cost savings compared with the existing methods used in content delivery networks. In addition, we also find the optimal strategy and its benefits can be affected by a list of system parameters, including the unit cost of different resources, the hop distance to the origin server, the Zipf parameter of users’ request patterns, and the settings of different bitrate versions for one segment.

76 citations


Cites background from "A server's perspective of Internet ..."

  • ...Thus, it is necessary to have computing and storage resources to transcode video contents and cache them at some intermediate nodes within the network [4]....

    [...]

Journal ArticleDOI
TL;DR: Among many potential applications of starvation analysis, it is shown how to apply it to optimize objective quality of experience (QoE) of media streaming, by exploiting the tradeoff between startup/rebuffering delay and starvations.
Abstract: Our purpose in this paper is to characterize buffer starvations for streaming services. The buffer is modeled as a FIFO queue with exponential service time and Poisson arrivals. When the buffer is empty, the service restarts after a certain amount of packets are prefetched. With this goal, we propose two approaches to obtain exact distribution of the number of buffer starvations, one of which is based on Ballot theorem, and the other uses recursive equations. The Ballot theorem approach gives an explicit result. We extend this approach to the scenario with a constant playback rate using Takacs Ballot theorem. The recursive approach, though not offering an explicit result, allows us to obtain the distribution of starvations with non-independent and identically distributed (i.i.d.) arrival process in which an ON/OFF bursty arrival process is considered. We further compute the starvation probability as a function of the amount of prefetched packets for a large number of files via a fluid analysis. Among many potential applications of starvation analysis, we show how to apply it to optimize objective quality of experience (QoE) of media streaming, by exploiting the tradeoff between startup/rebuffering delay and starvations.

58 citations

Journal ArticleDOI
TL;DR: This paper proposes an intelligent buffer management strategy called CBM (Cost-aware Buffer Management), for mobile video streaming applications that can provide provably performance guarantee with explicit bounds and experimental results show that CBM achieves significant gains over existing schemes.
Abstract: Mobile video traffic, owing to the rapid adoption of smartphones and tablets, has been growing exponentially in recent years and started to dominate the mobile Internet. In reality, mobile video applications commonly adopt buffering techniques to handle bandwidth fluctuation and minimize the impact of stochastic wireless channels on user experiences. However, recent measurement work reveals that mobile users tend to abort more frequently than PC users during viewing videos. Such a high abortion rate results in a significant wastage of buffered video data, which is directly translated into monetary and energy cost for mobile users. In this paper, we propose an intelligent buffer management strategy called CBM (Cost-aware Buffer Management), for mobile video streaming applications. Our purpose is to minimize cost induced by un-consumed video data while respecting certain user experience requirements. To this objective, we formulate the problem into a constrained stochastic optimization problem, and apply the Lyapunov optimization theory to derive the corresponding online strategy for cost minimization. Different from conventional heuristic-based strategies, our proposed CBM strategy can provide provably performance guarantee with explicit bounds. We also conduct extensive simulations to validate the effectiveness of our proposed strategy and our experimental results show that CBM achieves significant gains over existing schemes.

39 citations


Cites background from "A server's perspective of Internet ..."

  • ...[15] measured several key properties of mobile video streaming systems, including distribution of mobile devices, video length distribution, etc....

    [...]

