Measurement and Analysis of an Internet Streaming Service to Mobile Devices

Measurement and Analysis of an InternetStreaming Service to Mobile Devices

Yao Liu, Student Member, IEEE, Fei Li, Lei Guo, Member, IEEE, Bo Shen, Senior Member, IEEE,

Songqing Chen, Member, IEEE, and Yingjie Lan

Abstract—Receiving Internet streaming services on various mobile devices is getting increasingly popular, and cloud platforms have

also been gradually employed for delivering streaming services to mobile devices. While a number of studies have been conducted at the

client side to understand and characterize Internet mobile streaming delivery, little is known about the server side, particularly for the

recent cloud-based Internet mobile streaming delivery. In this work, we aim to investigate the Internet mobile streaming service at the

server side. For this purpose, we have collected a 4-month server-side log on the cloud (with 1,002 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 3,400 different hardware models with more than 100 different screen resolutions running 14 different mobile OS and three audio

codecs and four video codecs. 2) To deal with the device heterogeneity, CPU-intensive transcoding is used on the cloud 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 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 to reduce the total CPU cycle demand from the cloud. We show

that a policy considering different versions of a video altogether outperforms other intuitive ones when the cache size is limited.

Index Terms—Internet mobile streaming, heterogeneity, popularity, transcoding

Ç

1 INTRODUCTION

RECENTLY, mobile devices are getting increasing popular-ity. For example, according to Canalys, the total number

of smartphones sold worldwide in 2011 is 487.7 million [1],which is a 62.7 percent increase from the previous year (299.7million). By August 2012, 116.5 million people in the USowned smartphones [2].

Besides general web surfing on the Internet, these daysmore and more accesses from mobile devices are directed toall kinds of Internet streaming services. For example,YouTube [3] is among the earliest to provide streamingservices to mobile devices such as iPhone. Today both iOSand Android have native support for YouTube. Otherpopular streaming service providers, including Netflix [4]and Hulu [5], also provide streaming services to subscribedmobile users via APPs built in various mobile operatingsystems. Placeshifting services like Orb [6] and AirVideo [7]allow mobile users to access media content stored on theirhome computers. Qik [8] allows users to upload from

mobile devices and then broadcast the video content to theirfriends. Different from the above services, Vuclip [9] letsusers search and play all kinds of Internet videos on theirmobile devices regardless of their mobile device types.

To understand the key challenges of Internet mobilestreaming and the difference from traditional Internetstreaming, a number of studies have been performed. Astoday the majority of Internet mobile streaming services aredelivered in a client-server architecture, many studies havefocused on the resource consumption and streaming qualityreceived on the mobile device. For example, Xiao et al. [10]studied energy consumption when watching YouTube onmobile devices. Huang et al. [11] investigated fetchingpolicies of different mobile video players, and Finamoreet al. [12] examined the potential causes for inferiorstreaming quality of mobile YouTube accesses.

However, these studies mainly concern about the clientside by examining specific devices [10], [11] or via localexperiments [12]. As the key to the current Internet mobilestreaming delivery services, the server side plays a criticalrole in the entire streaming delivery process. Unfortunately,so far, little is known about the server side, possibly due tothe limited availability of data from the server side. This isparticularly true for the modern cloud-based Internetstreaming delivery services.

To provide in-depth understanding of the current cloud-based Internet mobile streaming services, in this study, we setto investigate the server side in streaming delivery to mobiledevices. For this purpose, we have analyzed a 4-month (fromNovember 2010 to February 2011) server-side workloadcollected from a top Internet mobile streaming serviceprovider. In this workload, there are about 480 million videosessions with about 1,002 Terabyte video traffic delivered.

2240 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 24, NO. 11, NOVEMBER 2013

. Y. Liu, F. Li, and S. Chen are with the Department of Computer Science,George Mason University, 4400 University Drive, MSN 4A5, Fairfax, VA22030. E-mail: {yliud, lifei, sqchen}@cs.gmu.edu.

. L. Guo is with the Department of Computer Science and Engineering, OhioState University, Columbus, OH 43210. E-mail: [email protected].

. B. Shen is with Vuclip, XinLab, Inc., 1551 McCarthy Boulevard, Suite 213Milpitas, CA 95035. E-mail: [email protected].

. Y. Lan is with the Guanghua School of Management, Peking University,Beijing, China 100871. E-mail: [email protected].

Manuscript received 22 June 2012; revised 28 Oct. 2012; accepted 8 Nov.2012; published online 28 Nov. 2012.Recommended for acceptance by D. Xuan.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference IEEECS Log Number TPDS-2012-06-0587.Digital Object Identifier no. 10.1109/TPDS.2012.324.

1045-9219/13/$31.00 � 2013 IEEE Published by the IEEE Computer Society

Through our analysis, we have a number of findings. Whilethe details are presented later in the paper, some highlightsare as follows:

. A unique challenge for Internet streaming deliveryto mobile devices is rooted from the fact that mobiledevices are very heterogeneous. In this workload,we find over 3,400 different hardware modelswith 109 different screen resolutions running 14different mobile OS and three audio codecs and fourvideo codecs. This greatly challenges the traditionalInternet streaming delivery infrastructure where thebottleneck often lies in the limited bandwidth.

. To deal with the device heterogeneity, runtime CPU-intensive transcoding is used by the cloud tocustomize a video to the appropriate versions onthe fly for different devices. A video clip could betranscoded into more than 40 different versions toserve requests from different devices.

