IEICE TRANS. COMMUN., VOL.E94{B, NO.1 JANUARY 2011
1
PAPER Special Section on Autonomous Decentralized Systems Technologies and Their Application to Networked Systems
Autonomous Node Allocation Technology for Assuring
Heterogeneous Streaming Service Under the Dynamic
Environment
Yefeng LIUy, Tatsuya TSUDAy, Nonmembers, Xiaodong LUy, Member,
and Kinji MORIy, Fellow
SUMMARY In Video-on-Demand (VOD) services, the play-
back continuity is one of the most crucial factors for end-user
to judge service quality. It is even more signicant than the
actual video image quality since new generation VoD users com-
monly have heterogeneous requirements on service according to
their context. Moreover, managing dynamic situations in VoD
service is always challenging, especially given the unpredictable
user preference and network condition involved. In this paper,
i) Autonomous Decentralized VoD System (ADVODS) has been
proposed to satisfy dierent levels of user's demands of service
quality and, ii) the Autonomous Node Allocation Technology
(ANAT) is proposed for assuring service continuity. With the
help of autonomous nodes and mobile agents, ANAT can ap-
plies dierent backup policies to users with dierent Service level
Agreements (SLA), and dynamically update the backup schema
to adapt the changing situations such as various service time or
congestion events. Drawing on the evaluation results this pa-
per shows that proposed system architecture has a better perfor-
mance on streaming service provision and continuity.
key words: Autonomous Decentralized VoD System; Service
Continuity Assurance; Successive Backup Strategy
1. Introduction
With the widespread adoption of high speed bandwidth
access and the rapidly expanded communication tech-
nologies, increasing number of customer have turned
their interests toward demand-oriented information ser-
vice. Video-on-Demand, a paradigm of such services,
is gaining vast popularity nowadays and still on the
rise. In the meanwhile, due to steady growth in both
owner quantity and computing capability of mobile con-
sumer electronics, increasingly out-of-home users would
like to enjoy VoD services through their portable de-
vices. Under such ubiquitous computing environment,
the Service Providers (SP) may opt to oer as hierar-
chical service since their users also have heterogeneous
service requirements.
From both VoD service provider's and consumer's
points of view, service continuance is one of the most
Manuscript received April 25, 2010.
Manuscript revised January 1, 2008.
Final manuscript received January 1, 2008.
yThe authors are with the Department of Computer Sci-
ence, Tokyo Institute of Technology, Tokyo 152-8522, Japan
Email: liu@mori.cs.titech.ac.jp
DOI: 10.1587/transcom.E94.B.1
important factors to judge service quality. Momentary
image impairment is acceptable upon most occasions;
however, sudden halt, pause or stop during playback
will be considerably annoying and result in a low service
Quality of User Experiences (QoE) [5]. Hence in this
paper the therm service quality is dened as service
continuity.
Signicant amount of research has been conducted
to design architecture to provide scalability, QoS and
fault-tolerance to VoD systems. [15][2][3] utilize the
distributed server architectures where multiple servers
were available for user. Capacity of such system can be
easily improved by adding additional servers. But the
centralized management is such systems' bottleneck,
e.g., centralization of server load detection at central
server result in longer response time. Current, only
a few studies [16][17] have investigated the continu-
ous service provision in the presence of server failure.
Bolosky et al. [16] from Microsoft proposed the Tiger
Video Fileserver which uses data mirroring to achieve
fault-tolerance. In their system every stripe unit has
double copies at server nodes, thus in the case of one
server node failure the system can still use the alterna-
tive copy of streaming data.
Unlike previous works [2][3][15], Autonomous De-
centralized VoD System (ADVODS) uses a fully de-
centralized architecture aiming to provide continuous
streaming service under a rapidly changing environ-
ment and to satisfy heterogeneous requirements from
end-user. In ADVODS, video data is encoded into lay-
ered data in dierent qualities. On SP end, with the
help of Mobile Agents, the service data is distributed to
separated nodes located in dierent logic levels. On the
other end, users can obtain the video in required qual-
ity by combining those layered data together. In addi-
tion, Autonomous Node Allocation Technology (ANAT)
is proposed to ensure service continuity under dynamic
situations{dierent service time, network congestion,
node failure, and so forth{via a SLA-centered backup
strategy.
