Modeling and Optimization of Resource Allocation in Cloud [PhD Thesis Progress 2]

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Modeling and Optimization of Resource Allocation in Cloud
PhD Thesis Progress – Second Report
Atakan Aral
Thesis Advisor: Asst. Prof. Dr. Tolga Ovatman
Istanbul Technical University – Department of Computer Engineering
June 22, 2015
Atakan Aral Modeling and Optimization of Resource Allocation in Cloud

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Outline
1 Introduction
Preliminary Information
Contribution to the Thesis
Time Plan
2 Algorithm Design
3 Evaluation
4 Results
5 Conclusion

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Time Plan
Outline
1 Introduction
Time Plan
2 Algorithm Design
3 Evaluation
4 Results
5 Conclusion

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Time Plan
Federated Cloud
Deﬁnition
Mechanisms and policies for scaling hosted services across multiple,
geographically distributed data centers and dynamically coordinating load
distribution among these data centers.
Aims an open and online cloud economy in which providers:
operate as parts of a market driven resource leasing federation;
can dynamically partner with each other to create a seemingly inﬁnite pool of IT
resources.
While users of cloud infrastructure:
avoid vendor lock-in and can easily hybridize their private data center;
can scale VMs across multiple IaaS providers in different geo-locations.

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Time Plan
Distributed VMs
Opportunities:
Available mechanisms and policies such as Federated Cloud;
Very high speed inter-DC communication technologies such as optical fiber;
Programming models that minimize size of data flow between nodes such as
MapReduce
Advantages:
fault tolerance
vendor independence
closer proximity to user base
cost benefits

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Time Plan
Distributed VMs
VM Placement Risks:
Cooperating VMs on distant DCs;
VMs far away from their user base;
VMs placed without considering different pricing strategies of vendors

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Time Plan
Algorithm Design and Implementation
Suggested Topology Based Matching (TBM) algorithm employs a graph
theoretical approach in combination with some heuristics.
Incremental development
Re-evaluation after each new improvement
to compare against baselines
to detect bottlenecks and other problems

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Time Plan
Evaluation
Bandwidth modeling
Cost modeling
Load modeling
Evaluation variables
7 baseline methods, 12 performance criteria, 4 variables

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Time Plan
Documentation
A conference paper to be presented, another journal paper being written
Application for TUBITAK 1002 - Short Term R&D Funding Program
Batch evaluation process which generates and logs results and charts for
each run
Revision control and documentation (https://github.com/atary/RalloCloud/)

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Time Plan
Gantt Chart
2015
1 2 3 4 5 6
Algorithm Design
Implementation
Evaluation
Modiﬁcation
Documentation

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Objective
To decrease deployment delay (by placing VMs close to the broker)
To decrease communication delay (by placing connected VMs to the
neighbour data centers)
To reduce resource costs (by balancing load and avoiding overload in any DC)

Introduction
Algorithm Design
Evaluation
Results
Conclusion
UML Activity Diagram

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Subgraph Matching
Search space is all possible injective matchings from the set of pattern nodes
to the set of target nodes.
Systematically explore the search space:
Start from an empty matching
Extend the partial matching by matching a non matched pattern node to a non
matched target node
Backtrack if some edges are not matched
Repeat until all pattern nodes are matched (success) or all matchings are
already explored (fail).
Filters are necessary to reduce the search space by pruning branches that do
not contain solutions.

Introduction
Algorithm Design
Evaluation
Results
Conclusion
LAD Filtering

Introduction
Algorithm Design
Evaluation
Results
Conclusion
LAD Filtering
D1 = D3 = D5 = D6 = A, B, C, D, E, F, G
D2 = D4 = A, B, D

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Bandwidth Modeling
Bandwidth capacities are modeled not in links but in DCs.
When a link is utilized, same amount of bandwidth is reduced from the DCs in
both sides of the link.
More generally, bandwidth capacities of all the nodes that are on the shortest
path are utilized.
Bandwidth request between two VMs is nonbifurcated. (No path-splitting)

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Bandwidth Modeling
VM1 VM3VM2
VM1
VM2VM3

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Cost Modeling
1 Fixed pricing based on memory, bandwidth and duration.
2 Dynamic pricing via Yield management
Increase the price of the resource that is running low in a DC
Cost = minCost + (maxCost − minCost) ∗ Util

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Load Modeling
Number of Clusters Based on the population density
around each location. Range: 1:16
Number of VMs Based on Poisson distribution: λ = 3
Cluster Topologies Either linear or complete
Arrival Times Uniform random in the range [0, 50)

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Evaluation Variables
VM memory request Data center memory capacity is 64x and each VM requires
memory allocation between 1x and 8x
Link bandwidth request Available bandwidth in each link is 80y and bandwidth
allocation to/from other VMs is between 1y and 8y.
Minimum number of requests Average number of requests from the least
populated location is in the range [2, 16] depending on this variable.
VM network intensity Ratio of local computation and inter-VM communication is
between 3 and 1/3.