. Compared to videos in traditional Internet stream-ing, mobile streaming video clips are typically ofmuch smaller size (with a median of 1.68 MBytes)and the video duration is shorter as well (with amedian of 2.7 minutes). Furthermore, the dailymobile user accesses are more skewed following aZipf-like distribution but users’ interests also shiftquickly, resulting in a stretched-exponential distri-bution in the long term.

To reduce the huge CPU cycles demanded for transcod-ing on the fly, we further explore caching on the cloud bytrading off storage for CPU cycles. Our study shows that apolicy that considers different versions of a video altogetheroutperforms other intuitive ones (e.g., a file based one)when the cache size is limited. As far as we know, we areamong the first to provide a server-side analysis on aVuclip-like Internet mobile streaming service. Our findingsprovide new insights and lay some foundations to improvethe current Internet mobile streaming delivery.

The remainder of the paper is organized as follows: Wedescribe some background and the workload overview inSection 2 and study the device hardware and softwareheterogeneity in Section 3. We examine various mobilevideo properties in Section 4 and study the access patternacross different ISPs in Section 5. We further explore thetradeoff between the storage and the CPU at the server sidein Section 6. Some related work is described in Section 7 andwe make concluding remarks in Section 8.

2 BACKGROUND AND WORKLOAD OVERVIEW

To investigate how current cloud-based Internet streamingservices are delivered to mobile devices, we have collectedserver-side log from private cloud rented by one of the

largest Internet mobile streaming service providers, Vuclip[9]. Vuclip provides mobile users with the search-and-delivery services. It allows users to search for and watchany videos on any video-enabled mobile phones and devices.

Different from many existing services that only providestreaming services to specific mobile devices, with thepowerful cloud platform, Vuclip can serve any type ofmobile devices that are capable of streaming playback.Vuclip allows any mobile user to search for interested videoavailable on the Internet, and transcodes them on-demandand on-the-fly based on the type of the mobile device usingthe cloud CPU cycles. As like most cloud-based services,Vuclip uses a client-server architecture (thus cloud andserver are exchangeable in this paper). To serve differenttypes of mobile devices, Vuclip employs on-demandtranscoding on the cloud. Transcoding is a process toconvert the requested video clip to the appropriate codecs,format, and size at runtime upon a request so that the videocan be properly rendered and played on the requestingmobile device. With elastic cloud resources, Vuclip trans-codes a video into different versions by choosing the bestaudio/video codecs, frame size, frame rate, and qualitylevel combination for the mobile device. According to ouranalysis, each video was accessed in more than twoversions on average (as shown in Tables 1 and 2), and themost popular video was accessed in 44 different transcodedversions (as shown later in Fig. 4).

To deliver video content, Vuclip uses the traditionalclient/server (C/S) architecture. The video file is deliveredvia pseudostreaming over HTTP. That is, when therequested content is available on the server, the client wouldissue an HTTP GET request to download the content. Avideo may be downloaded via several HTTP GET requestswith different partial ranges specified (i.e., range requests).To differentiate video requests from HTTP requests, wedefine a request as a single HTTP transfer between the clientand the server, and a session as the set of requests that areinvolved in downloading an entire video clip.

The 1-month workload we collected is from November1st to November 30th, 2010. In this log, there are about105 million sessions watching more than 4 million differentvideos. There are a total of about 192 million HTTPrequests. The total traffic delivered from the server in these30 days is about 212 TB. Table 1 gives a summary of thisworkload. Note that among all these requests, some arefrom desktop/laptop computers instead of mobile devices.To focus on the requests from mobile users, we differentiatethem in the server log through the User-Agent stringsspecified in each HTTP request. By analyzing the User-Agent, we find there is a total of 150,072 unique User-Agentstrings. Among them, 84,281 (56 percent) represent mobiledevices. However, examining the received requests, we find

LIU ET AL.: MEASUREMENT AND ANALYSIS OF AN INTERNET STREAMING SERVICE TO MOBILE DEVICES 2241

TABLE 1Summary of 1-Month Workload

TABLE 2Summary of 4-Month Workload

most of them come from User-Agent strings representingmobile users: more than 94 percent (181 million out of192 million) requests are from mobile devices.

Please refer to Section 1 and Figs. 1, 2, and 3of the supplementary file, which can be found on theComputer Society Digital Library at http://doi.ieeecomputersociety.org/10.1109/TPDS.2012.324, for howthe accesses change over one day, one week, and one month.

After our paper based on 1-month trace is initiallypublished at [13], we are able to further collect 3 moremonths trace from the Vuclip site. Table 2 gives a summaryof the workload. In this 4-month log, from November 2010to February 2011, there are more than 480 million sessionsaccessing more than 10 million unique videos. More than982 million HTTP requests were served by the server. Atotal of 1,002 TB traffic was delivered from the server. Tosupport heterogeneous mobile devices, videos are served indifferent versions. In the trace, we observed 50 differentformats, and each video is transcoded into 2.39 versions onaverage. As a result, more than 25 million unique files(versions) were requested during the 4-month period.

Please refer to Section 1 and Fig. 4 of the onlinesupplementary file for the hourly access pattern over the4-month period.