The organization of the paper is as follows. In Sec-
tion 2, the proposed streaming service system architec-
ture is introduced. In Section 3, the service continuity
assurance technology is discussed in detail. In Section
Copyright c
2011 The Institute of Electronics, Information and Communication Engineers

2
IEICE TRANS. COMMUN., VOL.E94{B, NO.1 JANUARY 2011
4, the simulation results are evaluated. Finally, a con-
clusion is given in Section 5.
2. Autonomous Decentralized VoD System
This section introduces the concept, architecture, and
components of ADVODS. The VoD system discussed
in this paper is available as web service on the Inter-
net, providing heterogeneous levels of service quality.
Users request dierent video qualities based on their
own interests. In this paper three qualities are dened,
namely gold, silver and bronze. Based on the fact that
there are more mobile device owners than ber cable
users in nowadays, it is assumed that the majority of
user request bronze-quality video service.
2.1 Layered Video Stream Data
Dividing bit stream into multiple layers already be-
comes common feature of current compression schemes.
Meaning streaming data can be divided into a base layer
alone with one or more enhanced layers. The base layer
data can be independently decoded, provide the min-
imum playable data size but lowest video quality; the
enhanced layer data have to be decoded together with
the base layer data and/or other enhanced layers data
to oer a higher video quality.
Based on the aforementioned assumption of exist-
ing three service layers, we also divide the video data
into three layers: Base data (Base), Enhancement1 data
(Enh1) and Enhancement2 data (Enh2). In this paper,
Base is provided to bronze-quality service, Base plus
Enh1 is provided to sliver-quality service, and add both
Enh1 and Enh2 to Base provides the highest denition
video to gold-quality service.
Fig. 1: ADVODS and its comparison with conventional architec-
ture in storage volume
2.2 ADVODS Architecture
The inspiration of ADVODS is from both Autonomous
Decentralized System (ADS) [1][6][7] and Faded Infor-
mation Field (FIF) concept [8][9][10][11], and addition-
ally takes advantage of the layered video stream data
structure and SLA business model. The design in-
tention of ADVODS architecture is to oer a layered
on-demand streaming service. Mobile agent technol-
ogy [12] is involved in order to autonomously allocate
service data for SP, quest demanded video for users,
and search for backup node. In ADVODS, SP gener-
ates Push Mobile Agents (Push-MAs) which initially
carry all layered video stream data. Then, Push-MAs
travel from SP to adjacent nodes and try to allocate
the higher layer data. Based on local situation these
nodes autonomously decide whether to store the data.
After that, Push-MAs remove, if any, stored data, and
move on to the neighbor nodes of current node. By
repeating the same process, Push-MAs can eventually
distribute the service data, from highest layer to the
base layer, to the nodes in the surrounding area of SP.
As can be seen from Figure 1, in the end a multi-level
structure is constructed, where at each level there are
always nodes with the same data. Our study result,
which is presented in later section Evaluation-I, indi-
cates when layers number is three, system can obtain
the best reliability. Furtermore, it is guaranteed that
no same data stored in dierent levels. No redundant
data in dierent level is an additional advantage of AD-
VODS architecture, the total storage volume is reduced
in comparison with the conventional system.
In terms of user end, service consumers need to
receive Base and/or enhanced layers streaming data in
order to see a video clip. Pull Mobile Agents (Pull-
MAs) are generated by users and move to edge nodes
with user's information including user's address and
requests. When Pull-MA arrives an appropriate edge
node then the Base data is available to the user. The
mobile agent will check if the user's requirements are
satised afterwards. If not, Pull-MA will keep on mov-
ing towards upper level nodes for higher quality video
data. This procedure continues until all user demands
are met or the Pull-MA reaches SP. The nodes se-
lected by Pull-MA will start provide services to the
user as soon as Pull-MA leaves. Here one fact should
be remained, that is nodes in dierent level also stores
stream data in dierent layer, so a user requests video
quality higher than Base will always be satised by
video streaming from multiple nodes from more than
one level.
A prototype implementation is developed, which
consists of a layer encoder and a layered-data-
processable multimedia player. The encoder encodes
MEPG-4 data into the three layers and, the video
player can not only decode Base data but also com-

LIU et al.: AUTONOMOUS NODE ALLOCATION TECHNOLOGY
3
Fig. 2: An example of video quality transition
bine data of multiple layers. Figure 2 gives an example
to demonstrate how the system overcomes serving node
failure scenario by dynamically adjusts the combination
of layered data.
