Resource Mapping Optimization for Distributed Cloud Services - PhD Thesis Defense
-
Upload
atakan-aral -
Category
Science
-
view
107 -
download
1
Transcript of Resource Mapping Optimization for Distributed Cloud Services - PhD Thesis Defense
Resource Mapping Optimization for Distributed Cloud Services
Resource Mapping Optimizationfor Distributed Cloud Services
PhD Thesis Defense
Atakan Aral
Thesis Advisor: Asst. Prof. Dr. Tolga Ovatman
Istanbul Technical University – Department of Computer Engineering
November 3, 2016
Resource Mapping Optimization for Distributed Cloud Services
Introduction
Motivation
Outline
1 IntroductionMotivationPreliminary InformationGeneral ProblemSolution Proposal
2 Topology MappingMotivating ExampleProblem Definition
Solution DetailsEvaluation
3 Replication ManagementMotivating ExamplesProblem DefinitionSolution DetailsEvaluation
4 Conclusion
Resource Mapping Optimization for Distributed Cloud Services
Introduction
Motivation
Motivation
Resource Mapping Optimization for Distributed Cloud Services
Introduction
Motivation
Motivation
Resource Mapping Optimization for Distributed Cloud Services
Introduction
Motivation
Motivation
Resource Mapping Optimization for Distributed Cloud Services
Introduction
Motivation
Motivation
Scientific computingInternet of ThingsFinance servicesHealth care systemse-Learning
(Mobile) image processing, VRSocial networkingOn-demand or live video streamingOnline gaming. . .
Real time, distributed, data- and computation-intensive cloud servicesneed low latency and low-cost communication.
Resource Mapping Optimization for Distributed Cloud Services
Introduction
Preliminary Information
Outline
1 IntroductionMotivationPreliminary InformationGeneral ProblemSolution Proposal
2 Topology MappingMotivating ExampleProblem Definition
Solution DetailsEvaluation
3 Replication ManagementMotivating ExamplesProblem DefinitionSolution DetailsEvaluation
4 Conclusion
Resource Mapping Optimization for Distributed Cloud Services
Introduction
Preliminary Information
Cloud Computing
DefinitionApplications and services that run on a distributed network using virtualizedresources and accessed by common Internet protocols and networking standards.
Broad network access: Platform-independent, via standard methodsMeasured service: Pay-per-use, e.g. amount of storage/processing power,number of transactions, bandwidth etc.On-demand self-service: No need to contact provider to provision resourcesRapid elasticity: Automatic scale up/out, illusion of infinite resourcesResource pooling: Abstraction, virtualization, multi-tenancy
Resource Mapping Optimization for Distributed Cloud Services
Introduction
Preliminary Information
Cloud Computing – Service Models
Resource Mapping Optimization for Distributed Cloud Services
Introduction
Preliminary Information
Cloud Computing – Deployment Models
Resource Mapping Optimization for Distributed Cloud Services
Introduction
Preliminary Information
Inter-Cloud
DefinitionA cloud model that allows on-demand reassignment of resources and transfer ofworkload through an interworking of cloud systems of different cloud providers.
Depending on the initiator of the Inter-Cloud endeavour:Cloud Federation: A voluntary collaboration between a group of cloudproviders to exchange their resourcesMulti-Cloud: An uninformed combination of multiple clouds by a serviceprovider to host the components of the service
Resource Mapping Optimization for Distributed Cloud Services
Introduction
Preliminary Information
Edge Computing
DefinitionPushing the frontier of computing applications, data, and services away fromcentralized nodes to the logical extremes of a network (e.g. mobile devices,sensors, micro data centers, cloudlets, routers, modems, . . .
Low latency access to powerful computing resourcesCyber-foraging: Computation or data offloading from mobile devices toservers for extending computation power, storage capacity and/or battery life.
Resource Mapping Optimization for Distributed Cloud Services
Introduction
General Problem
Outline
1 IntroductionMotivationPreliminary InformationGeneral ProblemSolution Proposal
2 Topology MappingMotivating ExampleProblem Definition
Solution DetailsEvaluation
3 Replication ManagementMotivating ExamplesProblem DefinitionSolution DetailsEvaluation
4 Conclusion
Resource Mapping Optimization for Distributed Cloud Services
Introduction
General Problem
General Problem
Deployment on multiple clouds can benefit cloud servicesAllows seemingly infinite scalability,Provides better geographical coverage,Avoids vendor lock-in and eases hybridization,Increases fault tolerance and availability,Allows exploiting differences in pricing schemes, . . .
Business solutions that allow basic inter-cloud deployment are already here,e.g. Nuvla, EGI, RightScale, Equinix, . . .However, such solutions leave cloud data center selection to user.There is no automatic mapping between provider resources and userrequirements.