3 CHARACTERIZATION OF MOBILE DEVICE

HETEROGENEITY

3.1 Mobile System Heterogeneity

To provide Internet streaming services to all kinds of mobiledevices like Vuclip, a unique challenge is the heterogeneityamong mobile devices. Different from the precodingapproach that was taken by many other service providersto serve specific types of mobile devices, Vuclip employstranscoding, a technique that can customize the video into aproper format for the requesting mobile device at runtime.Although transcoding is very flexible and desirable to serveheterogeneous mobile devices, transcoding demands hugeCPU cycles on the fly, where the support with cloudresources is critical.

To get a realistic picture of the mobile device hetero-geneity, we retrieve detailed device information fromWURFL [14] based on the User-Agents information wehave extracted from the 4-month server log. Amongthe 229,333 User-Agents that represent mobile devices,we are able to get the brand and model information frommore than 196,832 (85.8 percent) distinct User-Agentstrings. The rest only have browser information.

As shown in Table 3, accesses to Vuclip in these 4 monthscame from 3,465 different device models. These deviceshave different screen sizes that can support video playbackwith different resolution rates. Delving into this, we findthat these devices have 109 different resolutions (width and

height combinations), ranging from 84� 48 to 1,600� 1,200.Fig. 1 shows the most popular resolutions, including320� 240, 480� 360, and 480� 320. They also run on14 different mobile operating systems.

3.2 Audio/Video Codec Heterogeneity

To play video on a mobile device, both audio and videocodecs are required. On different devices, the supportedcodecs may be different as well. Such heterogeneity wouldfurther increase the load for the server if the server conductstranscoding for the mobile device. Note that if suchtranscoding is done at the client side, it would lead toexcessive battery power consumption.

To examine the codec heterogeneity, we further look intothe supported audio/video codecs on these 3,465 hardwaremodels. We find that typically there are three audio codecsbeing used, namely AAC, AMR, and WMA, and there arefour video codecs being used, namely, H.263, H.264,MPEG-4, and WMV. Figs. 2 and 3 show the popularity ofthese codecs. As shown in these figures, AMR is the mostpopular audio codec, as more than 57.6 percent devicessupport it, and H.263 and MPEG-4 are the most popularvideo codecs.

With three audio codecs and four video codecs, weexpect a total of 12 combinations of different audio/videocodecs. In practice, however, not all these combination ofthe audio and video codec are used. In the workload, weonly find five combinations. Table 4 shows the fivevideoþaudio encoding schemes used. For the more than10 million (10,779,818) unique videos that were accessedin our 4-month log, Table 4 shows that H.263þAMR andMPEG4þAAC are the most popular encoding schemes,accounting for 79.6 percent of total viewing sessions. Thisis not surprising as H.263 and MPEG4 are the mostwidely supported video codecs on the 3,465 models ofmobile devices.

In addition to different codecs, video files are alsoencoded into two different formats, i.e., two types of


Fig. 1. Most popular resolutions.

TABLE 3System Heterogeneity of Mobile Devices

Fig. 2. Audio codecs.

containers, 3GP and ASF. 3GP is the 3GPP file format,which is a multimedia container format defined by theThird Generation Partnership Project (3GPP) for 3G UMTSmultimedia services. 3GP is often used on 3G mobilephones. On the other hand, ASF (advanced systems/streaming format) belongs to Microsoft Media frameworkand it is a proprietary digital audio/digital video containerformat. Apparently, 3GP is much more widely used inpractice than ASF for mobile videos.

Besides the above hardware and software heterogeneity,mobile devices may have different network speed, due tovarious reasons, such as accessing through cellular networkor WiFi. To support different mobile Internet access speed,Vuclip also transcodes video clips into three differentquality levels: low quality, high quality, and WiFi quality.Table 5 shows the corresponding range of object encodingrate for different quality levels. Consider the variety ofresolutions, videos are also customized into three differentframe widths: 176, 320, and 360. As we can observe from thetable that a larger resolution (width) video does notnecessarily come with a high encoding rate. On the otherhand, a video with a high encoding rate typically comeswith a larger resolution.

Table 6 further shows the number of videos accessedbetween November 2010 and February 2011 that are of low,high, and WiFi quality as well as the number of theirrequested sessions. The videos in high quality are mostlyrequested: more than 73.8 percent viewing sessions are forvideos encoded with high quality. Consider that Vucliptranscodes the video content on-demand, it is not surprisingthat 90 percent of video contents have at least one versionencoded with high quality. WiFi quality, however, is the

least requested quality level. This is likely due to therelatively slow mobile accessing speed and tiered data planbilling model today.

Since Vuclip transcodes the original video to accom-modate mobile devices with different codecs, frame width,and quality level, for the ease of presentation, we useversions to refer to different transcoded video files for eachvideo in the remainder of the paper. On the other hand, weuse videos to refer to a set of video clips that correspond tothe same content. Fig. 4 shows the probability massfunction of the number of versions each video has. Asshown in the figure, in this workload, about 57 percentvideos have only one version, and about 3.3 percent videosare accessed in 10 or more versions. The largest versionnumber is 44.

4 CHARACTERIZATION OF MOBILE STREAMING

VIDEOS

The previous section has shown that mobile deviceheterogeneity is a great challenge to the service provider.With such a level of heterogeneity, what kind of videoclips are being served is of our great interest. In thissection, we further analyze the mobile video clips that wehave collected from the server log to reveal the commonswith and differences from the traditional Internet stream-ing content.

