Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Chahrazed LABBA
Dissertation Defense
Adaptive Deployment ofMulti-Agent Systems on Cloud
Environments
Mme. Narjès Bellamine Ben Saoud
Advisor
DISSERTATION DEFENSE 20 DECEMBER 2017
In collaboration with: Mme Julie Dugdale
15-03-2018Chahrazed LABBA
Introduction
State of the art
Contributions
Implementation and experimental results
Conclusion and future work
2
Outline
1
2
3
4
5
15-03-2018Chahrazed LABBA 3
Multi-agent systems
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ContextMotivation
• Multi-Agent Systems (MAS):
A set of agents, that interact with each other, situated in acommon environment, eventually, building or participating to anorganization. [Dem2000]
• Agent:
Autonomy, Reactivity, Proactivity, Social abilities
• Model and simulate complex socio-technical systems.
15-03-2018Chahrazed LABBA 4
Healthcare
NGO
ArmyTelco
operators
Fire-brigade
Police
An agent-based decision
support system for
emergency management
Compute
Storage
Network
Security
Compute
Storage
Network
Security
Compute
Storage
Network
Security
Availability
Compute
Storage
Network
Security
Availability
Compute
Storage
Network
Availability
Compute
Storage
Network
An Example of a Multi-Agent System
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ContextMotivation
15-03-2018Chahrazed LABBA 5
Deployment of MAS on Cloud infrastructures
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ContextMotivation
Distributed environment
Cluster grid
Cloud
Hybrid-cloud Inter-cloud
- Not in the reach of all the organizations- Require a considerable expertise to be
installed, maintained and operated- May not fulfill all the QoS requirements
─ Scalability─ Flexiblity─ QoS requirements
Cloud federation Multi-Cloud
How to achieve a cost efficient deployment of MAS on cloud environment?
15-03-2018Chahrazed LABBA
• A partitioning Problem
6
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ContextMotivation
Multi-Agent System partitioning on clouds
• Assign agents to low
cost cloud resources
• Maintain reduced
communication costs
• Fulfill QoS needs
• Large number of
partitioning algorithms
Among all the partitioning options, which ones are more
suitable to partition a given MAS on a Given Cloud?
15-03-2018Chahrazed LABBA 7
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ContextMotivation
Complicated cloud infrastructure
Provider A
Inter-Cloud
Provider BProvider C
R1 R2
R3
R1R2
R1 R2
Availability
zones
Security
Availability
Security
Availability
Security
Availability
Dynamic agents‘ requirements Frequent load-balance issues
Multiple cloud providers Various cost models Different regions and zones Heterogeneous resources
For a given MAS(scenario), which cloud configuration settings
are suitable to achieve reduced deployment costs?
15-03-2018Chahrazed LABBA 8
Problem Statement: Depolyment of MAS on clouds
MASDeployment
on Clouds
MAS Properties Cloud issues
Predict and optimize
(Plan)
Large-scaleQoS
requirements
Partitioning
options
communication
Complicated
infrastructure
Competitive
third parties
costs models
Conflicting
Deployment needsMultiple deployment
solutions
How to provide a cost-efficient adaptive deployment of a given
multi-agent system on a given cloud environment?
Cloud settings
15-03-2018Chahrazed LABBA 9
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Partitioning ApproachesPrediction-based approches
Research Context
Partitioning approaches
• Multi-Agent Systems
• Cloud Applications
Prediction & estimation- based approach
• Multi-Agent Systems
• Cloud applications
15-03-2018Chahrazed LABBA
• MAS classification
MAS_S MAS_C MAS_H
10
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Partitioning ApproachesPrediction-based approches
Partitioning approaches for MAS (1/3)
3 MAS types(Ngobye et all, 2010)
• NI-MAS
• SI-MAS
• CI-MAS
2 MAS types(Salem et all, 2015)
• Closed MAS
• Open MAS
2 MAS types(Glavic, 2006)
• Independent MAS
• Cooperative MAS
3 MAS types(Ferber, 1999)
• Purely communicative
• Purely situated
• Hybrid
MAS with only
Indirect interactions
MAS with only
direct interactions
MAS with Both (in)
direct interactions
- Environment + Non direct
interactions
- Different types of
interactions
- Cooperate through
interactions
- Cooperation is not
an obligation
- Interaction without
intension to do so
- Homogenous/
Heterogeneous non
Interaction
- Homogenous/
Heterogeneous interaction
- Direct interaction
- Indirect interaction
- Direct and indirect
interaction
15-03-2018Chahrazed LABBA 11
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Partitioning ApproachesPrediction-based approches
Partitioning approaches for MAS (2/3)
Cluster-Based
Approaches
Grid-Based
Approaches
Graph-Based
Approaches
An operational framework for MAS partitioning ×Focus on one MAS type.× Restricted number of partitioning algorithms.×Consider one single target infrastructure.× Results can not be generalized to other distributed environments and MAS types.