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Performance Criteria and Baseline Heuristics
Arbitrary Next-ﬁt (ANF)
Load Balancing (LBG)
Random Choice (RAN)
Latency based Next-ﬁt (LNF)
VM Deployment Latency (Seconds)
VM Communication Latency (Seconds)
Task Completion Time (Hours)
Throughput (MIPS)
Rejection Rate (%)
Cost ($)

Introduction
Algorithm Design
Evaluation
Results
Conclusion
VM memory request
0,00
50,00
100,00
150,00
200,00
250,00
300,00
1 2 3 4 5 6 7 8
VMDeploymentLatency(Seconds)
VM RAM
ANF LBG RAN TBF LNF
0,0
0,5
1,0
1,5
2,0
2,5
3,0
1 2 3 4 5 6 7 8
VMCommunicationLatency(Seconds)
VM RAM
ANF LBG RAN TBF LNF

Introduction
Algorithm Design
Evaluation
Results
Conclusion
VM memory request
0
50
100
150
200
250
300
350
400
1 2 3 4 5 6 7 8
TaskCompletionTime(Hours)
VM RAM
ANF LBG RAN TBF LNF
0
500
1000
1500
2000
2500
3000
1 2 3 4 5 6 7 8
Throughput(MIPS)
VM RAM
ANF LBG RAN TBF LNF

Introduction
Algorithm Design
Evaluation
Results
Conclusion
VM memory request
0
10
20
30
40
50
60
70
80
90
100
1 2 3 4 5 6 7 8
RejectionRate(%)
VM RAM
ANF LBG RAN TBF LNF
0
10000
20000
30000
40000
50000
60000
1 2 3 4 5 6 7 8
Cost($)
VM RAM
ANF LBG RAN TBF LNF

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Link bandwidth request
0,0
0,5
1,0
1,5
2,0
2,5
1 2 3 4 5 6 7 8
VMCommunicationLatency(Seconds)
Link BW
ANF LBG RAN TBF LNF
0
10000
20000
30000
40000
50000
60000
1 2 3 4 5 6 7 8
Cost($)
Link BW
ANF LBG RAN TBF LNF

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Minimum number of requests
0
500
1000
1500
2000
2500
1 2 3 4 5 6 7 8
Throughput(MIPS)
Number of Requests
ANF LBG RAN TBF LNF

Introduction
Algorithm Design
Evaluation
Results
Conclusion
VM network intensity
0
50
100
150
200
250
300
350
1 2 3 4 5 6 7 8
TaskCompletionTime(Hours)
Network intensity
ANF LBG RAN TBF LNF
0
10000
20000
30000
40000
50000
60000
70000
80000
90000
1 2 3 4 5 6 7 8
Cost($)
Network intensity
ANF LBG RAN TBF LNF

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Publications
Aral, A. and Ovatman, T. (2014). Improving resource utilization in cloud
environments using application placement heuristics. In Proceedings of the
4th International Conference on Cloud Computing and Services Science
(CLOSER), pages 527–534.
Aral, A. and Ovatman, T. (2015). Subgraph matching for resource allocation in
the federated cloud environment. In Proceedings of 8th IEEE International
Conference on Cloud Computing (IEEE CLOUD). (to appear)
Aral, A. and Ovatman, T. (2015). Graph theoretical allocation of map reduce
clusters in federated cloud. (for journal submission)

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Planned Studies
Algorithm
Additional constraints (jurisdiction, partially known topology)
Vertical scaling support
Hybrid cloud support
Homeomorphism
Connected components
Evaluation
Signiﬁcance study
Evaluation with topology improvements
Multi-objective optimization
Dynamic heuristic selection, meta-heuristics

Introduction
Algorithm Design
Evaluation
Results
Conclusion
Thank you for your time.

Modeling and Optimization of Resource Allocation in Cloud [PhD Thesis Progress 2]

Recommended

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Modeling and Optimization of Resource Allocation in Cloud [PhD Thesis Progress 2] (20)

More from AtakanAral (14)

Recently uploaded (20)

Modeling and Optimization of Resource Allocation in Cloud [PhD Thesis Progress 2]