4.1 Video Playback Duration, File Size, and Formats

Fig. 5 depicts the distribution of video playback duration inseconds. In this figure, videos that were accessed inNovember 2010 are sorted in decreasing order of theplayback duration, and the y-axis is in log scale. As shownin the figure, video clips accessed by mobile users aremostly short in terms of playback duration: more than97 percent videos are less than 10 minutes long, and themedian playback duration is 162 seconds (less than3 minutes). Compared to the longer duration of traditionalInternet streaming video clips, such a shorter durationmakes it more feasible for mobile devices because videostreaming consumes a lot of limited resources on mobiledevices, including the network for data receiving, the CPUfor decoding, and the display for rendering. Such resource


TABLE 4Video Codecs (4 Months)

TABLE 5Video Resolution and Encoding Rate

TABLE 6Video Quality (4 Months)

Fig. 4. # of versions per video (PMF).

Fig. 3. Video codecs.

consumption can drain the limited battery power supply ata very high rate.

Correspondingly, Fig. 6 shows the file size (bytes)distribution. Again, we sort the video files (versions) basedon their sizes in decreasing order. As shown in Fig. 6, thevideo file distribution is similar to that of the duration asshown in Fig. 5. Note that, here in this figure, each videomay have been accessed in several versions in differentformats and file sizes. As we can see, most video filesaccessed by mobile devices are smaller than 8 MBytes, witha mean file size of 2.78 MBytes and the median file size1.68 MBytes. This shows that videos accessed by mobiledevices are mostly small in terms of bytes. This can reducethe total network transmission for downloading the videofile. Note the network interface card could consume 30 to40 percent of the total battery power consumed during astreaming session to a mobile device [15], [16].

Moreover, compared to the size of the traditionalInternet video files [17], [18], the size distribution we findin this server log is much smaller. This provides a greatopportunity for reducing the transcoding cost as we discusslater in Section 6.

We then examine how many different versions eachformat accounts for. Fig. 7 shows the result. In this figure,the y-axis shows the total number of versions that are usingeach format in log scale, while the x-axis shows the differentformats ranked by their corresponding number of versions.We find that some formats are significantly more popular

than others. Fig. 8 further shows the popularity distributionof accesses for different formats used by Vuclip. Similar toFig. 7, some formats are in orders of magnitude morepopular than others. During 4 months, accesses for the top-10 formats as shown in Table 7 account for 94 percent oftotal sessions. These results are as expected because of thedifferent popularity of different audio/video codecs sup-ported by different mobile devices.

4.2 Popularity of Mobile Videos

Fig. 9a shows the popularity pattern of videos accessed site-wide on November 1st. In this figure, the x-axis representsvideos ranked by the number of requested sessions indecreasing order, plotted in log-scale, while the y-axisrepresents the number of viewing sessions of this video,also plotted in log scale. This figure shows that, in log-logscale, the popularity distribution of videos accessed can bewell fitted with a Zipf-like distribution

yi /1

i�;

where i is the popularity rank of the video, yi is thenumber of requested sessions for the video, and � is theskewness parameter. Moreover, we find � ¼ 0:955 fits ourdata very well with the goodness of fit value R2 very closeto 1, indicating the popularity distribution is not onlyZipf-like, but also very close to the Zipf’s law where� ¼ 1. Similar patterns have been found for the other daysin the workload.

The Zipf-like distribution is known to be efficient inmodeling web traffic, and is the premise for efficient webcaching. Specifically, � is an indicator of request concen-tration, and proxy caching can be more efficient with alarger � value. For example, it was reported in [19] that �varies between 0.64 and 0.83 for web traffic, while it tendsto be smaller for media traffic (for example, work [18]


Fig. 5. Video playback duration distribution (1 month).

Fig. 6. Video file size distribution (1 month).

Fig. 7. # of versions per format (4 months).

Fig. 8. # of sessions per format (4 months).

TABLE 7Most Popular Formats

reports 0.56 for YouTube traffic). Different from previousmeasurement studies where data were collected at edgelocations (e.g., one university campus), the mobile videoaccesses are highly concentrated at the server side asindicated by the larger � value observed. Such discrepancyis reasonable as collecting traffic at edge locations can onlyreflect the access pattern of users from one specific area(e.g., one university campus), while the server logs canprovide a complete and global view of the videopopularity. Furthermore, the more concentrated accessesalso mean caching at the server side is more effective thancaching at the edge/client side, if caching at the server sideis needed. Note for content delivery, caching at the serverside is typically not for reducing network traffic as cachingat the client side.

While Fig. 9a shows short-term (one day) popularitydistribution, we further examine the distribution of popu-larity over longer term. Fig. 9b shows the correspondingpopularity distribution over 4 months of time. In this figure,the left y-axis is in powered scale while the right y-axis is inlog scale. The x-axis is in log scale as well. As shown in thefigure, the video popularity deviates from a straight line inlog-log scale, meaning not a Zipf-like distribution. Instead,they can roughly be fitted with a stretched exponential (SE)distribution as shown by the left y-axis in powered (by aconstant c) scale [20]. With SE distribution, the rankdistribution function can be expressed as

yci ¼ �a log iþ b:

An SE distribution is fit by several parameters as shown inthe figure. For example, parameter c is also called thestretch factor, which characterizes the median file size ofworkload [20]. It was reported that for media workloadswith a median file size <5 MBytes , the stretch factoris �0:2. Our analysis confirms this with a c of 0.065 and amedian file size of 1.68 MBytes. Parameter a in an SEdistribution increases with the duration of workload aswell as the ratio of media request rate to new content birthrate, and it causes the distribution to deviate from astraight line in log-log scale. Please refer to Section 2.1 ofthe online supplementary file for video popularity dis-tributions over different periods of time.

