acm nossdav 2008 - kalman graffi - load balancing for multimedia streaming in heterogeneous...
DESCRIPTION
Multimedia streaming of mostly user generated content is an ongoing trend, not only since the upcoming of Last.fm and YouTube. A distributed decentralized multimedia streaming architecture can spread the (traffic) costs to the user nodes, but requires to provide for load balancing and consider the heterogeneity of the participating nodes. We propose a DHT-based information gathering and analyzing architecture which controls the streaming request assignment in the system and thoroughly evaluate it in comparison to a distributed stateless strategy. We evaluated the impact of the key parameters in the allocation function which considers the capabilities of the nodes and their contribution to the system. Identifying the quality-bandwidth tradeoffs of the information gathering system, we show that with our proposed system a 53% better load balancing can be reached and the efficiency of the system is significantly improved.TRANSCRIPT
KG_0805_ACM_NOSSDAV_Presentation_06.ppt
KOM - Multimedia Communications LabProf. Dr.-Ing. Ralf Steinmetz (Director)
Dept. of Electrical Engineering and Information TechnologyDept. of Computer Science (adjunct Professor)
TUD – Technische Universität Darmstadt Merckstr. 25, D-64283 Darmstadt, Germany
Tel.+49 6151 166150, Fax. +49 6151 166152 www.KOM.tu-darmstadt.de
© author(s) of these slides 2008 including research results of the research network KOM and TU Darmstadt otherwise as specified at the respective slide
httc – Hessian Telemedia Technology
Competence-Center e.V - www.httc.de
Dipl.-Math., Dipl.-Inform. Kalman [email protected]
12. April 2023
Load Balancing for Multimedia Streaming in Heterogeneous Peer-to-Peer Systems
Kalman Graffi, Sebastian Kaune, Konstantin Pussep, Aleksandra Kovacevic, Ralf Steinmetz
KOM – Multimedia Communications Lab 2
Outline
Introduction to P2P-based Multimedia Streaming
Model of Content Block based Streaming
Architecture for Load Balanced P2P Streaming Requirements System Architecture Task Allocation using a Scoring Function
Evaluation Simulation Setup Metrics Results
Conclusion
KOM – Multimedia Communications Lab 3
Introduction to P2P Multimedia Streaming
Large-scale multimedia streaming scenario A lot of users, sharing multimedia content between each other Peer-to-peer technology keeps the costs down How to balance the load on the peers and to support heterogeneity?
Streaming Architecture
KOM – Multimedia Communications Lab 4
Model of Content Block based Streaming
Streamable multimedia content C is split in m blocks with unique IDs
Model:
Two sets per block ID with peerIDs:
Owners and Demanders
Within an architecture that is distributed scalable
Typical system load: Maximum 100 providers per block
MM content CB1 B2 B3 B4 B5
B5
B4
B3
B2
B1
B5
B4
B3
B2
B1
Demand Set Wi
Owner Set Hi
P1
P2
P3
P4
P5
P6
P7
P1
P2
P3
P4
P5
P6
P7
KOM – Multimedia Communications Lab 5
Matching Problem
Content block based allocation of
block provider to block requester
Matching problem:What is the best strategy to matchpeers from Wi to Hi
considering: Peer load Heterogeneity of the peers
Leaving peers the freedom to define maximum load levels considers their current load
Leading to Balanced load regarding the
stream provision in the system
P3
P4
P5
P3
P4
P5
P1
P2
P6
P7
Block iRequester
Block iProvider
Block iRequester
Block iProvider
P1
P2
P6
P7
KOM – Multimedia Communications Lab 6
Outline
Introduction to P2P Streaming
Model of Content Block based Streaming
Architecture for Load Balanced P2P Streaming Requirements System Architecture Task Allocation using a Scoring Function
Evaluation Simulation Setup Metrics Results
Conclusion
KOM – Multimedia Communications Lab 7
Requirements on the Architecture Design
Functional requirements Users can announce their multimedia content Users find providers for desired content using the architecture Users can define a maximum load they are willing to carry The architecture solves the matching problem
Non-functional requirements No central element is used The matching problem is solved in a load balancing way
Supporting peer heterogeneity: considering the individual load Load balancing: aiming at an equal utilization of the block providers
Focus is on non-functional requirements Optimizing system wide aspects Leaving the path of traditional P2P patterns
KOM – Multimedia Communications Lab 8
Main Idea of the Solution
Optimizing system wide aspects Supporting peer heterogeneity Achieving load balancing
Main Idea Give the freedom of contribution back to peers
Peers define maximum load to carry Current capabilities are considered Strong peers add more, weak peers less
Give the control of load back to the system: System decides on load dispatching
Load balancing is a system wide state
In contrary to individual peers choosing other peers to contribute
KOM – Multimedia Communications Lab 9
Distributed Hybrid System Architecture
Network Structure: All entities are participants (peers) in a DHT DHT peers store per block ID (i)
a set of block providers (Hi)
Announcement: All peers (from Hi) register themselves for all of their
content blocks (Bi) at the peer responsible for the corresponding block ID (Ri)
Periodically, peers from Hi update their state at Ri
Peers chose themselves how much to announce
Motivation for design decisions: DHTs allow for role assignment Block ID mapping to peers is easy Hybrid approach:
centralize overview, distribute load
P3
P4
P5
Block 3requester
Block 3provider
P1
P2
P6
P7
DHT
10
Responsible for block 3
Peer Qual./Load
P1
P2
P6
okgoodweak
okP7
KOM – Multimedia Communications Lab 10
Distributed Hybrid System Architecture
Query Requesting peers (from Wi) ask Ri for
a content provider from Hi In the example: peer 5 asks for block 3
Matching on Ri
Ri uses a scoring function and the states of the peers in Hi to determine the most suitable provider
Ri answers with address of one provider, that is then used to stream the content
Load balancing aspect Block maintaining peer (Ri) focuses on a fair load allocation Various optimization strategies can be adopted
DHT
10
Responsible for block 3
Peer Qual./Load
P1
P2
P6
okgoodweak
okP7
Get provider for block 3Provider to
use: peer 2
KOM – Multimedia Communications Lab 11
Task Allocation using a Scoring Function
Question: How to derive “quality” of the providers?