Resource Mapping Optimization for Distributed Cloud Services
Introduction
General Problem
General Problem (continued)
Resource mapping in distributed cloud is a complex problem due to factorssuch as:
Multiple objectives and perspectives (QoS, access latency, throughput,availability, cost, profit, . . . ),Constraints (SLAs, processing capacity, network bandwidth, energyconsumption, . . . ),Geographical distribution and nonuniform network conditions,Heterogeneity and dynamicity of both entities (i.e. resources and requests),The large number of the entities (especially in the case of Edge Computing),Inter-dependencies among the components that constitute a cloud service (e.g.Virtual Machines, Databases).
Resource Mapping Optimization for Distributed Cloud Services
Introduction
Solution Proposal
Outline
1 IntroductionMotivationPreliminary InformationGeneral ProblemSolution Proposal
2 Topology MappingMotivating ExampleProblem Definition
Solution DetailsEvaluation
3 Replication ManagementMotivating ExamplesProblem DefinitionSolution DetailsEvaluation
4 Conclusion
Resource Mapping Optimization for Distributed Cloud Services
Introduction
Solution Proposal
Proposed Solution
HypothesisEmploying resource mapping algorithms that consider and utilize structuralcharacteristics of the distributed cloud services would increase the QoSexperienced by service users in a cost-efficient way.
QoS indicators are: (i) access latency to the service, (ii) access latencybetween dependent components, and (iii) access latency to the data.Cost indicators are: (i) VM provisioning cost, (ii) data storage cost, and (iii)data transfer (bandwidth) cost.
VMnVM1 VM2...
Service ModelEntity
Cloud based Service SaaS
Resource Mapper PaaS
Topology Mapper Replication Manager
• Replica Placement
• Replica Discovery
• Network Embedding
• Resource Allocation
DB
Cloud Infrastructure IaaS
DC1
DC2
DC4 DC5
DC3
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Motivating Example
Outline
1 IntroductionMotivationPreliminary InformationGeneral ProblemSolution Proposal
2 Topology MappingMotivating ExampleProblem Definition
Solution DetailsEvaluation
3 Replication ManagementMotivating ExamplesProblem DefinitionSolution DetailsEvaluation
4 Conclusion
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Motivating Example
Motivating Example
A small cloud-based navigation service for public transportationTwo-tier architecture: User interface and route planning
VM 1: 8 CPU cores, 16 GB memory, 2 TB HDD storageVM 2: 32 CPU cores, 60 GB memory, 80 GB SSD storage
500 Mbps of dedicated bandwidth between the two tiers is desired
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Motivating Example
Motivating Example (continued)
First tier is replicated in twoseparate DCs to improve:
AvailabilityFault toleranceProximity
Second tier is not replicateddue to:
Economical constraintsUncriticality to the service
But it is still deployed on athird DC because of:
Pricing differencesFairness
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Motivating Example
Motivating Example (continued)
There are 5 (federated) cloud providers in the area.
Not all have dedicatednetwork connectionsbetween them.Heterogeneous bandwidthcapacities and latenciesHeterogeneous resourcecapacities and loads
How to map user virtual machines to cloud data centers?
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Problem Definition
Outline
1 IntroductionMotivationPreliminary InformationGeneral ProblemSolution Proposal
2 Topology MappingMotivating ExampleProblem Definition
Solution DetailsEvaluation
3 Replication ManagementMotivating ExamplesProblem DefinitionSolution DetailsEvaluation
4 Conclusion
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Problem Definition
Virtual Machine Cluster Embedding
Mapping VM clusters across inter-cloud infrastructure (node & edge mapping)
Reduce inter-cloud latencyReduce makespanReduce resource costs
Optimize bandwidthutilizationIncrease system throughputIncrease acceptance rateIncrease revenue
CP4
CP1
CP5
CP2
CP3
VM1 VM3VM2
CP4
CP1
CP5
CP2VM1
VM2VM3
CP3
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Problem Definition
Virtual Machine Cluster Embedding (continued)
GC =(
VC ,EC ,AVC ,A
EC
)GF =
(VF ,EF ,AV
F ,AEF
)∀ (v ∈ VC) ∃
(v ′ ∈ VF
)| v 7→ v ′
∀ (e ∈ EC) ∃(E ′ ⊆ EF
)| e 7→ E ′
Validity conditions:1 Capacity sufficiency2 Path creation3 No cycle forming
CP4
CP1
CP5
CP2
CP3
VM1 VM3VM2
CP4
CP1
CP5
CP2VM1
VM2VM3
CP3
f = (VM1 7→ CP3,VM2 7→ CP5,VM3 7→ CP4)
g =(D1,2 7→ {C1,3,C3,5},D2,3 7→ {C4,5}
)
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Solution Details
Outline
1 IntroductionMotivationPreliminary InformationGeneral ProblemSolution Proposal
2 Topology MappingMotivating ExampleProblem Definition
Solution DetailsEvaluation
3 Replication ManagementMotivating ExamplesProblem DefinitionSolution DetailsEvaluation
4 Conclusion
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Solution Details
Topology based Mapping (TBM) Algorithm
Map VM clusters to the subgraphs of the federation topology that areisomorphic to their topology.