No overall study on the appropriateness of partitioning algorithms according to the variety of MAS types
Vigueras et al,2009
Wang et al, 2009/ 2012
15-03-2018Chahrazed LABBA 12
× do not consider the cloud characteristics:compute costs, communication costs, resource
heterogeneity, variety of cloud providers ×Vendor lock-in problem (single cloud provider)
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Partitioning ApproachesPrediction-based approches
Partitioning approaches for MAS (3/3)
15-03-2018Chahrazed LABBA 13
Partitioning approaches for Cloud Applications
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Partitioning ApproachesPrediction-based approches
Optimize the deployment costsSupport various QoS requirements× Dedicated for a given application type×Insufficient for some MAS types×Low granularities that induce more complexity to the partitioning
15-03-2018Chahrazed LABBA
• Predict the performance and the correctness of the MAS
• Detect load-balance issues
• Evaluate partitioning mechanisms for MAS on clusters
14
Prediction approaches for MAS
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Partitioning ApproachesPrediction-based Approches
Prediction at the MAS level
Prediction at the infrastructure level
×No support for cloud environments
15-03-2018Chahrazed LABBA 15
Scheduling
Provisioning
Service Selection
Estimate and optimize the cloud costs ×Cover few QoS requirements: CPU and RAM×Support a given cloud environment×Support a given application type
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Partitioning ApproachesPrediction-based approches
Prediction approaches for cloud applications
15-03-2018Chahrazed LABBA
1
• Recommend partitioning algorithms to develop new distributed MAS
• Deal with the different MAS types as well as cloud environments
• Analyze the existing partitioning algorithms
• Give guidlines to develop new distributed systems
2
• optimize MAS partitioning on cloud environments
• Consider cloud specifications
• Consider MAS characteristics
3
• Prediction approach
• Deal with different partitioning options
• Deal with different cloud settings
• A cost model to compare deployment solutions
16
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Objectives
15-03-2018Chahrazed LABBA 17
Framework for MAS deployment on clouds
• Pre-deployment method
• Deployment process
Framework to support MAS deployment on clouds
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Contributions
Pre
-de
plo
ym
en
t
me
tho
dd
ep
loym
en
t
pro
ce
ss
A pre-selection Process
A prediction Process
Deployment Service
An operational framework for MAS partitioning
• Analyze MAS type
• Recommend partitioning algorithms
New extended partitioning algorithms
• New graph-based partitioning algorithms
A prediction process
• Profiling phase
• Cost deployment model
15-03-2018Chahrazed LABBA 18
Framework to Support MAS Deployment on Clouds
Pre
-sel
ecti
on
pro
cess
Pre
dic
tio
n
Pro
cess
Pre
-dep
loy
men
t M
eth
od
Dep
loy
men
t
Pro
cess
User
specifications
Scenario Definition
Partitioning algorithms
Library
Algorithm1
Algorithm2
.