2.3 Subsystems
ADVODS consists of four autonomous subsystems:
Push-MA, Pull-MA, Node and DISC-MA.
 Push-MA: SPs create push mobile agents, and del-
egate them with the task of allocate information
(or also known as the information fading). Push-
MA autonomously make the decision of which data
to carry, to place or to delete according to SP's re-
quirements of reliability, node's local status as well
as the popularity and importance of the informa-
tion.
 Pull-MA: Users generate pull mobile agents. A
Pull-MA has the knowledge of a specic user's de-
mand, and responsible for searching the subscribed
data. It autonomously coordinates with nodes to
step-by-step navigates its way to the demanded re-
source.
 Node: Node is an agent execution platform and
major components of system. Node keep a check
on the system's condition by analyzing their local
information, and cooperate with Pull-MA, Push-
MA, and DISC-MA to support their assignments.
 DISC-MA: Discovery Mobile Agent is employed by
user to search for the most appropriate backup
node within a limited area. Details of the func-
tionality of DISC-MA are given in the next section
as part of ANAT.
Each of the subsystems has the autonomous con-
trollability and coordinability to control itself and to
cooperate with the other subsystems under an evolving
environment.
3. Autonomous Node Allocation Technology
In reality, failures may occur at any moment and any
part of system. When failure happens on serving nodes,
it will cause decrement of video quality, or even worse,
stop the playback. Autonomous Node Allocation Tech-
nology (ANAT) is proposed to overcome this issue thus
ensures service continuity. This backup strategy is so-
called a successive operation because it is a hybrid be-
tween static and dynamic process, where system itself
autonomously makes runtime decision of which mode
to active according to real-time network conditions.
Failure occur on dierent layers result in dierent
aections, hence can be further classied into two types:
rst, the failed node is at non-Base layers. In this case,
user lost part or all of the enhanced streaming data, as a
result the video quality decreases after buered data is
used up, whereas the service can continue, and system
can recover from failure easily by asking other higher
level node to carry on the serving, because system still
holds the user's information at least in the Base layer
node. The other type of failure, when the fail node
is a Base node and providing bronze-quality streaming
service, is more dicult to deal with. In such situation
system is hard to be recovered since only the failed node
has the user information. This paper puts main focus
on the second type of single failure.
Two distinct levels of user class are dened by as-
signing with corresponding SLA: normal and VIP. The
solutions for dierent users are various. For VIP class,
the redundant hot-standby node mechanism is utilized.
This technology performs well in general scenarios, but
the cost is expensive and may cause network congestion
if the amount of user becomes large. This in turn im-
plies that this 1+1 redundant node strategy is unsuit-
able for normal class user. However, if the alternative
node notion is not exploited, the system has to nd an-
other server node after the failure which will delay the
recovery time and force user to replay video clip from
the very beginning.
As aforementioned, system is unable to recover
from Base node failure event due to lost of user informa-
tion. Therefore, for the purpose of ensuring service con-
tinuity, it is necessary to maintain an alternative node
which contains user information. In ANAT, a N+1
backup node strategy is employed by normal members.
This N+1 backup technology uses reservation mecha-
nism to nd backup nodes before failures actual happen
in order to have a shorter recovery time. When failure
is detected by system the registered backup node will
continue to provide service to the aected normal mem-
bers. Compared to 1+1 redundancy solution, reserva-
tion approach provides a lower service continuity level
but create less overhead trac and signicantly reduce
nodes' load since the reservation can be taken for multi-
ple user per Base node. ANAT is a combination of two
sub-technologies, Autonomous Node Search as well as
Autonomous Node Update. They are introduced in fol-

4
IEICE TRANS. COMMUN., VOL.E94{B, NO.1 JANUARY 2011
lowing subsections.
3.1 Autonomous Node Search Technology
Discovery Mobile Agent (DISC-MA) is employed to
search an area within a certain distance from start
node. The mobile agents are able to change between
two states: Search and Reply. Agents decide their
own search route, recorded information, and maximum
search area. Nodes can therefore not only get infor-
mation about other nodes from passed agents, but also
provide such data to other agents which arrive after-
wards.