Mapping function f is injective and ∀ (e ∈ EC) ∃ (E ′ ⊆ EF ) | e 7→ E ′∧ |E ′| = 1
In case of multiple alternatives, choose by average latency to the user.Fall back to a greedy heuristic in case of failure.
Instead, map to a homeomorphic subgraph where |E ′| ≥ 1
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Solution Details
Topology based Mapping (TBM) Algorithm (continued)
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Solution Details
Topology based Mapping (TBM) Algorithm (continued)
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Solution Details
Topology based Mapping (TBM) Algorithm (continued)
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Solution Details
Topology based Mapping (TBM) Algorithm (continued)
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Evaluation
Outline
1 IntroductionMotivationPreliminary InformationGeneral ProblemSolution Proposal
2 Topology MappingMotivating ExampleProblem Definition
Solution DetailsEvaluation
3 Replication ManagementMotivating ExamplesProblem DefinitionSolution DetailsEvaluation
4 Conclusion
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Evaluation
Experimental Setup
Simulated on the RalloCloud framework which is based on CloudSim.Federation topology is taken from the FEDERICA project and contains 14clouds across Europe.Virtual machine clusters are randomly generated based on population density.Simulation period is 50 hours.
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Evaluation
Results (1/4)
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Evaluation
Results (2/4)
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Evaluation
Results (3/4)
Resource Mapping Optimization for Distributed Cloud Services
Topology Mapping
Evaluation
Results (4/4)
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Motivating Examples
Outline
1 IntroductionMotivationPreliminary InformationGeneral ProblemSolution Proposal
2 Topology MappingMotivating ExampleProblem Definition
Solution DetailsEvaluation
3 Replication ManagementMotivating ExamplesProblem DefinitionSolution DetailsEvaluation
4 Conclusion
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Motivating Examples
Sport Event
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Motivating Examples
Sport Event
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Motivating Examples
Sport Event
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Motivating Examples
Traffic
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Motivating Examples
Traffic
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Motivating Examples
Motivating Examples
Edge Computing provides low-latency access to computing resources formobile code offloading.However, many services need to access data that is stored centrally due to:
Limited storage capacity of the edge entitiesEconomic constraintsAvailability for offline analysisSimpler maintenance and concurrency control
High latency to central storage harms QoS.How to decide the number and location of data replicas so that the dataaccess latency is decreased in a cost efficient way?
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Motivating Examples
Locality of Reference
The patterns that the user data accesses exhibit: Temporal Locality, SpatialLocality, and Geographical Locality
[0,0
.5)
[0.5,1
)
[1,1
.5)
[1.5,2
)
[2,2
.5)
[2.5,3
)
[3,3
.5)
[3.5,4
)
[4,4
.5)
[4.5,5
)
[5,5
.5)
[5.5,6
)
[6,6
.5)
[6.5,7
)
[7,7
.5)
[7.5,8
)
[8,8
.5)
[8.5,9
)
[9,9
.5)
[9.5,1
0)
[10,1
0.5)
[10.5,1
1)
[11,1
1.5)
[11.5,1
2)
[12,1
2.5)
[12.5,1
3)
[13,1
3.5)
[13.5,1
4)
[14,1
4.5)
[14.5,1
5)
[15,1
5.5)
[15.5,1
6)
[16,1
6.5)
[16.5,1
7)
[17,1
7.5)
[17.5,1
8)
[18,1
8.5)
[18.5,1
9)
[19,1
9.5)
[19.5,2
0)
0
2
4
6
·104
Distance (1, 000× km)
Num
ber
ofR
eque
stPa
irs
The CAIDA Anonymized Internet Traces 2015 Dataset (2.3 Billion IPv4 packets)
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Problem Definition
Outline
1 IntroductionMotivationPreliminary InformationGeneral ProblemSolution Proposal
2 Topology MappingMotivating ExampleProblem Definition
Solution DetailsEvaluation
3 Replication ManagementMotivating ExamplesProblem DefinitionSolution DetailsEvaluation
4 Conclusion
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Problem Definition
Edge Replica Placement
a
c
b d cs
1 Which data objects to replicate?2 When to create/destroy a replica?3 How many replicas for each object?
4 Where to store each replica?5 How to redirect requests to the
closest replica?
In order to minimize average replica-to-client distance in a bandwidth-and cost-effective way.
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Problem Definition
Facility Location Problem
minimize :∑
j
fj · Yj +∑
i
∑j
hi · dij · Xij
fj = unit_pricej · replica_size · epochhi = num_requestsi · replica_sizedij = latencyij · λ
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Problem Definition
Requirements
Centralized solutions are not feasible due to the large number of entities.The solution must be distributed with minimal input and communication.
Edge users continuously enter and leave the network topology throughconnections with nonuniform latencies.
The solution must be dynamic and online.Edge entities should be aware of the closest replica when they need a certaindata object.
A Replica Discovery technique is necessary.