Algorithm n
Sequential simulation
Profiling phase (2)
Pre-Selection
Service (1)
Deployment Service (6)
Candidate Partitioning
Algorithm
PA_1,PA_2, ..,PA_k
Cloud Specifications
Algorithms
invocation
Service (4)Decision Making
Service (5)
Monitoring
Service (3)
PA_1 PA_2PA_k
Deployment specifications Architecture
Target Cloud
Pre-deployment
method
Deployment
Pre-selection
Process
Prediction
Process
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment
Overview of the Framework
15-03-2018Chahrazed LABBA 19
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment
An operational Framework for MAS
Partitioning
Enter MAS
Specifications
Specify Deployment
Infrastructure
Analyze MAS
Specifications
Determine MAS
Type
Recommend a set
Of candidate partitioning
algorithms
Pre-selection process
- Criteria
- Reference combinations
- Matching grid
15-03-2018Chahrazed LABBA 20
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment
Analyze MAS types
Situated MAS
Communicative
MAS
Hybrid MAS
15-03-2018Chahrazed LABBA 21
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment
A matching Grid for recommending partitioning algorithms
15-03-2018Chahrazed LABBA 22
Extended Partitioning Algorithms
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment
15-03-2018Chahrazed LABBA 23
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment
Graph partitioning on cloud Infrastructure
Objective function to minimize the inter-cloud
Communication costs
Agents
Interactions
Agents requirements
Amount of communicated data
CPU requirements
RAM requirements
Security
Availability
Bandwidth
15-03-2018Chahrazed LABBA 24
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment
Graph partitioning algorithm for cloud Infrastructures
Assign each agent to a cloud provider and a resource
Apply Graph partitioning algorithm
Resource allocation
Initial deployment
SolutionAdaptive deployment
15-03-2018Chahrazed LABBA 25
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment
Generation of a deployment Graph
Cloud1Cloud2
VM2
VM1
VM1
VM2
VM3
0,5 Go1 Go
Generated deployment “nested” graph for inter-clouds
On-premisePublic Clouds
VM1
VM2
VM2
VM1Private cloudCloud1
Cloud20,2 Go
0,5 Go0,9 Go
0,5 Go
0,9 Go
0,5 Go0,5 Go
0,3 Go
Generated deployment “nested” graph for Hybrid-clouds
15-03-2018Chahrazed LABBA 26
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment
Extended FM algorithms (1/3)
Takes into consideration the agents QoS requirement while searching for minimal costs.Minimizes the communication costs instead of their numbers. No random initial partitioning
Compute the gain
Move the agent with
Maximum gain (feasible move)
Update the gain
15-03-2018Chahrazed LABBA
• Cloud Environment E-FM (C2E-FM)
• Inter-Cloud E-FM (ICE-FM)
• Hybrid-Cloud E-FM (HCE-FM)
27
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment
Extended FM algorithms (2/3)
15-03-2018Chahrazed LABBA 28
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment
PMFM and Extended PMFM algorithms
• Cloud Environment E-PMFM (C2E-PMFM) : Calls C2E-FM
• Inter-Cloud E-PMFM (IC2E-PMFM): Calls ICE-FM
Divide the graph into k bloks
Creates pairs of blocks
Call FM algorithm
15-03-2018Chahrazed LABBA
• The new algorithms
29
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment
Proposed Partitioning algorithms (1/2)
15-03-2018Chahrazed LABBA 30
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment
Proposed Partitioning algorithms (2/2)
15-03-2018Chahrazed LABBA 31
Prediction Process: Estimate and optimize
Deployment costs
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment costs
15-03-2018Chahrazed LABBA 32
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment costs
Description of the full prediction Process
15-03-2018Chahrazed LABBA 33
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment
Cost model for comparing deployment solutions
Communication
cost +
Migration
cost +
Execution time
cost∑ cost of all the inter-messages
15-03-2018Chahrazed LABBA
• Migration costs
• Execution Time costs
34
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
Framework for MAS deployment on cloudsAn operational Framework for MAS partitioningExtended Partitioning AlgorithmsPrediction Process: estimate and optimize deployment
Cost model for comparing deployment solutions
∑ cost of all the migrations
15-03-2018Chahrazed LABBA 35
Experiments and implementation
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ImplementationCase study1: Emergency managementCase study 2: Dynamic business processes
15-03-2018Chahrazed LABBA
• Agent Technology (JADE)
Pre-selection process
Prediction process
Emergency management simulator
Dynamic business process
• Programming language:
Partitioning algorithms
36
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ImplementationCase study 1: Emergency managementCase study 2: Dynamic business processes
Implementation
15-03-2018Chahrazed LABBA 37
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ImplementationCase study 1: Emergency managementCase study 2: Dynamic business processes
Experimental settings
Experiment 1: compare HCPAdata, HCPAcost
And HCPAedge
• Inter-communication costs
• Overall deployment costs
Experiment 2: Impact of cloud configurations
on the overall deployment costs
• Hybrid cloud settings
• Cloud providers
• Availability zones
• Bandwidth
15-03-2018Chahrazed LABBA 38
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ImplementationCase study 1: Emergency managementCase study 2: Dynamic business processes
Experiment 1
- HCPAcost algorithm outperforms HCPAedge
and HCPAdata in terms of both communication
and overall deployment costs.
- HCP Aedge results in increased inter
communication and deployment costs.