The stretched exponential distribution has been used

to characterize many natural and economic phenomena,as well as the access patterns of Internet media traffic. It

was shown that under an SE distribution, media caching is

much less efficient than under a Zipf distribution [20].

This poses a new challenge if long-term caching is neededon the server side.

4.3 Popularity of Different Video Versions

As videos are accessed in different versions, we next

examine the popularity distribution of video versions. We

find that in short term, unlike videos, the popularity

distribution is not Zipf-like. However, over longer terms,

the popularity pattern evolves into an SE distribution. For

details, please refer to Section 2.2 of the online supple-

mentary file.

4.4 Popularity Evolution

We further examine how the popularity changes over time.

We find that new video clips are generated at a high rate,

and therefore, user interests shift quickly. For details, please

refer to Section 2.3 of the online supplementary file.

4.5 Correlation between Popularity and VideoLength

We have shown in Section 4.1 that the mobile video files

are often short and the video popularity in a short-termhas a Zipf-like distribution. Thus, one may wonder

whether shorter videos get more accesses. To examine

this, we group video files into 1 minute interval based on

their lengths. Fig. 10 shows the total number of requested

sessions decreases as the video length increases, andvideos shorter than 5 minutes account for about 60 percent

total requests. We further calculate the correlation between

video popularity and video length, and examine if shorter

videos tend to be more popular. Our results show that thecorrelation coefficient is 0.006, which indicates the correla-

tion is weak. We have also conducted similar tests based

on versions, and the results are similar. This indicates the

large percentage of requests for short videos are due to

the large population of short video contents instead ofuser preferences.


Fig. 9. Video popularity distribution.

5 CHARACTERIZATION OF ACCESSES BASED ON

ISPS

In this section, we further examine if the accesses at ISP-

level exhibit similar characteristics as observed at the

server-side.

5.1 Popularity Distribution within ISP

We focus on the top six ISPs that contributed to the most

traffic volume in 1 month. For the ease of presentation, we

anonymize the identities of the ISPs, and refer to them as

ISP #1 to ISP #6 ranked by the amount of their traffic. Fig. 11

shows the popularity of videos accessed on November 1st

across six ISPs. The rest days all exhibit similar popularity

pattern. Despite the fact that the total number of requests

from six ISPs is different, the unique number of videos

accessed is in the same order of magnitude.

It is shown that the access patterns of ISPs #3, #4, #5, and#6 are also Zipf-like, which can be well fitted with R2 veryclose to 1. � values are smaller than as observed site-wide,varying between 0.66 and 0.72, which is close to traces fromweb proxy [19], and higher than observed from other mediasystems [18]. This shows that mobile video accesses are lessconcentrated at ISP side compared to at the server side (seeSection 4.2).

The access patterns of ISPs #1 and #2 cannot be well-fitted with Zipf distribution, because the top videos areaccessed more frequently than the Zipf model predicts.

As one of the concerns about serving videos to mobileusers is the device heterogeneity, we further examine howusers from different ISPs access the videos. Table 8 showsthe average number of versions per video accessed fromeach ISP in one month. We find that versions accessed byeach ISP is smaller compared to site-wide (2.31 versionsper video as we show in Section 3). This indicates thatthere is less diversity within ISP compared to site-wide. Asa result, the daily distribution of version popularity per ISPis very similar to the popularity of video popularity shownin Fig. 11.

5.2 Popularity Evolution within ISP

We find that user interests shift more quickly within ISPsthan site-wide. For details, please refer to Section 3.1 of theonline supplementary file.

6 TRADEOFF BETWEEN CPU AND STORAGE

As aforementioned, to conduct on-demand transcoding toserve all kinds of heterogeneous mobile devices, Vuclip hasto rely on a rented private cloud platform. While the cloudcan provide elastic services with sufficient CPU cycles for


Fig. 10. Percent of sessions for videos of different length (minute).

Fig. 11. Daily video popularity distribution per ISP.

online transcoding, it challenges service providers both

technically and economically with the growing popularity

of Internet mobile streaming services. This is because

Vuclip will be charged more if more CPU cycles have been

used. It is, thus, very desirable to reduce the huge demand

of CPU cycles for such transcoding.In the previous section, we have shown that the mobile

users’ accesses are more concentrated (skewed) than thosein the traditional Internet streaming services. Thus, cachingat the server side, sometimes called reverse caching, couldbe explored to temporarily cache some transcoded objectsso that on-the-fly transcoding would not be necessary ifthe same type of mobile devices access the same video.Such full-object caching is possible for mobile videosbecause mobile video objects are typically smaller with amedian of 1.68 MBytes as we showed before. Note thatdifferent from the traditional caching objectives such asweb proxy caching for reducing the network traffic,caching at the server side here is to reduce the CPU cyclesdemanded for transcoding. That is, a tradeoff between thestorage and the CPU.

On the other hand, we have also shown that mobileusers’ interests shift quickly (Section 2.3 of the onlinesupplementary file). This provides some hints for cachereplacement. Regardless whether the cache is implementedvia disk and/or memory, the cache replacement policy isthe key to the cache performance. Typical cache replace-ment strategies (such as popularity-based policies) maywork, however, the complexity added by Vuclip-likeservices comes from the fact that a video often has multipletranscoded versions. Intuitively, these different versionscould be considered as separate objects in the cache.However, if we consider video popularity, different ver-sions of a same video are internally related. Therefore, wenext explore different replacement strategies via simula-tions for Vuclip-like systems.