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Solution Details
Outline
1 IntroductionMotivationPreliminary InformationGeneral ProblemSolution Proposal
2 Topology MappingMotivating ExampleProblem Definition
Solution DetailsEvaluation
3 Replication ManagementMotivating ExamplesProblem DefinitionSolution DetailsEvaluation
4 Conclusion
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Solution Details
Decentralized Replica Placement (D-ReP) Algorithm
Storage nodes that host replicas act as local optimizers.They evaluate experienced demand, storage cost as well as expected latencyimprovement to carry out either:
Duplication num_requestsknh · latencynh · λ > unit_pricen · epochMigration (num_reqsknh −
∑i∈Ni 6=n
num_reqskih) · latencynh · λ >(unit_pricen − unit_priceh
)· epoch
Removal∑
i∈N num_requestskih < original_num_requestsh · αThe replicas are incrementally pushed from the central storage to the edge.λ allows user to control the trade-off between cost- and latency-optimization.
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Solution Details
Replica Discovery
Only the most relevant nodes are notified of the replica creations or removals.Temporal locality: requesters of the source during the n most recent epochsGeographical locality: m-hop sphere of the destination
Each active node keeps a Known Replica Locations (KRL) table.
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Evaluation
Outline
1 IntroductionMotivationPreliminary InformationGeneral ProblemSolution Proposal
2 Topology MappingMotivating ExampleProblem Definition
Solution DetailsEvaluation
3 Replication ManagementMotivating ExamplesProblem DefinitionSolution DetailsEvaluation
4 Conclusion
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Evaluation
Experimental Setup
Simulated on the RalloCloud framework which is based on CloudSim.The CAIDA Anonymized Internet Traces 2015 Dataset: IPv4 packets datafrom February 19, 2015 between 13:00-14:00 (UTC) which contains morethan 2.3 Billion recordsGeoLite2 IP geolocation databaseSynthetic workloads based on uniform, exponential, normal, Chi-squared, andPareto distributions of request locationsBarabási–Albert scale-free network generation model: 1000 nodes, 2994edges, and a heavy-tailed distribution of bandwidth in [10, 1024] mbpsAmazon Web Services S3 prices
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Evaluation
Results (1/4)
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Evaluation
Results (2/4)
Resource Mapping Optimization for Distributed Cloud Services
Replication Management
Evaluation
Results (3/4)
Resource Mapping Optimization for Distributed Cloud Services
Conclusion
Outline
1 IntroductionMotivationPreliminary InformationGeneral ProblemSolution Proposal
2 Topology MappingMotivating ExampleProblem Definition
Solution DetailsEvaluation
3 Replication ManagementMotivating ExamplesProblem DefinitionSolution DetailsEvaluation
4 Conclusion
Resource Mapping Optimization for Distributed Cloud Services
Conclusion
Contribution
Topology Mapping Algorithm1 4 5
First attempt to employ subgraph isomorphism to find a injective match betweenvirtual and physical cloud/grid topologiesFirst to explicitly evaluate network latency for the VMNE
Minimum Span Heuristic3 5
Novel heuristic algorithm to defer MIPDecentralized Replica Placement Algorithm2 5
First completely decentralized, partial-knowledge replica placement algorithmKRL based Discovery Technique2
Novel replica discovery technique with low overheadRalloCloud: A simulation environment for Inter-Cloud resource mapping1
First simulator for inter-cloud systems
Resource Mapping Optimization for Distributed Cloud Services
Conclusion
Publications
1 Aral, A., and Ovatman, T. 2016. Network-Aware Embedding of Virtual Machine Clusters onto FederatedCloud Infrastructure. The Journal of Systems and Software, vol. 120, pp. 89-104. DOI:10.1016/j.jss.2016.07.007
2 Aral, A., and Ovatman, T. 2017?. A Decentralized Replica Placement Algorithm for Edge Computing.Under review in The IEEE Transactions on Parallel and Distributed Systems.
3 Aral, A., and Ovatman, T. 2014. Improving Resource Utilization in Cloud Environments using ApplicationPlacement Heuristics. In 4th Int’l. Conf. on Cloud Comp. and Services Science (CLOSER 2014), pp.527-534.
4 Aral, A., and Ovatman, T. 2015. Subgraph Matching for Resource Allocation in the Federated CloudEnvironment. In IEEE 8th International Conference on Cloud Computing (IEEE CLOUD 2015), pp.1033-1036.
5 Aral, A. 2016. Network-Aware Resource Allocation in Distributed Clouds. IEEE International Conferenceon Cloud Engineering (IC2E 2016), Doctoral Symposium.
Resource Mapping Optimization for Distributed Cloud Services
Conclusion
Future Work
Standardization of cloud APIsReal-time performance and cost guaranteesService self-awarenessLoad balancing and fairness between cloud providersFault tolerance and availability
Resource Mapping Optimization for Distributed Cloud Services
Conclusion
Thank you for your time.