References
More filters
Proceedings ArticleDOI
21 Mar 1999
TL;DR: This paper investigates the page request distribution seen by Web proxy caches using traces from a variety of sources and considers a simple model where the Web accesses are independent and the reference probability of the documents follows a Zipf-like distribution, suggesting that the various observed properties of hit-ratios and temporal locality are indeed inherent to Web accesse observed by proxies.
Abstract: This paper addresses two unresolved issues about Web caching. The first issue is whether Web requests from a fixed user community are distributed according to Zipf's (1929) law. The second issue relates to a number of studies on the characteristics of Web proxy traces, which have shown that the hit-ratios and temporal locality of the traces exhibit certain asymptotic properties that are uniform across the different sets of the traces. In particular, the question is whether these properties are inherent to Web accesses or whether they are simply an artifact of the traces. An answer to these unresolved issues will facilitate both Web cache resource planning and cache hierarchy design. We show that the answers to the two questions are related. We first investigate the page request distribution seen by Web proxy caches using traces from a variety of sources. We find that the distribution does not follow Zipf's law precisely, but instead follows a Zipf-like distribution with the exponent varying from trace to trace. Furthermore, we find that there is only (i) a weak correlation between the access frequency of a Web page and its size and (ii) a weak correlation between access frequency and its rate of change. We then consider a simple model where the Web accesses are independent and the reference probability of the documents follows a Zipf-like distribution. We find that the model yields asymptotic behaviour that are consistent with the experimental observations, suggesting that the various observed properties of hit-ratios and temporal locality are indeed inherent to Web accesses observed by proxies. Finally, we revisit Web cache replacement algorithms and show that the algorithm that is suggested by this simple model performs best on real trace data. The results indicate that while page requests do indeed reveal short-term correlations and other structures, a simple model for an independent request stream following a Zipf-like distribution is sufficient to capture certain asymptotic properties observed at Web proxies.

3,582 citations


"A server's perspective of Internet ..." refers background in this paper

  • ...For example, it was reported in [18] that α varies between 0....

    [...]

Proceedings ArticleDOI
Meeyoung Cha1, Haewoon Kwak2, Pablo Rodriguez1, Yong-Yeol Ahn2, Sue Moon2 
24 Oct 2007
TL;DR: In this article, the authors analyzed YouTube, the world's largest UGC VoD system, and provided an in-depth study of the popularity life cycle of videos, intrinsic statistical properties of requests and their relationship with video age, and the level of content aliasing or of illegal content.
Abstract: User Generated Content (UGC) is re-shaping the way people watch video and TV, with millions of video producers and consumers. In particular, UGC sites are creating new viewing patterns and social interactions, empowering users to be more creative, and developing new business opportunities. To better understand the impact of UGC systems, we have analyzed YouTube, the world's largest UGC VoD system. Based on a large amount of data collected, we provide an in-depth study of YouTube and other similar UGC systems. In particular, we study the popularity life-cycle of videos, the intrinsic statistical properties of requests and their relationship with video age, and the level of content aliasing or of illegal content in the system. We also provide insights on the potential for more efficient UGC VoD systems (e.g. utilizing P2P techniques or making better use of caching). Finally, we discuss the opportunities to leverage the latent demand for niche videos that are not reached today due to information filtering effects or other system scarcity distortions. Overall, we believe that the results presented in this paper are crucial in understanding UGC systems and can provide valuable information to ISPs, site administrators, and content owners with major commercial and technical implications.

1,713 citations

Proceedings ArticleDOI
24 Oct 2007
TL;DR: This paper presents a traffic characterization study of the popular video sharing service, YouTube, and finds that as with the traditional Web, caching could improve the end user experience, reduce network bandwidth consumption, and reduce the load on YouTube's core server infrastructure.
Abstract: This paper presents a traffic characterization study of the popular video sharing service, YouTube. Over a three month period we observed almost 25 million transactions between users on an edge network and YouTube, including more than 600,000 video downloads. We also monitored the globally popular videos over this period of time.In the paper we examine usage patterns, file properties, popularity and referencing characteristics, and transfer behaviors of YouTube, and compare them to traditional Web and media streaming workload characteristics. We conclude the paper with a discussion of the implications of the observed characteristics. For example, we find that as with the traditional Web, caching could improve the end user experience, reduce network bandwidth consumption, and reduce the load on YouTube's core server infrastructure. Unlike traditional Web caching, Web 2.0 provides additional meta-data that should be exploited to improve the effectiveness of strategies like caching.

990 citations


"A server's perspective of Internet ..." refers background in this paper

  • ...83 for web traffic, while it tends to be smaller for media traffic (for example, work [17] reports 0....

    [...]

  • ...Work [17] examined the traffic characteristics of YouTube at a campus edge network....

    [...]

