Cost-Efficient Scheduling Of Elastic Processes In Hybrid Clouds Philipp Hoenisch, Christoph Hochreiner, Dieter Schuller, Stefan Schulte, Jan Mendling, Schahram Dustdar 

Outline ‣Introduction ‣Motivation ‣Approach ‣Evaluation ‣Conclusion 2 

Introduction 

Business Process Management Execute a sequence of software services to provide a value added service to customers [wes07]. 4 Evaluation Process Mining Business Activity Monitoring Enactment Operation Monitoring Maintenance Configuration System Selection Implementation Test and Deployment Design Business Process Identification and Modeling Analysis Validation Simulation Verification Administration and Stakeholders 

Business Process Management Execute a sequence of software services to provide a value added service to customers [wes07]. 4 Evaluation Process Mining Business Activity Monitoring Enactment Operation Monitoring Maintenance Configuration System Selection Implementation Test and Deployment Design Business Process Identification and Modeling Analysis Validation Simulation Verification Administration and Stakeholders 

Elastic Processes [dus11] Enabled by elasticity aspects from Cloud Computing. 5 Resource Elasticity Quality Elasticity Cost Elasticity 

Motivation 

VM 1 VM 2 VM 3 VM 4 VM 5 VM 6 VM N Private Cloud Public Cloud Motivational Scenario 7 Service 1 Service 2 Service 1 Service 2 Service 1 Service 2 Service 1 Service 2 

Service 1 Service 2 Service 1 Service 2 Service 1 Service 2 Motivational Scenario 8 Service 1 Service 2 Service 5 Service 3 Service 4 Service 6 Service 1 Service 2 VM 1 VM 2 VM 3 VM 4 VM 5 VM 6 VM N Private Cloud Public Cloud Service 1 Service 2 Service 3 Service 4 Service 5 Service 6 

Service 1 Service 2 Service 1 Service 2 Service 1 Service 2 Motivational Scenario 8 Service 1 Service 2 Service 5 Service 3 Service 4 Service 6 Service 1 Service 2 VM 1 VM 2 VM 3 VM 4 VM 5 VM 6 VM N Private Cloud Public Cloud Service 1 Service 2 Service 3 Service 4 Service 5 Service 6 

Service 1 Service 2 Service 1 Service 2 Service 1 Service 2 Motivational Scenario 8 Service 1 Service 2 Service 5 Service 3 Service 4 Service 6 Service 1 Service 2 VM 1 VM 2 VM 3 VM 4 VM 5 VM 6 VM N Private Cloud Public Cloud Service 1 Service 2 Service 3 Service 4 Service 5 Service 6 

Service 1 Service 2 Service 1 Service 2 Service 1 Service 2 Motivational Scenario 8 Service 1 Service 2 Service 5 Service 3 Service 4 Service 6 Service 1 Service 2 VM 1 VM 2 VM 3 VM 4 VM 5 VM 6 VM N Private Cloud Public Cloud Service 1 Service 2 Service 3 Service 4 Service 5 Service 6 

Service 1 Service 2 Service 1 Service 2 Service 1 Service 2 Motivational Scenario 8 Service 1 Service 2 Service 5 Service 3 Service 4 Service 6 Service 1 Service 2 VM 1 VM 2 VM 3 VM 4 VM 5 VM 6 VM N Private Cloud Public Cloud Service 1 Service 2 Service 3 Service 4 Service 5 Service 6 

Service 1 Service 2 Service 1 Service 2 Service 1 Service 2 Motivational Scenario 8 Service 1 Service 2 Service 5 Service 3 Service 4 Service 6 Service 1 Service 2 VM 1 VM 2 VM 3 VM 4 VM 5 VM 6 VM N Private Cloud Public Cloud Service 1 Service 2 Service 3 Service 4 Service 5 Service 6 

Motivational Scenario 9 Service 1 Service 2 Service 5 Service 3 Service 4 Service 6 VM 1 VM 2 VM 3 VM 4 VM 5 VM 6 VM N Private Cloud Public Cloud Service 1 Service 2 Service 3 Service 4 Service 5 Service 6 