Resource Mapping Optimization for Distributed Cloud Services
Literature Review
Appendices
5 Literature Review
6 RalloCloud
7 Algorithm Details
8 Intra-Cloud Mapping
9 Subgraph Matching
10 Complete Results
Resource Mapping Optimization for Distributed Cloud Services
Literature Review
Virtual Network Embedding
Cloud Federation Embedding Entity Simultaneous NW-AwareTBM 3 3 3 VMs 3 3
[30] 3 3 3 VMs[31] 3 3 3 VMs 3 3
[32] 3 3 3 Serv. 3
[33] 3 3 3 Tasks[34] 3 3 3 VMs 3 3
[35] 3 3 3 VNs 3
[22] 3 3 3 VMs 3
[36] 3 3 3 VMs 3
[37] 3 3 3 VNs 3
[25] 3 3 3 VMs 3 3
[38] 3 3 3 VNs 3 3
Resource Mapping Optimization for Distributed Cloud Services
Literature Review
Virtual Network Embedding
Cloud Federation Embedding Entity Simultaneous NW-AwareTBM 3 3 3 VMs 3 3
[39] 3 VNs[40] 3 VNs 3 3
[41] 3 VMs 3 3
[42] 3 VNs 3
[43] 3 3 VNs 3 3
[44] 3 3 VMs 3 3
[45] 3 3 VNs 3 3
[46] 3 3 VMs 3 3
[47] 3 3 VNs 3
[48] 3 3 Tasks 3 3
[49] 3 3 VNs 3
Resource Mapping Optimization for Distributed Cloud Services
Literature Review
Virtual Network Embedding
Cloud Federation Embedding Entity Simultaneous NW-AwareTBM 3 3 3 VMs 3 3
[50] 3 3 VNs 3
[51] 3 3 VMs 3 3
[52] 3 3 Jobs[53] 3 3 WEs 3 3
Resource Mapping Optimization for Distributed Cloud Services
Literature Review
Virtual Network Embedding
[25] Tordsson, J., Montero, R.S., Moreno-Vozmediano, R. and Llorente, I.M. (2012). Cloud brokering
mechanisms for optimized placement of virtual machines across multiple providers, Future Generation
Computer Systems, 28(2), 358–367.
Exact solution with IP, cloud brokerage and uniform interface[35] Leivadeas, A., Papagianni, C. and Papavassiliou, S. (2013). Efficient resource mapping framework over
networked clouds via iterated local search-based request partitioning, IEEE Transactions on Parallel and
Distributed Systems, 24(6), 1077–1086.
Graph partitioning, IP based embedding, no latency consideration[38] Xin, Y., Baldine, I., Mandal, A., Heermann, C., Chase, J. and Yumerefendi, A. (2011). Embedding virtual
topologies in networked clouds, Proceedings of the 6th International Conference on Future Internet
Technologies, ACM, pp.26–29.
Connected components of the VM cluster topology are merged,isomorphic subgraph to the resulting graph is sought
Resource Mapping Optimization for Distributed Cloud Services
Literature Review
Replica Placement (Centralized)
P D Environment Topology Objectives[68] Web Tree Proximity[69] CDN Unrestricted Proximity[70] N/A Unrestricted Proximity[71] Data Grid Tree Proximity, Cost, Load Balance[72] N/A Tree Proximity[73] N/A Unrestricted Proximity[65] Cloud Unrestricted Proximity[74] Cloud Unrestricted Bandwidth, Load Balance[67] Cloud Unrestricted Proximity, Cost[75] 3 N/A Unrestricted Availability[76] 3 Data Grid Multi-Tier Proximity, Cost, Bandwidth[77] 3 CDN Unrestricted Proximity, Cost
Resource Mapping Optimization for Distributed Cloud Services
Literature Review
Replica Placement (Centralized)
P D Environment Topology Objectives[63] 3 Cloud Unrestricted Prox., Bandwidth, Load Balance[78] 3 Data Grid Multi-Tier Proximity[64] 3 Data Grid Unrestricted Prox., Bandwidth, Availability[66] 3 Cloud-CDN Unrestricted Proximity, Cost[79] 3 Data Grid Tree Prox., Bandwidth, Availability[80] 3 Data Grid Multi-Tier Proximity, Bandwidth[81] 3 Cloud Tree Proximity, Availability[82] 3 Cloud Unrestricted Bandwidth, Load Balance[83] 3 Cloud-CDN Unrestricted Cost, Availability[84] 3 Cloud (Mobile) Unrestricted Proximity, Availability
Resource Mapping Optimization for Distributed Cloud Services
Literature Review
Replica Placement (Decentralized)
P D Environment Topology Objectives[89] Cloud-CDN Unrestricted Proximity[90] 3 Cloud Complete Prox., Cost, Bandwidth, Avail.[91] 3 Data Grid Unrestricted Prox., Cost, Bandwidth, Avail.[92] 3 P2P Unrestricted Cost, Bandwidth, Availability[93] 3 Web Unrestricted Proximity, Bandwidth[94] 3 N/A Multi-Tier Proximity, Bandwidth[88] 3 3 Web Unrestricted Prox., Cost, Load B., Bandwidth[95] 3 3 P2P Unrestricted Proximity, Cost[87] 3 3 Web Unrestricted Proximity, Cost[96] 3 3 Web Unrestricted Proximity, Cost
D-ReP 3 3 Cloud Unrestricted Proximity, Cost, Bandwidth
Resource Mapping Optimization for Distributed Cloud Services
Literature Review
Replica Placement
[90] Bonvin, N., Papaioannou, T.G. and Aberer, K. (2010). A self-organized, fault-tolerant and scalable replica-
tion scheme for cloud storage, Proceedings of the 1st ACM symposium on Cloud computing, pp. 205–216.
