1. Robust Cloud Resource Provisioning
for Cloud Computing Environments
presented by
Sivadon Chaisiri1, Bu-Sung Lee1,2, and Dusit Niyato1
1
School of Computer Engineering, Nanyang Technological University, Singapore
2
HP Labs Singapore
presented in
IEEE International Conference on Service-Oriented Computing and Applications
(SOCA’10) Perth, Australia, December 14, 2010

2. Outline
●
Overview of Cloud Computing
●
Provisioning plans
●
Challenge of Resource Provisioning
●
Robust Cloud Resource Provisioning
●
Modeling the RCRP
●
Formulating the RCRP
●
Numerical Studies
●
Conclusion
2

3. Overview of Cloud Computing
Hardware
Software infrastructure
Pool of resources
Cloud Computing
• Large distributed system
Physical compute resources • Large pool of resources
Storage Network
• Multiple providers
• Amazon EC2
Cloud Cloud Cloud Cloud Cloud • GoGrid
provider provider provider provider provider
• Rackspace
Cloud Computing • Virtualization (e.g., IaaS)
• Internet access
• Pay-per-use basis
• Provisioning plans
Cloud
consumer
Cloud
consumer
Cloud
consumer
Cloud
consumer
Cloud
consumer
• On-demand
• Reservation
3

4. Provisioning Plans
●
On-demand plan offered by Amazon EC2
●
Reservation plan offered by Amazon EC2
●
Reservation can reduce the total provisioning cost
●
On-demand (Small Instance): 0.085x365x24 = $744.60 for 1yr contract
●
Reservation: 227.50+(0.03x365x24) = $490.30 for 1yr contract or
34.15% cheaper but 49.04% cheaper for 3yr contract
4

5. Challenge of Resource Provision
●
Goal: How many VMs do we need to provision in advance to
minimize the total cost under uncertainty?
●
Challenge:
●
Multivariate uncertainty e.g., price, demand, availability, etc.
●
Unavoidable under- and overprovisioning costs
●
Multiple providers + service-level-agreements (SLAs)
Decision Realization
Reserve N VMs Utilize N VMs (no more cost)
Actual demand is N VMs
(a) Best provisioning
Decision Realization Decision Realization
No need
Reserve N VMs Utilize N/2 VMs
Reserve N VMs Provision 2N VMs on-demand
Utilize N VMs on-demand provisioning
(on demand cost) (oversubscribed cost)
Actual demand is 3N VMs Actual demand is N/2 VMs
(b) Underprovisioning problem (c) Overprovisioning problem
5

6. Robust Cloud Resource Provisioning
●
RCRP algorithm is proposed
●
Minimize the expected resource provisioning cost
●
Reduce on-demand & oversubscribed costs
●
Consider multivariate uncertainty
●
Meet the decision maker’s risk preference: most decision
makers are risk averse
●
Two types of robustness
●
Solution robustness: solution is almost optimal
●
Model robustness: penalty is almost avoided
6

7. Modeling the RCRP
●
System model of cloud computing
7

8. Modeling the RCRP (cont...)
●
Multiple IaaS-based cloud providers
●
Provisioning plans: reservation & on-demand
●
Each cloud provider offers different plans, prices, and
service-level-agreement (SLA)
●
VM class = group of VMs executing the same job
●
Each VM class requires different resources
●
Demand = the number of VMs of specific VM class
required to execute the cloud consumer's job
8

9. Modeling the RCRP (cont...)
●
Provisioning phases: reservation, expending, on-demand
●
Two provisioning stages (namely first and second)
●
Uncertain parameter is described by probability distribution
●
Realization = observed uncertain parameter
●
Recourse action = the action corresponding to certain
realization
●
(Optimal) Solution consists of
●
The number of reserved VMs provisioned for each VM class
●
A collection of recourse actions
9

10. Formulating the RCRP
• Complete RCRP model
10

11. Formulating the RCRP (cont…)
• Multi-criteria optimization
• Total resource provisioning cost:
• Solution robustness: cost of deviation with weight :
• Model robustness: penalty function cost with weight :
11

12. Formulating the RCRP (cont...)
Solution robustness Model robustness
●
Adjustment of weights to meet the risk preference
• Weighting to adjust the solution robustness
• Guideline for adjusting the model robustness
●
Weighting and to adjust the model robustness:
●
: overprovisioning weights
●
: underprovisioning weights
●
Simplifying over- and underprovisioning weights
●
Let
●
where
12

13. Numerical Studies: Parameter Setting
• Two VM classes (I1 and I2) require difference resources
• Max resource capacity offered by cloud providers (J1 to J4):
● J1 (private cloud) offers limited resources but zero on-demand cost
● J2 to J4 (public clouds) offer abundant resources
• Pricing defined by each cloud provider:
• Three types of uncertain parameters are considered
– Types: user's demand, resource price, resource availability
– Each type is described by different probability distribution
• RCRP and other models are implemented and solved by GAMS/CPLEX
13

14. Numerical Studies: Results
14

15. Numerical Studies: Results (cont...)
●
Comparison between RCRP and others
●
Summary of the comparison:
●
NoRes yields the highest total cost
●
MaxRes has zero on-demand but highest oversubscribed
●
EVU gains the lowest oversubscribed but high on-demand
●
OVMP achieves the minimum total cost
●
RCRP is more flexibly controlled and it can achieve the
total cost close to OVMP
15

16. Conclusion
●
Due to uncertainty, inefficiency of resource
provisioning can lead to very expensive costs
●
RCRP is proposed to minimize the total provisioning
cost, while uncertainty is considered
●
RCRP can achieve both solution- and model-
robustness
●
RCRP can meet decision makers' risk preferences
●
RCRP can be applied in the real practice
●
Future work: sampling techniques and real practice
will be performed
16

17. THANK YOU
17

18. Formulating the RCRP (cont…)
• Stochastic programming (SP) model
• This SP could only satisfy low-risk decisions
• SP cannot be adjusted to meet the risk preference
18

19. Numerical Studies: Results (cont...)
How to choose the appropriate solution?
1) Apply goal programming based on a predefined goal such as
●
Expected reservation cost <= $1,200
●
Expected on-demand cost <= $1,000
●
Stand deviation of RO must be less than SP
2) Vary the weights and solve the RCRP until the goal is met
Selected solution: = 1 and =1
19