  • ...Moreover, compared to the size of the traditional Internet video files [16], [17], the size distribution we find in this server log is much smaller....

    [...]

Proceedings ArticleDOI
18 Apr 2006
TL;DR: This study focuses on user behavior, content access patterns, and their implications on the design of multimedia streaming systems, and introduces a modified Poisson distribution that more accurately models the observations.
Abstract: Video-on-demand over IP (VOD) is one of the best-known examples of "next-generation" Internet applications cited as a goal by networking and multimedia researchers. Without empirical data, researchers have generally relied on simulated models to drive their design and developmental efforts. In this paper, we present one of the first measurement studies of a large VOD system, using data covering 219 days and more than 150,000 users in a VOD system deployed by China Telecom. Our study focuses on user behavior, content access patterns, and their implications on the design of multimedia streaming systems. Our results also show that when used to model the user-arrival rate, the traditional Poisson model is conservative and overestimates the probability of large arrival groups. We introduce a modified Poisson distribution that more accurately models our observations. We also observe a surprising result, that video session lengths has a weak inverse correlation with the video's popularity. Finally, we gain better understanding of the sources of video popularity through analysis of a number of internal and external factors.

728 citations


"A server's perspective of Internet ..." refers background in this paper

  • ...presented the access logs of a live VoD system [20], which shows different user and content properties compared to [16]....

    [...]

  • ...Previous studies on a video-on-demand (VoD) system report the daily rate of change in Top-100 videos is less than 15% [16]....

    [...]

  • ...[16] examined server logs of a traditional VoD system with a total of over 6700 unique videos, and analyzed user access patterns, session length, and video popularity....

    [...]

  • ...Moreover, compared to the size of the traditional Internet video files [16], [17], the size distribution we find in this server log is much smaller....

    [...]

Journal ArticleDOI
17 Aug 2008
TL;DR: The challenges and the architectural design issues of a large-scale P2P-VoD system based on the experiences of a real system deployed by PPLive are discussed and a number of results on user behavior, various system performance metrics, including user satisfaction, are presented.
Abstract: P2P file downloading and streaming have already become very popular Internet applications. These systems dramatically reduce the server loading, and provide a platform for scalable content distribution, as long as there is interest for the content. P2P-based video-on-demand (P2P-VoD) is a new challenge for the P2P technology. Unlike streaming live content, P2P-VoD has less synchrony in the users sharing video content, therefore it is much more difficult to alleviate the server loading and at the same time maintaining the streaming performance. To compensate, a small storage is contributed by every peer, and new mechanisms for coordinating content replication, content discovery, and peer scheduling are carefully designed. In this paper, we describe and discuss the challenges and the architectural design issues of a large-scale P2P-VoD system based on the experiences of a real system deployed by PPLive. The system is also designed and instrumented with monitoring capability to measure both system and component specific performance metrics (for design improvements) as well as user satisfaction. After analyzing a large amount of collected data, we present a number of results on user behavior, various system performance metrics, including user satisfaction, and discuss what we observe based on the system design. The study of a real life system provides valuable insights for the future development of P2P-VoD technology.

618 citations


"A server's perspective of Internet ..." refers background in this paper

  • ...conducted a large scale measurement to study the PPLive-based on-demand streaming [23]....

    [...]

Frequently Asked Questions (1)
Q1. What have the authors contributed in "A server’s perspective of internet streaming delivery to mobile devices" ?

In this work, the authors set to investigate the Internet mobile streaming service at the server side. For this purpose, the authors have collected a one-month server log ( with 212 TB delivered video traffic ) from a top Internet mobile streaming service provider serving worldwide mobile users. In this workload, the authors find over 2800 different hardware models with about 100 different screen resolutions running 14 different mobile OS and 3 audio codecs and 4 video codecs. Furthermore, the daily mobile user accesses are more skewed following a Zipf-like distribution but users ’ interests also quickly shift. Considering the huge demand of CPU cycles for online transcoding, the authors further examine serverside caching in order to reduce CPU cycle demand. The authors show that a policy considering different versions of a video altogether outperforms other intuitive ones when the cache size is limited.