Game theoretical, replicate/migrate autonomously, each node is aware ofother replicas, prices, demand, bandwidth, . . .
[95] Shen, H. (2010). An efficient and adaptive decentralized file replication algorithm in P2P file sharing
systems, IEEE Transactions on Parallel and Distributed Systems, 21(6), 827–840.
Replicas on data access path intersections (traffic hubs), traffic analysis[87] Pantazopoulos, P., Karaliopoulos, M. and Stavrakakis, I. (2014). Distributed placement of autonomic
internet services, IEEE Transactions on Parallel and Distributed Systems, 25(7), 1702–1712.
[96] Smaragdakis, G., Laoutaris, N., Oikonomou, K., Stavrakakis, I. and Bestavros, A. (2014). Distributed
server migration for scalable Internet service deployment, IEEE/ACM Trans. on NW, 22(3), 917–930.
Solve FLP in an r-ball, centrality, demand and FLP decisions are broadcasted
Resource Mapping Optimization for Distributed Cloud Services
RalloCloud
Appendices
5 Literature Review
6 RalloCloud
7 Algorithm Details
8 Intra-Cloud Mapping
9 Subgraph Matching
10 Complete Results
Resource Mapping Optimization for Distributed Cloud Services
RalloCloud
Entity Relationship Diagram
Resource Mapping Optimization for Distributed Cloud Services
RalloCloud
Network Modeling
Requested bandwidth between two VMs is allocated from all edges on theshortest path between the clouds to which two VMs are deployed.Three types of latency are considered:
Deployment Latency = M +∑i∈P1
Li +SB
Communication Latency =∑i∈P2
Li +DB
Data Access Latency =∑i∈P3
Li +DB
Resource Mapping Optimization for Distributed Cloud Services
RalloCloud
Cost Modeling
Static Pricing and Trough Filling strategies
Total Service Cost =∑i∈VC
∑j∈AV
C
resource_sizeij · unit_priceij · durationij
+∑i∈EC
resource_sizei · unit_pricei · durationi
+∑
i
replica_sizei · unit_pricei · durationi
Resource Mapping Optimization for Distributed Cloud Services
RalloCloud
Performance Criteria
Criteria
Temporal
Latency
Cloud-to-User Latency
Hop Count
Inter-Cloud Latency
Duration
Completion Time
Execution Time
Makespan
SLA Violations
VolumetricalThroughput
Utilization Rate
Economical
Benefit-Cost Ratio
Resource Cost
Revenue
DistributionalDistribution Rate
Load Balance
OtherAcceptance/Rejection Rate
Fairness
Resource Mapping Optimization for Distributed Cloud Services
Algorithm Details
Appendices
5 Literature Review
6 RalloCloud
7 Algorithm Details
8 Intra-Cloud Mapping
9 Subgraph Matching
10 Complete Results
Resource Mapping Optimization for Distributed Cloud Services
Algorithm Details
TBM
Resource Mapping Optimization for Distributed Cloud Services
Algorithm Details
Replica Discovery
Resource Mapping Optimization for Distributed Cloud Services
Algorithm Details
D-ReP
a
b
c
d
f
e
a1
a2
a3
a4
e1
b3
b2
b1
e4
e2
e3
f3
d1
f1
f2
d2
c1 c2
Resource Mapping Optimization for Distributed Cloud Services
Algorithm Details
User demand locations
a
b
c
d
f
e
a1
a2
a3
a4
e1
b3
b2
b1
e4
e2
e3
f3
d1
f1
f2
d2
c1 c2
Resource Mapping Optimization for Distributed Cloud Services
Algorithm Details
ITERATION 1d: User demand received from c and f
a
b
c
d
f
e
a1
a2
a3
a4
e1
b3
b2
b1
e4
e2
e3
f3
d1
f1
f2
d2
c1 c2
Resource Mapping Optimization for Distributed Cloud Services
Algorithm Details
ITERATION 1d: Cache creation decision
a
b
c
d
f
e
a1
a2
a3
a4
e1
b3
b2
b1
e4
e2
e3
f3
d1
f1
f2
d2
c1 c2