As the goal of this technology is to guarantee ser-
vice continuity, the user buer size logically becomes
the bottom line of recovery time. This means the se-
lected backup node should located within a reasonable
distance from primary node, otherwise the backup will
not be able to carry on service before user run out of
buered data. Based on such principle, a corresponding
timeout schema is applied, which aiming to limited the
search time and area in an appropriate scope and toler-
ant agent loss. The work-
ow of backup node searching
is introduced below. In this paper we assume that the
processing time of DISC-MA at a node and the travel
time of agent from Base node to user is ignorable:
i Base node starts searching process as soon as it be-
gins to serve an user, for example, in Fig.3 node S
begins the search by generating search-status DISC-
MAs and send them to all neighboring nodes. More-
over, a timer is set by start node with the value
Timeoutori = BufferSize 2 (1)
where Timeoutori is Timeout of the agents (at start
node), Buffer is buer size of end user. Mobile
agents also record timeout value. If no agent comes
back by timeout, start node restarts the searching
but accept any delayed reply. In case of no valid
response for three searches, Base node gives up dis-
covering.
ii DISC-MAs move back and forth between the Base
layer and Enh1 layer in order to nd a vacant Base
node within a limited distance from user. As shown
in Fig.3, when agent arrive a node with multiple out
going links, more DISC-MAs are generated. Every
time an agent passes by, node set a timer for that
agent with timeout value
Timeout = TimeoutL TravelL>T  2 (2)
where Timeout is the timeout for this node,
timeoutL is the timeout value at previous node, and
TravelL>T is the travel time from last node to this
node. The latter two parameters are provided by
DISC-MA.
Fig. 3: Autonomous Node Search - Search Stage
iii DISC-MAs record information of Base node they
have passed, and share this information to other
nodes along their routes. Nodes keep such infor-
mation until the same agent comes back or time-
out alarms. By doing so, an Agent-Node-Agent co-
operation is established. When other agents from
dierent directions arrive this node, they can also
share such information from node, and make its own
searching more ecient.
iv DISC-MAs stop in three situations: a) no further
out going link, e.g., agent a in Fig.3, b) total travel
time equals to user buer size, e.g., agents b and e in
Fig.3, or, c) agent arrives a node has been passed by
other DISC-MA (of same user), e.g., agents c and d
in Fig.3. The logical topology of DISC-MAs' routes
will ultimately become a Directed Acyclic Graph
(DAG) [13].
v Figure 4 demonstrates how nodes and agents be-
have after agents stopped. DISC-MA rst changes
its status to Reply and then moves same path back
towards where it was generated, decides the best
candidate backup node along its route based on
recorded information and make a tentative reser-
vation with it.
vi Every node has to wait for all Replay DISC-MAs (it
has sent out) come back and selects the best can-
didate node among these agents. Then, this node
forward one agent with the best result, and ask all
other agents to cancel their tentative reservations.
vii By repeating this process, the Base node who
started the searching can eventually decide the best
backup node and make a ocial reservation with it.
This alternative node search strategy ensures the
best candidate within a reasonable search range can
be found. There is one shortage, when single user is
searching for alternative node, the procedure may be

LIU et al.: AUTONOMOUS NODE ALLOCATION TECHNOLOGY
5
Fig. 4: Autonomous Node Search - Reply Stage
costly in response time owing to the waiting of Re-
ply DISC-MA. However, if we consider the situation
when multiple user start discovery in same time pe-
riod, the eciency of this method increase rapidly as
the number of requester grows because, with the help
of node-agent cooperation, the information from dier-
ent part of network can be shared among neighboring
subsystems, thus discovery mobile agents are able to
get information of whole discovery area without actu-
ally collecting data from every individual node within
its search area.
3.2 Autonomous Node Update Technology
When Base nodes sucient in free capacity, the backup
performs in a static way { user's backup is handled by
the selected alternative node. On the contrary, in the
case of some Base nodes become lack of capacity, a
dynamic update policy is required.
3.2.1 Priority of SLA
We rst divide users into two group, one consists the
ones who have already started service and other one
includes the new users who have not. Base on the
assumed goal (service continuity should be rst guar-
anteed rather than other parameters, say, video start
delay), the former group's user always have higher pri-
ority level than user in other group.
In addition, a Base node is possible to have four
types of Service Level Agreement (SLA) [14]: vip,
vipdual, normal, and normaldual, thereinto, vipdual and
normaldual are respectively the VIP user served by this
Base node as redundant, and the normal user served by
this Base node as backup. Each type of SLA owns a
sequential ordered priority:
vip > normal > vipdual > normaldual
The order is decided with respect to service continuity
assurance, e.g., when a vipdual requests a full load Base
node which is providing service to vip and normal, this
vipdual user (rather than the normal) has to search
for other server node because the serving continuity is
aected if the normal user lose its server node.