Idea: Map each content provider to a linear scale Consider network status, load, heterogeneity and online duration Allocate tasks to peers based on their position on the scale Load is balanced by peers (Ri) responsible for the content
Peer (p) related information parameters, time dependent1. Allocated tasks for the system
2. Estimation of local tasks / load
3. Bandwidth quality of the peer
4. Online time of the peer
Scoring function determines most suitable provider
)(tI LTP)(tI BqP
)(tI ATP
)(tI OTP
)()()()(),( 4321 tItItItItpc OTP
BqP
LTP
ATPs RTimePeerstpcs :),(
KOM – Multimedia Communications Lab 12
Outline
Introduction to P2P Streaming
Model of Content Block based Streaming
Architecture for Load Balanced P2P Streaming Requirements System Architecture Task Allocation using a Scoring Function
Evaluation Simulation Setup Metrics Results
Conclusion
KOM – Multimedia Communications Lab 13
Simulation Setup and Metrics
Simulation environment PeerfactSim.KOM – the Peer-to-Peer Systems Simulator Focus on block based access:
25, 50, 75 and 100 peers interested per multimedia block 25 – 100 requests in total
Reference solution: random task assignment
Metrics Costs for one assignment: Cost for all assignments: Sum of all costs S(.) using strategy RND or SF Profit:
Measuring load balancing: standard deviation in the load distribution
)()()()(),( 4321 tItItItItpc OTP
BqP
LTP
ATPs
)(
)()(
SFS
SFSRNDS
KOM – Multimedia Communications Lab 14
Finding suitable Parameter Settings
For the setting of α1, α2, α3 and α4 :
Fix α3 (Bq) and α4 (OT) to 25%
Scenario: 100 peers, 100 requests
Observation: Best results for α1 = 45% and α2 = 5%
Load is balanced (low deviation)
Using this paramter setting Leads to more balanced load Higher „profit“ in comparison to
random load allocation
setup1 setup2 setup3 setup4 setup5
α1 / AT 5% 15% 25% 35% 45%
α2 / LT 45% 35% 25% 15% 5%
Profit 65% 62% 56% 62% 77%
KOM – Multimedia Communications Lab 15
Improved Load Balance
Metric for load balance
Simulation setup Peers, requests: 25/25,50/50,75/75,100/100 Varying allocation strategy: SF, RND
Observation SF leads to decreased deviation in peer load Load balancing is improved up to 53%
KOM – Multimedia Communications Lab 16
Tradeoff between Costs and Performance
Scoring function requires information on the peers (AT, LT, Bq, OT)
Generates traffic costs Fresh information generates more costs Does optimal update interval exist?
Simulation setup 100 peers, 100 requests Varying update interval Values vary continuously between 0 and 2
Observation High update intervals lead (as expected) to
Less traffic overhead Older information
But: No turning point identifiable Costs for performance nearly linear
KOM – Multimedia Communications Lab 17
Outline
Introduction to P2P Streaming
Model of Content Block based Streaming
Architecture for Load Balanced P2P Streaming Requirements System Architecture Task Allocation using a Scoring Function
Evaluation Simulation Setup Metrics Results
Conclusion
KOM – Multimedia Communications Lab 18
Conclusion
Architecture for multimedia content streaming Applicable on any DHT, load-balanced, supporting peer heterogeneity Designated peers are responsible for dispatching load to content providers They use a scoring function, which considers
Peer load (Active Tasks, Local Tasks) Peer heterogeneity (Bandwidth quality, Online Time)
Improved system performance: System controls load allocation, as load balancing is a system-wide metric In comparison to a random load dispatcher,
The scoring function results in up to 109% better choices The load deviation in the system is reduced by 53%
KOM – Multimedia Communications Lab 19
Further Questions ?
Peers
αβλ
μ
Parameters
f(α, β)=…=xg(λ, μ)=…=yh(α, λ)=…=z
ModelsInterpreted state
Architecture
Information Management Architecture
Analysis, Modeling and Interpretation
Using Info. to Gain
Efficiency