Resource Mapping Optimization for Distributed Cloud Services
Algorithm Details
ITERATION 2f: Migration decision
a
b
c
d
f
e
a1
a2
a3
a4
e1
b3
b2
b1
e4
e2
e3
f3
d1
f1
f2
d2
c1 c2
Resource Mapping Optimization for Distributed Cloud Services
Algorithm Details
ITERATION 2c: Duplication decision
a
b
c
d
f
e
a1
a2
a3
a4
e1
b3
b2
b1
e4
e2
e3
f3
d1
f1
f2
d2
c1 c2
Resource Mapping Optimization for Distributed Cloud Services
Algorithm Details
ITERATION 3e: Migration decision
a
b
c
d
f
e
a1
a2
a3
a4
e1
b3
b2
b1
e4
e2
e3
f3
d1
f1
f2
d2
c1 c2
Resource Mapping Optimization for Distributed Cloud Services
Algorithm Details
ITERATION 3a: Migration decision
a
b
c
d
f
e
a1
a2
a3
a4
e1
b3
b2
b1
e4
e2
e3
f3
d1
f1
f2
d2
c1 c2
Resource Mapping Optimization for Distributed Cloud Services
Algorithm Details
ITERATION 3c: Removal decision
a
b
c
d
f
e
a1
a2
a3
a4
e1
b3
b2
b1
e4
e2
e3
f3
d1
f1
f2
d2
c1 c2
Resource Mapping Optimization for Distributed Cloud Services
Intra-Cloud Mapping
Appendices
5 Literature Review
6 RalloCloud
7 Algorithm Details
8 Intra-Cloud Mapping
9 Subgraph Matching
10 Complete Results
Resource Mapping Optimization for Distributed Cloud Services
Intra-Cloud Mapping
Virtual Machine Cluster Embedding
Allocation of PMs to VMs in a single cloud data center.
Increase resourceutilizationIncrease data centerthroughputIncrease acceptance rateIncrease revenue
#PMs∧i=0
#VMs∑j=0
A[i][j] = 1
#PMs∧
i=0
#Res∧j=0
#VMs∑k=0
A[i][k ]× RR[j][k ]
≤ AR[i][j]
Resource Mapping Optimization for Distributed Cloud Services
Intra-Cloud Mapping
Minimum Span Heuristic
Map a VM to the PM with the maximum resource utilization evenness.Span(p) = maxi∈[1,m] (ri)−mini∈[1,m] (ri)
Fall back to a Mixed Integer Programming (MIP) solution in case of failure.
Resource Mapping Optimization for Distributed Cloud Services
Intra-Cloud Mapping
Pseudo Code
Resource Mapping Optimization for Distributed Cloud Services
Intra-Cloud Mapping
Other Heuristics
SD(v) =
√√√√ m∑i=1
(ri − r̄)2
CD(v) =
∑mi=1∑m
j=1 |ri − rj |2
DM(v) =n∑
i=1
(ri − min
j∈[1,m]
(rj))
SK (v) =
√√√√ m∑i=1
( ri
r̄− 1)2
Resource Mapping Optimization for Distributed Cloud Services
Intra-Cloud Mapping
Results (1/2)
0%
1%
2%
3%
4%
5%
6%
7%
8%
9%
10%
11%
12%
13%
14%
15%
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
3 4 5 6 7 8 9 10 11 12
Impr
ovem
ent
Rat
e
Num
ber
of A
ssig
ned
App
lica
tion
s
VM Count
GR HR Improvement
PM
VM
s
10.8% perfectplacement12.1% more VMs34.5% lessmigrations
Resource Mapping Optimization for Distributed Cloud Services
Intra-Cloud Mapping
Results (2/2)
Resource Mapping Optimization for Distributed Cloud Services
Subgraph Matching
Appendices
5 Literature Review
6 RalloCloud
7 Algorithm Details
8 Intra-Cloud Mapping
9 Subgraph Matching
10 Complete Results
Resource Mapping Optimization for Distributed Cloud Services
Subgraph Matching
Subgraph Matching
Search space is all possible injective matchings from the set of pattern nodesto the set of target nodes.Systematically explore the search space:
Start from an empty matchingExtend the partial matching by matching a non matched pattern node to a nonmatched target nodeBacktrack if some edges are not matchedRepeat until all pattern nodes are matched (success) or all matchings arealready explored (fail).
Filters are necessary to reduce the search space by pruning branches that donot contain solutions.