6.1 Replacement Strategies

A simple strategy is to ignore the internal relationship ofdifferent versions of a video, and consider each version as adistinct object. Under this assumption, the existing webproxy cache replacement policies can be adopted. Since weare dealing with video objects, we thus first consider aversion-popularity-based replacement policy, in which autility function is defined as the ratio of the version accessnumber to the storage size occupied by that version. Theversion with the least utility is the victim to be purged fromthe cache.

On the contrary, if we consider that different versions ofa video are related because along the diminishing popular-ity of a video, all of its versions may get fewer and feweraccesses, we can also consider a policy in which all differentversions of a video are bundled together as one object in thecache. Taking the popularity of this object as the sum ofpopularity of all its versions, we can design a video-popularity-based replacement policy, in which the entireobject (with all versions) is replaced once it is identified asthe victim based on the utility function defined as the ratioof the total access number to this video to the total size ofoccupied storage.

With the above two strategies, naturally, one maywonder if a hybrid policy could perform better. That is,neither considering the version independently as theversion-popularity-based policy does, nor considering allversions of a video as one object as the video-popularity-based policy does. In a hybrid strategy, a utility functioncan be defined for each video as in the video-popularity-based replacement policy, but the victim is the least popularversion of the least popular video.

6.2 Simulation Results

To study the effectiveness of these different strategies, weconduct trace-driven simulations using the collected work-load to compare their performance. With the accesses ofNovember 1st, Fig. 12a shows that when the cache size issmaller than 27 percent of the total size of accessed objects ofthat day, a video-popularity-based replacement policy canachieve the highest cache hit rate, and save roughly about55 percent CPU cycles, while a version-popularity-basedstrategy performs the worst. This is likely due to the highlyconcentrated video access pattern as shown in Fig. 9acompared to less concentrated version access pattern shownin Fig. 6a of the online supplementary file. The hybrid strategyperforms consistently worse than the video-popularity based


TABLE 8Average # of Versions per ISP

Fig. 12. Cache performance over different time.

one, but still a little better than the version-popularity-basedpolicy. These results are consistent with our analysis on thevideo and version popularity. When the cache size increases,the version-popularity-based policy has more flexibility inchoosing the best version to cache, and thus achieves the bestcache hit ratio among the three.

Fig. 12b further shows the results when one-week tracefrom November 1st to November 8th was simulated. Thecache size percentage used in this simulation is based ontotal accessed objects of November 1st. We find that thecache performance over one week can reach as high as thatof a day. However, the cache performance over a month asshown in Fig. 12c is worse. This is because with an SEdistribution for monthly video accesses, caching is muchless efficient than with a Zipf distribution for daily accesses.

7 RELATED WORK

The Internet has witnessed the sharp increase of Internetvideo traffic in the recent years with all kinds of Internetstreaming systems, such as VoD and Internet P2P-basedstreaming systems. Lots of research has been conducted tostudy these Internet streaming systems. For example, Yuet al. [17] examined server logs of a traditional VoDsystem with a total of over 6,700 unique videos, andanalyzed user access patterns, session length, and videopopularity. Yin et al. [21] presented the access logs of alive VoD system, which shows different user and contentproperties compared to [17]. Krishnappa et al. [22]collected Hulu traffic at a campus edge network, andexamined the potential benefits of performing caching andprefetching at edge networks. For P2P-based streamingsystems, Wu et al. [23] investigated P2P streamingtopologies in UUSee. Huang et al. [24] conducted a largescale measurement to study the PPLive-based on-demandstreaming. More recently, taking advantage of the lowrunning cost and elasticity of cloud computing, videoservice providers today serve videos from multiple datacenters in the cloud. For example, Netflix is running all itsoperations today on Amazon EC2 [25]. YouTube alsoleverages Google’s CDN for video streaming [26].

Along the increasing popularity of user-generated con-tent (UGC), studies have also been conducted to character-ize UGC videos. For example, Cha et al. [27] studied userbehaviors and video popularity of YouTube, and comparedthem with non-UGC content from Netflix. Work [18]examined the traffic characteristics of YouTube at a campusedge network.

With the rapid increase of Internet-capable mobiledevices in recent years, mobile Internet video services andaccesses are surging. A few studies have been conducted toinvestigate the performance of mobile streaming applica-tions. Focusing on the resource utilization for receivingstreaming data on mobile devices, Xiao et al. [10] studiedthe power consumption of mobile YouTube. Finamore et al.[12] collected traffic from several edge locations andstudied the potential reasons for the inferior streamingexperience of mobile YouTube users. Previously, we havealso conducted measurements to study the resourceutilization of different streaming approaches to mobiledevices [16]. Furthermore, to save battery power, we haddesigned and implemented BlueStreaming, a system thatcan leverage low-power of Bluetooth to help P2P streaming

to mobile devices [28]. Collecting data from an ISPproviding cellular data services, Erman et al. [29] analyzemobile video traffic, focusing on different streamingmethods used, prevailing encoding rates, and user beha-viors. However, their study is limited to one ISP. Li et al.[30] collect server-side log of an iOS application that usesHTTP live streaming (HLS) for video accesses, and analyzethe user behaviors and access patterns. Different from theirstudy that focuses on iOS and HLS only, we base our studyon a video service that can be accessed via a much largervariety of mobile devices.

