IEEE Globecom 2013 Workshop on Cloud Computing Systems, Networks, and Applications SLA-driven Dynamic Resource Provisioning for Service Provider in Cloud Computing Yongyi Ran *, Jian Yang, Shuben Zhang, Hongsheng Xi University of Science and Technology of China 12/9/2013 1/15/2014 1
Outline Motivation System Model SLA-Driven Algorithm Performance Evaluation Conclusion 1/15/2014 2
Motivation Characteristic of Cloud Computing Virtualization technology (Resource VM) Cost: pay-per-use, cost saving Scalability and elasticity: on-demand Maintenance is easier Application programming interface (API) accessibility Other: Multitenancy, Reliability, etc. Virtual Machine, Pay-per-User, On-demand, Dynamic Resource Provisioning 1/15/2014 3
Motivation Challenges for Dynamic Resource Provisioning Demand: Fluctuation and uncertainty, hard to predict accurately Cost saving: Hourly fee Over-provisioning, under-provisioning Guaranteeing SLA / QoS: Availability, response time, etc. Challenge: Dynamic Resource Provisioning Strategy Making a Good Tradeoff between Saving Cost and Guaranteeing SLA 1/15/2014 4
Motivation Related Works Resource Provisioning / Allocation Training an offline table for looking up the optimum number of instances for a special workload. [Hong et al. 2012] Disadvantage: Need know the workload statistical characteristic previously For multiple cloud providers, formulating and solving stochastic integer programming with a multistage recourse. [Sivadon et al. 2012] Disadvantage: computation complexity of this method is too high Online, dynamic, requiring no prior workload. knowledge of 1/15/2014 5
Motivation Related Works Service Agreement Level (SLA) In previous works, SLA often specifies : Responsibilities, guarantees, warranties, performance levels in terms of availability, response time, etc. Defining the unavailability probability of the service as a metric of SLA unavailability probability indicates the probability that the workload exceeds the serving capacity 1/15/2014 6
Motivation Our Main Works SLA-driven dynamic VMs provisioning strategy Formulating this problem as minimizing the number of purchased VMs subject to a SLA requirement in terms of the unavailability probability. An online unavailability probability estimation model based on the large deviation principle. An event-based implementation procedure. Experiments are carried out based on two real workload traces. 1/15/2014 7
System Model System Architecture for Dynamic Resource provisioning periodically calculate E-SLA, and then compare with the negotiated SLA (N-SLA) to decide the increase or decrease of the number of VMs. 1/15/2014 8
System Model VM Type & Price 1/15/2014 9
System Model SLA-Based Mathematical Model 1/15/2014 10
SLA-Driven Algorithm SLA-driven Dynamic Resource Provisioning Strategy Unavailability Probability Estimation 1/15/2014 11
SLA-Driven Algorithm SLA-driven Dynamic Resource Provisioning Strategy Unavailability Probability Estimation 1/15/2014 12
SLA-Driven Algorithm SLA-driven Dynamic Resource Provisioning Strategy Unavailability Probability Estimation 1/15/2014 13
SLA-Driven Algorithm SLA-driven Dynamic Resource Provisioning Strategy The Moment Generating Function M(θ) 1/15/2014 14
SLA-Driven Algorithm SLA-driven Dynamic Resource Provisioning Strategy The Moment Generating Function M(θ) 1/15/2014 15
SLA-Driven Algorithm Event-based Implementation Procedure for Dynamic Resource Provisioning 1/15/2014 16
Performance Evaluation Comparing Algorithm the ARMA-based (Autoregressive Moving Average model, ARMA) strategy the ARMA-based strategy with Margins (ARMA- M) the Reactive Strategy the Reactive Strategy with Margins (Reactive-M) 1/15/2014 17
Performance Evaluation Performance Comparison 1/15/2014 18
Performance Evaluation 1/15/2014 19
Performance Evaluation 1/15/2014 20
Performance Evaluation Conclusion Resource Provisioning / Allocation By defining the unavailability probability of the service as a metric of SLA. A SLA-driven dynamic VMs provisioning strategy based on the large deviation principle. Proactively estimate the unavailability probability of the service and adjust the resources (VMs) according to the result of the comparison between the E-SLA and N-SLA. Online, dynamic, requiring no prior workload. knowledge of 1/15/2014 21
Many Thanks! 12/5/2013 22