Resource Mapping Optimization for Distributed Cloud Services
Subgraph Matching
LAD Filtering
Resource Mapping Optimization for Distributed Cloud Services
Subgraph Matching
LAD Filtering
D1 = D3 = D5 = D6 = A,B,C,D,E ,F ,G
D2 = D4 = A,B,D
Resource Mapping Optimization for Distributed Cloud Services
Complete Results
Appendices
5 Literature Review
6 RalloCloud
7 Algorithm Details
8 Intra-Cloud Mapping
9 Subgraph Matching
10 Complete Results
Resource Mapping Optimization for Distributed Cloud Services
Complete Results
Complete Results
Resource Mapping Optimization for Distributed Cloud Services
Complete Results
Complete Results
Resource Mapping Optimization for Distributed Cloud Services
Complete Results
Complete Results
Resource Mapping Optimization for Distributed Cloud Services
Complete Results
Complete Results
Resource Mapping Optimization for Distributed Cloud Services
Complete Results
Complete Results
Resource Mapping Optimization for Distributed Cloud Services
Complete Results
Complete Results
Resource Mapping Optimization for Distributed Cloud Services
Complete Results
Complete Results
Resource Mapping Optimization for Distributed Cloud Services
Complete Results
Complete Results
Resource Mapping Optimization for Distributed Cloud Services
Complete Results
Complete Results
Resource Mapping Optimization for Distributed Cloud Services
Complete Results
Complete Results
[0,2
5)
[25,
50)
[50,
75)
[75,
100)
[100,1
25)
[125,1
50)
[150
,175)
[175
,200)
[200,2
25)
[225,2
50)
[250,2
75)
[275
,300)
[300
,325
)
[325,3
50)
[350,3
75)
[375,4
00)
[400
,425)
[425,4
50)
[450,4
75)
[475,5
00)
[500
,525)
[525
,550
)
[550,5
75)
[575,6
00)
[600,6
25)
[625
,650)
[650,6
75)
[675,7
00)
[700,7
25)
[725
,750)
[750
,775
)
[775,8
00)
[800,8
25)
[825,8
50)
[850
,875)
[875,9
00)
[900,9
25)
[925,9
50)
[950
,975)
[975
,1,0
00)
0
500
1,000
Generated Grid Coordinates
Num
bero
fReq
uest
Pair
s
Uniform DistributionExponential Distribution
Resource Mapping Optimization for Distributed Cloud Services
Complete Results
Complete Results
[0,2
5)
[25,
50)
[50,
75)
[75,
100)
[100,1
25)
[125,1
50)
[150
,175)
[175
,200)
[200,2
25)
[225,2
50)
[250,2
75)
[275
,300)
[300
,325
)
[325,3
50)
[350,3
75)
[375,4
00)
[400
,425)
[425,4
50)
[450,4
75)
[475,5
00)
[500
,525)
[525
,550
)
[550,5
75)
[575,6
00)
[600,6
25)
[625
,650)
[650,6
75)
[675,7
00)
[700,7
25)
[725
,750)
[750
,775
)
[775,8
00)
[800,8
25)
[825,8
50)
[850
,875)
[875,9
00)
[900,9
25)
[925,9
50)
[950
,975)
[975
,1,0
00)
0
1,000
2,000
Generated Grid Coordinates
Num
bero
fReq
uest
Pair
s
Normal Distribution (σ = 50)Normal Distribution (σ = 100)Normal Distribution (σ = 150)
Resource Mapping Optimization for Distributed Cloud Services
Complete Results
Complete Results
[0,2
5)
[25,
50)
[50,
75)
[75,
100)
[100,1
25)
[125,1
50)
[150
,175)
[175
,200)
[200,2
25)
[225,2
50)
[250,2
75)
[275
,300)
[300
,325
)
[325,3
50)
[350,3
75)
[375,4
00)
[400
,425)
[425,4
50)
[450,4
75)
[475,5
00)
[500
,525)
[525
,550
)
[550,5
75)
[575,6
00)
[600,6
25)
[625
,650)
[650,6
75)
[675,7
00)
[700,7
25)
[725
,750)
[750
,775
)
[775,8
00)
[800,8
25)
[825,8
50)
[850
,875)
[875,9
00)
[900,9
25)
[925,9
50)
[950
,975)
[975
,1,0
00)
0
2,000
4,000
Generated Grid Coordinates
Num
bero
fReq
uest
Pair
s
Chi-Squared Distribution (degree o f f reedom = 1)Chi-Squared Distribution (degree o f f reedom = 2)Chi-Squared Distribution (degree o f f reedom = 3)
Resource Mapping Optimization for Distributed Cloud Services
Complete Results
Complete Results
[0,2
5)
[25,
50)
[50,
75)
[75,
100)
[100,1
25)
[125,1
50)
[150
,175)
[175
,200)
[200,2
25)
[225,2
50)
[250,2
75)
[275
,300)
[300
,325
)
[325,3
50)
[350,3
75)
[375,4
00)
[400
,425)
[425,4
50)
[450,4
75)
[475,5
00)
[500
,525)
[525
,550
)
[550,5
75)
[575,6
00)
[600,6
25)
[625
,650)
[650,6
75)
[675,7
00)
[700,7
25)
[725
,750)
[750
,775
)
[775,8
00)
[800,8
25)
[825,8
50)
[850
,875)
[875,9
00)
[900,9
25)
[925,9
50)
[950
,975)
[975
,1,0
00)
0
1,000
2,000
Generated Grid Coordinates
Num
bero
fReq
uest
Pair
s
Pareto Distribution (scale = 1, shape = 3)Pareto Distribution (scale = 1, shape = 4)Pareto Distribution (scale = 1, shape = 5)
Resource Mapping Optimization for Distributed Cloud Services
Complete Results
Complete Results