In this study, we have investigated the commons anddifferences of mobile Internet streaming services with/fromthe traditional Internet streaming services. Our studyreveals a critical challenge in Internet mobile streamingservices is the hardware and software heterogeneity ofmobile devices. Our analysis also shows different accesspatterns of mobile videos from traditional Internet stream-ing videos. Furthermore, we have also shown that cachingat the server side with a proper replacement policy cansignificantly reduce the resource consumption for Vuclip-like Internet streaming systems in dealing with hetero-geneity. An earlier version [13] of this manuscript ispublished in the proceedings of INFOCOM 2012.

8 CONCLUSION

The wide adoption of mobile devices in practice has madepervasive Internet streaming possible. With elastic re-sources, cloud platforms have also been increasinglyemployed for Internet streaming delivery. While a numberof studies have been conducted to examine the streamingservices from the client’s perspective, in this work, wehave studied the Internet mobile streaming services fromthe server side via 4-month server-side log collected on acloud from one of the largest Internet mobile streamingservice providers. Through detailed analysis, we haveshown the great hardware and software heterogeneity ofmobile devices, different characteristics of mobile videos,and different user access patterns from those in traditionalInternet streaming services. As the great challenge thatVuclip-like system faces is the huge demand of CPUresources for online transcoding to deal with heterogene-ity, we show that caching at the server side with a properreplacement policy can effectively tradeoff limited storagesize for great savings on CPU cycles. These resultsprovide some basic guidelines for building and optimizingfuture Internet mobile streaming systems.

ACKNOWLEDGMENTS

The authors appreciate constructive comments from anon-ymous referees. This work was partially supported by USNational Science Foundation (NSF) under grants CNS-0746649, CNS-1117300, CCF-0915681, and CCF-1146578.

REFERENCES

[1] “Smart Phones Overtake Client PCs in 2011,” http://www.canalys.com/newsroom/smart-phones-overtake-client-pcs-2011,2013.

[2] “Comscore Reports Aug. 2012 U.S. Mobile Subscriber MarketShare,” http://www.comscore.com/Press_Events/Press_Releases/2012/10/comScore_Re ports_Aug._2012_U.S._Mobile_Subscriber_Market_Share, 2013.

[3] “Mobile YouTube,” http://m.youtube.com, 2013.


[4] “Netflix ,” http://www.netflix.com, 2013.[5] “Hulu,” http://www.hulu.com, 2013.[6] “OrbLive,” http://itunes.apple.com/app/id290195003, 2013.[7] “Air Video,” http://itunes.apple.com/app/id306550020, 2013.[8] “Qik,” http://www.qik.com, 2013.[9] “Vuclip,” http://m.vulip.com/, 2013.[10] Y. Xiao, R. Kalyanaraman, and A. Yla-Jaaski, “Energy Consump-

tion of Mobile YouTube: Quantitative Measurement and Analy-sis,” Proc. Second Int’l Conf. Next Generation Mobile Applications,Services, and Technologies (NGMAST), 2008.

[11] J. Huang, Q. Xu, B. Tiwana, Z.M. Mao, M. Zhang, and P. Bahl,“Anatomizing Application Performance Differences on Smart-phones,” Proc. Eighth Int’l Conf. Mobile Systems, Applications, andServices (MobiSys), 2010.

[12] A. Finamore, M. Mellia, M. Munafo, R. Torres, and S.G. Rao,“YouTube Everywhere: Impact of Device and InfrastructureSynergies on User Experience,” Proc. ACM SIGCOMM Conf.Internet Measurement Conf. (IMC), 2011.

[13] Y. Liu, F. Li, L. Guo, B. Shen, and S. Chen, “A Server’s Perspectiveof Internet Streaming Delivery to Mobile Devices,” Proc. IEEEINFOCOM, 2012.

[14] “WURFL,” http://wurfl.sourceforge.net/, 2013.[15] J. Adams and G. Muntean, “Adaptive-Buffer Power Save

Mechanism for Mobile Multimedia Streaming,” Proc. IEEE Int’lConf. Comm. (ICC), 2007.

[16] Y. Liu, L. Guo, F. Li, and S. Chen, “An Empirical Evaluation ofBattery Power Consumption for Streaming Data Transmission toMobile Devices,” Proc. ACM Multimedia, 2011.

[17] H. Yu, D. Zheng, B.Y. Zhao, and W. Zheng, “Understanding UserBehavior in Large-Scale Video-on-Demand Systems,” Proc. FirstACM SIGOPS/EuroSys European Conf. Computer Systems (EuroSys),2006.

[18] P. Gill, M. Arlitt, Z. Li, and A. Manhanti, “YouTube TrafficCharacterization: A View from the Edge,” Proc. Seventh ACMSIGCOMM Conf. Internet Measurement (IMC), 2007.

[19] L. Breslau, P. Cao, L. Fan, G. Phillips, and S. Shenker, “WebCaching and Zipf-Like Distributions: Evidence and Implications,”Proc. IEEE INFOCOM, 1999.

[20] L. Guo, E. Tan, S. Chen, Z. Xiao, and X. Zhang, “The StretchedExponential Distribution of Internet Media Access Patterns,”Proc. 27th ACM Symp. Principles of Distributed Computing (PODC),2008.