3.2.2 Dynamic Update
Backup update is triggered by four situations: rst,
when a Base node's free capacity became lower than
certain level; second, when a node crashed and its user's
original backup nodes have already become their pri-
mary serving node; third, when a node's backup node
crashed; and last, when a normal backup starts to pro-
vide service, all other normal backup user on the same
node have to nd another backup node due to N+1
backup strategy. With regards to node's condition and
user's priorities, a case-by-case discussion is presented
in this section.
If node's capacity lower than a decided bottom
line but not fully consumed, i) any further user request
which has lower priority than the lowest priority user
on this node will be refused, ii) any further user request
which has has an equal priority compared to the low-
est priority user on this node will be accepted, iii) any
further user request which has a higher priority will be
accepted, meanwhile the node will force its lowest pri-
ority user (if more than one such kind of users, choose
the user who has the least remaining playback time)
to leave by dispatching Re-DISC MA. Re-DISC MA
works the same way as DISC-MA, the only dierence
is such agent report search result to both start node
and the Base node of the user who was forced to leave.
When received result, start node deletes the lowest pri-
ority user's data; and the user updates its backup node
information.
On the other hand, if a node is currently working
on full load, i) any further user request which has a
lower or equal priority than the user with lowest priority
on this node will be refused, ii) any further user request
which has a higher priority will be accepted, the Re-
DISC MA for lowest priority user is assigned but the
result is not expected by this node anymore, the user's
data will be deleted at once.
Above-mentioned cases covers all possible situa-
tions, each user is assured of a corresponding service
continuity level according to the SLA-based update pri-
ority.
4. Evaluations
This section introduces evaluation result of proposed
system and technology.
4.1 Evaluation - I
The optimal number of layers of system is concerned

6
IEICE TRANS. COMMUN., VOL.E94{B, NO.1 JANUARY 2011
and evaluated as there is a clear trade o between re-
liability of two mobile agents and reliability of nodes.
The reliability of system is dened by four factors: the
reliability of the Rush-MA, the Pull-MA, distributed
video data and the node.
 RPush(Reliability of Push-MA) is the probability
that Push-MA allocates the layered video stream
data in a node and it is kept from corruption. It
depends on the volume of video data:
RPush(v) = 
v (3)
 is reliability of video per gigabyte, v is the data
volume (GB).
 RPull(Reliability of Pull-MA) is the probability of
a Pull-MA reaches a node which stores the re-
quested video data:
RPull(v) = 
s (4)
 is the per-step Pull-MA reliability, and s is the
number of Pull-MA steps.
 RData(Reliability of distributed video data) is the
probability of successfully streaming the data from
a node to the user:
RData(v) = 
v (5)
 RNode(Reliability of Node) is the probability of
nodes in operation:
RNode =
n1X
k=0

n
k

nk(1
)k (6)
is the reliability of each one, and n is the number
of each user's available nodes that includes replace-
ment nodes.
Therefore the reliability of layered video stream
data is
RBase;Ehn1;2 = RPush
RPull
RData
Rnode (7)
Eventually, the service reliability of each quality
video service is the composed of reliability of each lay-
ered video data:
 Service reliability of high-quality service:
RH = RBase
REhn1
REhn2 (8)
 Service reliability of middle-quality service:
RM = RBase
REhn1 (9)
 Service reliability of low-quality service:
RL = RBase (10)
The initialization and result of evaluation is shown
in Figure 5, which indicates three layers architecture
Fig. 5: Service reliability
can oer our system the best reliability.
4.2 Evaluation - II
A simulator is developed to oer a platform for au-
tonomous streaming service provision and utilization.
The proposed technologies are integrated in the simu-
lator; their functionality and eectiveness are veried
using extensive tests.
The overall goal for the experimental scenario is:
evaluate the service continuity level of normal members
with the condition of keeping service continuity of VIP
member non-impact.
4.2.1 Simulation Scenario
In simulation, system recovers from node failure with
the adoption of selected backup nodes. Following oper-
ation sequence amplies how system behaviors in this
simulation.
