end-to-end analysis of distributed video-on-demand systems p. mundur, r. simon, and a. k. sood ieee...
Post on 20-Dec-2015
214 views
TRANSCRIPT
End-to-End Analysis of Distributed Video-on-Demand Systems
P. Mundur, R. Simon, and A. K. SoodIEEE Transactions on Multimedia, Vol. 6, No. 1, Feb 2004
Presented by Ho Tsz Kin10/03/2004
Agenda
Introduction
System Model
Admission Control
Request Handling
Performance Evaluation
Conclusion
Introduction
Distributed VoD architecture with replication
Model and analyze subsystems – server, network, client
Design End-to-end admission control techniques
Request handling Strategies
ObjectiveMinimize the number of blocked requests
Hierarchical VoD ArchitectureMirrored sites
Provide video delivery to many user population
Cluster of video servers
Limited bandwidth
Replicated videoNo contention
Decoding, buffering and display
Contention exists
Streaming Model
Primary service provider
Delivery over local network only
Secondary service providerDelivery over high speed networkRedirecting to other remote sites
Admission control test
Server and Network Model
High capacity and bandwidth disk array Double buffer scheme
Order serving by the disk is not important
(b, r, p, M) Traffic regulatorb: token bucket sizer: token accumulation ratep: peak rate of NICM: max. packet size for the flow
Server and Network Model
(b, r, p, M) Traffic regulatorIn any interval x 0, bytes at its output is min(M+px, b+rx)Control the burstiness of the flow into the network
WFQ schedulingImplemented in the networkProvide a firm per-packet end-to-end delay bound on a per-link and per-routing path
RSVP
Reserve resources along the path of the requests
Admission ControlDerive admission control conditions at the server and networkUsing (b, r, p, M) traffic regulator, and WFQ
b = Br, r = Rr, p defaults infinityMax. bounded end-to-end delay
Retrieval block size
Network reserved rate
Routing path with J links
Max packet size
Overall bandwidth on jth link
Max packet size permitted in network(MTU)
Admission Control
New request will be admitted only if
Reserved rate for new request
1i new
T
r r ji
R R A
Admission control tests run on link by link basis over the routing path
Request Handling
redirectblocked request at one resource is simply redirected to other resourceshigher implementation overheadadditional setup time
split-basedfixed load-sharing among resourcessimpler implementationdifficult to determine efficient splits
Performance EvaluationSimulation Data and Model
Metrics: Blocking rate, Blocking probability
n e tw o rk2n e tw o rk1
L o c a ln e tw o rk
...L o c a l c lu s te r
...
R e m o te c lu s te r1
...
R e m o te c lu s te r2
Performance EvaluationAdmission control test
Request handling policiesRedirect
Local Remote1 Remote2
Request
Blocked
Performance EvaluationSplit-x-y
• First split between local and remote clusters in ratio of x and (100 - x)• Further split between remote cluster1 and cluster2 in ratio of y and (100 - y)
Split-redirect -x-y• Blocked requests in local
are redirected to remotecluster1
• Blocked requests inremote cluster1 are NOTredirected to remotecluster2
Performance EvaluationSingle rate playback service (8Mbps)750 streams by local server, 311 streams per remote clusterReplicated locally (2.5TB storing 800GB)
Performance Evaluation
CrossoverBy proportion of traffic directed toward the local cluster
DivergenceBy proportion of trafficdirected toward the remote traffic
Performance EvaluationEfficient split
Choosing split value that are close to the proportion of resource capacitiesPossible only if the portion of remote cluster capacity is knownIn general, difficult todetermine
One-level redirectionalready achieves better
Performance EvaluationReplication issue1st scenario: single server in the local cluster, top 30 videos stored locally2nd scenario: five servers in the local cluster, top 30 videos stored locally, 5 times replication allowed3rd scenario: five servers in the local cluster, complete video collection stored locally
Performance Evaluation
FairnessOnly partial list of top videos are stored on the local cluster with five video server
Class1: top 20% videos
Class2 & 3: other 80%
Redirect at 1000 requestsper hour
Conclusion & DiscussionAnalyzed VoD system with a hierarchy of servers and network elementsDerived the admission control conditionUsing extensive simulation, designed and evaluated several request handling policiesRedirect will be more suitableDiscussion
QoS guaranteed networksOther network traffic exists, and more remote sitesDifficult to determine the split percentageReliability issue