15-03-2018Chahrazed LABBA 39
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ImplementationCase study 1: Emergency managementCase study 2: Dynamic business processes
Experiment 2
- HCPAcost algorithm
- Availability zone R1
- 100 Mbit/s of bandwidth (on/off premises)
- 300 Mbit/s of bandwidth (inter-cloud)
- Setting 1, setting 2, setting 3
Setting 1 provides better results in
terms of overall deployment costs
- HCPAcost algorithm
- Availability zone R1
- 100 Mbit/s of bandwidth (on/off premises)
- Hybrid setting1
- P1(vm1), P2(vm2), P3(vm3)
The use of P2 provides better results
in terms of overall deployment costs
- HCPAcost algorithm
- 100 Mbit/s of bandwidth (on/off premises)
- Hybrid setting1
- P2
- Availability zone R1, R2, R3
The use of R1 provides better results
in terms of overall deployment costs
- HCPAcost algorithm
- Hybrid setting1
- P2
- Availability zone R1
- Bandwidth 300Mbit/s, 100Mbit/s, 500Mbit/s
The use of B1 provides better results
in terms of overall deployment costs
15-03-2018Chahrazed LABBA 40
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ImplementationCase study 1: Emergency managementCase study 2: Dynamic business processes
Input for first fit algorithm
15-03-2018Chahrazed LABBA
• ICPAcost Vs ICPAdata
41
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ImplementationCase study 1: Emergency managementCase study 2: Dynamic business processes
Experiment 1: Towards an enhanced initial partitioning (1/2)
• Regardless the type of the infrastructure as well as the initial partitions
ICPAcost outperforms ICPAdata
15-03-2018Chahrazed LABBA 42
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ImplementationCase study 1: Emergency managementCase study 2: Dynamic business processes
Experiment 2: Towards an enhanced initial partitioning (2/2)
• The initial partitioning is
impacted by different criteria:
Type of infrastructure
Number of allowed
instances
The order of the agents,
cloud and VM lists
15-03-2018Chahrazed LABBA 43
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ImplementationCase study 1: Emergency managementCase study 2: Dynamic business processes
Experiment 2: ICPAcost Vs a Naive approach
• The use of our algorithm provides better results in terms of deployment costs
regardless the used initial partitions.
15-03-2018Chahrazed LABBA 44
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ConclusionFuture work
Conclusion and Future Work
15-03-2018Chahrazed LABBA
• Framework to support MAS deployment on cloudenvironments
o Pre-Deployment method
o Deployment Process
• An operational framework for MAS partitioning
o Analyze the MAS type through defined criteria
o Determine the MAS type
o Recommend a set of candidate partitioning algorithms
• New graph-based partitioning algorithms
• A prediction process
o Profiling phase
o Deployment costs models
45
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ConclusionFuture work
15-03-2018Chahrazed LABBA
• Real deployment
• New partitioning algorithms
• Cloud Resource description
• Impact of agent development framework
• Support more distributed systems
46
Introduction
State of the art
Contributions
Experiments and implementation
Conclusion et Future work
ConclusionFuture work
15-03-2018Chahrazed LABBA
International journal papers:1. Chahrazed Labba, Narjes Bellamine Ben Saoud, Julie Dugdale, A predictive ap- proachfor the efficient distribution of agent-based systems on a hybrid-cloud, In FutureGeneration Computer Systems, 2017, ISSN 0167-739X, (impact factor: 3.997)
International conference papers1. Chahrazed Labba, Nour, Assy, Narjes Bellamine Ben Saoud, Walid Gaaloul,: “Adap- tiveDeployment of Service-based processes into cloud federations”, The 18th Inter-national Conference on Web Information Systems Engineering, WISE 2017, accepted paper.(Rank A)
2. Chahrazed Labba, Noura ben Saleh, Narjes Bellamine Ben Saoud: “An Agent-based Meta-Model for Response Organization Structures”, The 4 th International Confer- ence onInformation Systems for Crisis Response and Management in Mediterranean CountriesISCRAM-med 2017.
3. Chahrazed Labba, Narjes Bellamine Ben Saoud, Julie Dugdale: “Towards a con- ceptualframework to support adaptative agent-based systems partitioning”. SNPD 2015: 692-696(Rank C).
4. Chahrazed Labba and Narjes Bellamine Ben Saoud: “Cost-Based Assesment Of PartitioningAlgorithms Of Agent-Based Systems On Hybrid Cloud Environments”, EMSS 2015, Italy(Rank B).
5. Chahrazed Labba, Narjes Bellamine Ben Saoud, Karim Chine: “Towards Large-Scale CloudBased Emergency Management Simulation : SimGenis Revisited". ISCRAM- med 2014: 13-20.
47
Publications List
15-03-2018Chahrazed LABBA 48
Thank you for your attention !