Every node broadcasts a periodic message with
content code Survival Signal to all connected nodes to
report its alive status. Consequently, adjacent nodes of
a fail node can aware the failure event by not receive
Survival Signal from this node after a certain period of
time Td. It is an art in selecting the value of Td, overall
recovery time can be reduced by limiting Td, whereas
more trac is created, and vice versa. However, this
trade-o is not detailed discussed in this paper since it
is not related to the main topic of our paper. The neigh-
boring node, which detected the failure, broadcasts a
Failure-detected message. Any Base node receives the
Failure-detected message will search its reservation ta-
ble to check whether itself is responsible for the faulty
node. If the failed node is listed in this node's reser-
vation table, it sends Delivery-start message to user
and continues to provide service, and starts re-discovery
process for other normal backup users in its own reser-
vation table.
4.2.2 Initialization
The system parameters and their initial values are

LIU et al.: AUTONOMOUS NODE ALLOCATION TECHNOLOGY
7
listed in table1.
Table 1: PARAMETERS FOR TESTING
Parameter Description Value
NBase Base node number 60
NEnh1 Enh1 node number 30
NEnh2 Enh2 node number 15
Capacity Maximum delivery number 30
Tprocess Processing time 0.4
Ttrans Communication time 0.1
Tbuffer Buer length 15
In this simulation, users randomly decide their tar-
get nodes, and simulator will trigger node failure when
the system becomes stable. Here we dene the service
continuity of normal user as below
ServiceContinuity =
NonstopNormal
Normal
(11)
NonstopNormal is the number of normal members,
whose service is continuously provided, Normal is the
total number of normal members.
4.2.3 Simulation Results
The simulation results of systems without and with pro-
posed technology are illustrated in Figure 6 and Figure
7, respectively.
We observe from Figure 6 that, without proposed
technology, when total number of user is more than
1400, and around 20% of them are VIP class user, ser-
vice continuity of normal user drops to lower than 40%.
And service continuity becomes lower than 25% when
there are 60% out of 1300 user are VIP member. The
service continuity is guaranteed when user number be-
low 800.
Fig. 6: Simulation result of system without proposed technology
Oppositely, as can be found from Figure 7, pro-
posed technology has made an obvious improvement
on service continuity when the total number of users
higher than 800. Service continuity can stay around
60% when VIP ratio is 20% and number of user rises to
1600. The service continuity is (basically) guaranteed
when user number under 1000.
Fig. 7: Simulation result of system with proposed technology
The evolution result of proposed system shows that
when the total number of user increases, the rapid de-
creasing of service continuity for both VIP and normal
members is prevented. As seen by evolution, ADVODS
with proposed technology is proved to be better perfor-
mance in service provision and continuity over conven-
tional system.
5. Conclusion
Based on the analysis of demand-oriented streaming
service, Autonomous Decentralized VoD System has
been presented to provide continuous service and sat-
isfy heterogeneous requirements from end user under an
ubiquitous environment. In ADVODS, layered stream
data structure and Service Level Agreement concept
are utilized. In consequence, streaming service provi-
sion is optimized. Architecturally speaking, in multi-
tier streaming system the Base is the most important
layer to guarantee service continuity. Therefore Au-
tonomous Node Allocation Technology with a succes-
sive backup schema is proposed in this paper, in order
to ensure the accessibility of Base nodes. In ANAT,
VIP members are allowed to have 1+1 redundant hot
stand by nodes at Base layer. Unfortunately, however,
this solution is not provided to normal class user as
it may cause network congestion in whole system. In-
stead, normal members exploit DISC-MA to reserves
N+1 alternative nodes. Moreover, an alternative node
update schema is proposed, which can autonomously
adjust system conguration according to current net-
work conditions for adapting the constantly changing
environment.
Drawing on the evaluation on developed prototype
system, it is proved that an improvement is made by
proposed system and technology in service continuity.
References
[1] K. Mori, Autonomous Decentralized Systems: Concept,
Data eld Architecture and Future Trends, In Proceedings
of IEEE International Symposium on Autonomous Decen-
tralized Systems (ISADS'93), pages 28-34, Kawasaki, Japn,
March 1993.
[2] J.Y. B. Lee, Parallel video servers: A tutorial, IEEE Trans-
actions on Multimedia, 5 (2): 20-28, 1998.
[3] S.H. G. Chan, Distributed servers architecture for networked
video services, IEEE/ACM Transactions on Networking, 9

8
IEICE TRANS. COMMUN., VOL.E94{B, NO.1 JANUARY 2011
(2): 125136, 2001.
[4] R. Younkin, A. Corriveau, P. Doherty, R. Salskov, Assessing
the Quality of User Experience., Intel Technology Journal,
February, 2007.
