ZHANG et al.: UNREELING XUNLEI KANKA N: UNDERSTANDING HYBRID CD N-P2P VOD STREAMING 231
TABLE I
E
XISTING MEASUREMENT STUDIES ON CDN/P2P-BASED VODSYSTEMS
B. CDN and P2P-Based Internet VoD Systems
Some measurement studies have been reported on the CDN
or P2P-based streaming systems. We summarize the measure-
ments of five existing system s (YouTube, Netflix, Hulu, PPLive
Vo D
5
and Joost) in the following aspects: 1) infrast ru c tur e;
2) content discovery; 3) content distribution. The comparison
results are presen ted in Ta bl e I.
The CDN-based VoD systems (e.g., You Tube, N et flix and
Hulu) share several comm on features. T hey are supported by
the large-scale commercial CDNs, which deploy thousands of
serversinvariousISPs.TheyrelyontheDNSinfrastructureto
provide request redirection. Videos are delivered based on the
HTTP protocol. Once a CDN server is assigned for a video re-
quest, it becomes the only source during this video streaming
session. Some mechanisms, such as the rate adaptation , are usu-
ally adopted to enhance the streaming performance.
Compared with the CDN-based VoD systems, the P2P-based
systems (e.g., PPLive VoD and Joost) deploy less infrastructure.
Despite the source ser ver s that stor e the original video files, only
dozens of channel servers or P2P trackers are provisioned in the
infrastructure. Trackers are responsible for replying the candi-
date peer swarm to every video request. The video data are usu-
ally delivered using the UDP protocol.
These issues are the keys to design Internet VoD system. In
Section III, we will compare the Kankan syste m with the ex-
isting CDN or P2P based systems in these issues.
C. User Behavior and Content Distribution in VoD Systems
The characteristics of user behaviors are important in evalu-
ating the performance in traditio nal VoD systems. There have
been some previous measurement studies on the user behaviors
in the existing Internet VoD systems, such as PPLive VoD [15],
[22] and YouTube [23]. However, in the P2P-based or hybrid
VoD systems, another important aspect of user beh avio rs is the
video distribution cached in peers. Since the contents in peer’s
cache can be shared for the future requests, the content avail-
ability in p eers plays an important role in a large-scale P2P VoD
streaming system. A detailed characterization of the content dis-
tribution in peer’s caches is helpful to quantify the benefits of
the P2P network in the hybrid VoD systems.
5
PPLive was renamed as PPTV in Dec. 2009. In order to be c ompatible with
the previous literature, we use PPLive in this paper.
To date, there have been few m easurem ent studies on the con-
tent distribution in peer caches except [15]. With the traces from
PPLive VoD, Huang et al. focused on three-level movie objects,
and i nvestigated the number of users owning different segm ents
of these movie objects. By comparing the trace-driven simula-
tion results against t he real request records, they evaluated the
replication policy in PPLive VoD. In Section VI, we wil l in -
vestigate the cached contents in the Kankan P2P network, and
present our analysis the content dynamics in peer caches.
III. K
ANKAN VoD STREAMING SYSTEM
A. Brief Introduction on Xunlei Kankan
Xunlei Limited is on e of the top 10 largest C h inese In-
ternet companies. One Xunlei’s major services is to provide
P2P-based file acceleration for the end-users. The P2P-based
Xunlei accelerator was reported to have close to 300 million
monthly users for the three-mon th period that ended on Mar.
2014 [24]. Another major service provided by Xunlei is the
Kankan VoD service. The Kankan player was initially produced
as a value-added service for Xunlei accelerator in 2007. After-
wards, Kankan became a stand-alone VoD system, which had its
own media player software and web site (w ww.kankan.com ).
According to the iResearch, Kankan was repo rted to have about
31.4 million daily unique users in 2012 [1], which was close to
the 34 m ill ion daily visits to PPLive in 2012 [25].
Despite the popularity of Xunlei and Kankan in China, it has
received relatively little attention in the research community
to date. Unlik e PPLive which was open in providing system
traces [15] or sharing system experience [22], not much about
Kankan has b een disclosed by Xu nlei company itself. On the
other hand, the encrypted proprietary protocols adopted in the
Kankan system make it difficult to reveal the system designs
without reverse engineering.
In our previous work on Xunlei measurement [2], we con-
ducted laborious protocol analysis and understood a large por-
tion of the Xu n lei protocol. We also obtained some preliminary
results on the Kankan P2P network network: 1) The Kankan
client adopts the same Xunlei P2P protocol with Xunlei accel-
erator; 2) 7.4 million unique Kankan peers were measured in a
week for just the TV show category.