[21] H. Yin, X. Liu, F. Qiu, N. Xia, C. Lin, H. Zhang, V. Sekar, and G.Min, “Inside the Birds Nest: Measurements of Large-Scale LiveVoD from the 2008 Olympics,” Proc. Ninth ACM SIGCOMM Conf.Internet Measurement (IMC), 2009.

[22] D.K. Krishnappa, S. Khemmarat, L. Gao, and M. Zink, “On theFeasibility of Prefetching and Caching for Online TV Services: AMeasurement Study on Hulu,” Proc. 12th Int’l Conf. Passive andActive Measurement (PAM), 2011.

[23] C. Wu, B. Li, and S. Zhao, “Exploring Large-Scale Peer-to-PeerLive Streaming,” IEEE Trans. Parallel and Distributed Systems, June2008.

[24] Y. Huang, T.Z. Fu, D.-M. Chiu, J.C. Lui, and C. Huang,“Challenges, Design and Analysis of a Large-Scale P2P-VoDSystem,” Proc. ACM SIGCOMM, Aug. 2008.

[25] “Netflix in the Cloud,” http://velocityconf.com/velocity2011/public/schedule/detail/17785, 2013.

[26] R. Torres, A. Finamore, M. Mellia, M. Munaf, and S. Rao,“Dissecting Video Server Selection Strategies in the YouTubeCDN,” Proc. IEEE Int’l Conf. Distributed Computing Systems(ICDCS), 2011.

[27] M. Cha, H. Kwak, P. Rodriguez, Y.-Y. Ahn, and S. Moon, “I Tube,You Tube, Everybody Tubes: Analyzing the Worlds Largest UserGenerated Content Video System,” Proc. Seventh ACM SIGCOMMConf. Internet Measurement (IMC), 2007.

[28] Y. Liu, F. Li, L. Guo, Y. Guo, and S. Chen, “BlueStreaming:Towards Power-Efficient Internet P2P Streaming to MobileDevices,” Proc. 19th ACM Int’l Conf. Multimedia (Multimedia), 2011.

[29] J. Erman, A. Gerber, K. Ramakrishnan, S. Sen, and O. Spatscheck,“Over the Top Video: The Gorilla in Cellular Networks,” Proc.ACM SIGCOMM Conf. Internet Measurement Conf. (IMC), 2011.

[30] Y. Li, Y. Zhang, and R. Yuan, “Measurement and Analysis of aLarge Scale Commercial Mobile Internet TV System,” Proc. ACMSIGCOMM Conf. Internet Measurement Conf. (IMC), 2011.

Yao Liu received the BS degree in computerscience from Nanjing University, China, in June2007. She is working toward the PhD degree incomputer science at George Mason University.Her research interests include multimedia sys-tems and peer-to-peer networks. She is astudent member of the IEEE.

Fei Li received the BS degree in computerscience from Jilin University in China, in July1997, and the MS, MPhil, and PhD degrees all incomputer science from Columbia University inFebruary 2002, in May 2007, and in February2008, respectively. He is an assistant professorof computer science at George Mason Univer-sity. He joined George Mason University inAugust 2007. His research interests includeonline and approximation algorithm design and

analysis, and applied algorithms for energy-efficient computing, net-worked systems, and cloud computing.

Lei Guo received the BS degree in spacephysics and the MS degree in computer sciencefrom the University of Science and Technologyof China, in 1996 and 2002, respectively, andthe PhD degree in computer science andengineering from the Ohio State University in2007. After that, he worked in Yahoo! andMicrosoft as a senior member of technical staffon the systems and algorithms of social searchand social networks. He is a research scientist in

Ohio State University. His research interests include big data processingand clouding computing, multimedia systems, mobile computing, socialnetworks, P2P networks, and Internet measurement and modeling. Heis a member of the IEEE.

Bo Shen received the BS degree in computerscience from Nanjing University of Aeronauticsand Astronautics, China, and the PhD degree incomputer science from Wayne State University,Detroit, Michigan. He is now the vice president ofengineering at Vuclip. Before that, he was asenior research scientist with Hewlett-PackardLaboratories. His research interests includemultimedia signal processing, multimedia net-working, and content distribution systems. He

has published more than 50 papers in prestigious technical journals andconferences. He holds seven US patents with many pending. He hasbeen on the editorial board for the IEEE Transactions on Multimediafrom 2006 to 2008. He served as the lead guest editor for IEEE TMMSpecial Section on Multimedia Applications in Mobile/Wireless Context.He also served on program committee for a number of technicalconferences including SIGMM. He is a senior member of the IEEE.


Songqing Chen received the BS and MSdegrees in computer science from HuazhongUniversity of Science and Technology, in 1997and 1999, respectively, and the PhD degree incomputer science from the College of Williamand Mary in 2004. He is currently an associateprofessor of computer science at George MasonUniversity. His research interests include theInternet content deliver systems, Internet mea-surement and modeling, operating systems and

system security, and distributed systems and high-performancecomputing. He is a recipient of the US NSF CAREER Award and theAFOSR YIP Award. He is a member of the IEEE.

Yingjie Lan is an assistant professor at PekingUniversity. His research interests include robustnetworks, competitive analysis of online algo-rithms, operations management, revenue man-agement, and supply chain management.

. For more information on this or any other computing topic,please visit our Digital Library at www.computer.org/publications/dlib.


Documents

Measurement and Analysis of an Internet Streaming Service to Mobile Devices