[5] I.L. Yen, R. Paul and K. Mori, Towards integrated methods
for high-assurance systems, IEEE Transactions on Comput-
ers, 31 (4): 32-34, April 1998.
[6] Mori, K., Yamashita, S., Nakanishi, H., Hayashi, K.,
Ohmachi, K., Hori, Y, Service accelerator (SEA) system for
supplying demand-oriented information services, In Proceed-
ings of the 3th International Symposium on Autonomous
Decentralized Systems (ISADS 1997), pages 129-136, Berlin,
Germany, 1997.
[7] K. Mori, Autonomous fading and navigation for information
allocation and search under evolving service system, In Pro-
ceedings of 1999 Asia Pacic Symposium on Information
and Telecommunication Technologies (APSITT '99), pages
326-330, August 1999.
[8] K. Mori, Autonomous decentralized systems: concept, data
eld architecture and future trends, In Proceedings of the
International Symposium on Autonomous Decentralized Sys-
tems (ISADS '93), pages. 28-34, 1993.
[9] X.D. Lu and K. Mori, Autonomous Decentralized VoD Sys-
tem Architecture and Fault-tolerant Technology to Assure
Continuous Service, In Proceedings of the 29th International
Conference on Distributed Computing Systems Workshops
(ICDCSW), Montreal, Canada, June 2009.
[10] X.D. Lu and K. Mori, Autonomous Decentralized VoD Ar-
chitecture to Achieve Service Assurance, In Proceedings of
the 28th International Conference on Distributed Comput-
ing Systems Workshops (ICDCSW), pages 557-562, Beijing,
China, June 2008.
[11] J. E. White, Mobile Agents, AAAI Press, The MIT Press,
1996.
[12] Bang-Jensen, Jrgen, 2.1 Acyclic Digraphs, Digraphs: The-
ory, Algorithms and Applications, pages 3234, Springer
Monographs in Mathematics (2nd ed.), Springer-Verlag, 2008
[13] A.N. Hiles, Service Level Agreements: Panacea or Pain?
The TQM Magazine 6 (2): 14-16, 1994
[14] H. Ma and K.G. Shin, Multicast video-on-demand services,
SIGCOMM Computer Communication Review, 32 (1): 3143,
2002.
[15] W. Bolosky, J. Draves, R. Fitzgerald, G.Gibson, M. Jones,
S. Levi, N. Mayhrvold, and R. Rashid, The tiger video le-
service, In Technical Report MSR-TR-96-09, Microsoft Re-
search, 1996
[16] T. Anker, D. Dolev, and I. Keidar, Fault tolerant video on
demand services, In Proceedings of the 19th International
Conference on Distributed Computing Systems (ICDCS '99),
pages 244-252, Austin, TX, USA, 31 May - 4 June, 1999.
Yefeng LIU joined Dept. of Com-
puter Science, Tokyo Institute of Tech-
nology as Ph.D candidate in 2009. He
received the B.S from University of Elec-
tronic Science and Technology of China
in 2007, and M.Sc from Chalmers Univer-
sity of Technology, Sweden in 2009. His
research interests include ubiquitous com-
puting system and distributed systems.
He is student member of IEEE and SIAM.
Tatsuya TSUDA joined Dept. of
Computer Science, Tokyo Institute of
Technology as Master candidate under su-
pervision of Prof. Kinji Mori in 2007. His
research interests include autonomous de-
centralized systems and distributed appli-
cation. He is student member of IEEE.
Xiaodong LU received the B.S.
degree in 1997, the M.S. degree in
2000, both from Lanzhou University,
P.R.China, and Ph.D degree from Tokyo
Tech, in 2005. Now he is an assistant pro-
fessor in the Dept. of Computer Science
at Tokyo Tech. His research interests in-
clude distributed and high-assurance in-
formation systems, and mobile agents. He
is member of IEEE, IEEE CS and IEICE.
Kinji MORI received the B.S., M.S.
and Ph.D degrees from Waseda Univer-
sity, Japan in 1969, 1971 and 1974 respec-
tively. From 1974 to 1997 he was in Hi-
tachi, Ltd. Then he joined the Dept. of
Computer Science of Tokyo Tech as a pro-
fessor. His research interests include the
distributed computing and fault-tolerant
computing. He is fellow of IEEE and IE-
ICE, member of JPSJ and SICE, Japan.