Motivational Scenario 9 Service 1 Service 2 Service 5 Service 3 Service 4 Service 6 VM 1 VM 2 VM 3 VM 4 VM 5 VM 6 VM N Private Cloud Public Cloud Service 1 Service 2 Service 3 Service 4 Service 5 Service 6 

Motivational Scenario 9 Service 1 Service 2 Service 5 Service 3 Service 4 Service 6 VM 1 VM 2 VM 3 VM 4 VM 5 VM 6 VM N Private Cloud Public Cloud Service 1 Service 2 Service 3 Service 4 Service 5 Service 6 

Motivational Scenario 9 Service 1 Service 2 Service 5 Service 3 Service 4 Service 6 VM 1 VM 2 VM 3 VM 4 VM 5 VM 6 VM N Private Cloud Public Cloud Service 1 Service 2 Service 3 Service 4 Service 5 Service 6 

Motivational Scenario 9 Service 1 Service 2 Service 5 Service 3 Service 4 Service 6 VM 1 VM 2 VM 3 VM 4 VM 5 VM 6 VM N Private Cloud Public Cloud Service 1 Service 2 Service 3 Service 4 Service 5 Service 6 

Challenges ‣Minimize the total process execution costs ‣ Collocate process steps with intense data interaction ‣ Achieve efficient resource usage across multiple clouds ‣ Avoid SLA violations ‣Manage a large number of concurrent processes ‣Prioritize private clouds for service execution 10 

Approach 

Service Instance Placement Problem (SIPP) [hoe15] Mixed Integer Linear Programming solution based on CPLEX 12 

Service Instance Placement Problem (SIPP) [hoe15] Mixed Integer Linear Programming solution based on CPLEX 12 ‣Minimize the total process execution costs ‣ Collocate process steps with intense data interaction ‣ Achieve efficient resource usage across multiple clouds ‣ Avoid SLA violations ‣Manage a large number of concurrent processes ‣Prioritize private clouds for service execution 

Service Instance Placement Problem (SIPP) [hoe15] Mixed Integer Linear Programming solution based on CPLEX 12 ‣Minimize the total process execution costs ‣ Collocate process steps with intense data interaction ‣ Achieve efficient resource usage across multiple clouds ‣ Avoid SLA violations ‣Manage a large number of concurrent processes ‣Prioritize private clouds for service execution within one cloud across multiple clouds 

Service Instance Placement Problem (SIPP) [hoe15] Mixed Integer Linear Programming solution based on CPLEX 12 ‣Minimize the total process execution costs ‣ Collocate process steps with intense data interaction ‣ Achieve efficient resource usage across multiple clouds ‣ Avoid SLA violations ‣Manage a large number of concurrent processes ‣Prioritize private clouds for service execution within one cloud ✓ across multiple clouds 

Service Instance Placement Problem (SIPP) [hoe15] Mixed Integer Linear Programming solution based on CPLEX 12 ‣Minimize the total process execution costs ‣ Collocate process steps with intense data interaction ‣ Achieve efficient resource usage across multiple clouds ‣ Avoid SLA violations ‣Manage a large number of concurrent processes ‣Prioritize private clouds for service execution ✓ within one cloud ✓ across multiple clouds 

Service Instance Placement Problem (SIPP) [hoe15] Mixed Integer Linear Programming solution based on CPLEX 12 ‣Minimize the total process execution costs ‣ Collocate process steps with intense data interaction ‣ Achieve efficient resource usage across multiple clouds ‣ Avoid SLA violations ‣Manage a large number of concurrent processes ‣Prioritize private clouds for service execution ✓ ✓ within one cloud ✓ across multiple clouds 

Extensions to SIPP ‣Extend the basic model to consider an arbitrary amount of Clouds ‣Consider data transfer costs ‣Allow for a prioritization of selected Clouds ‣Collocate process steps with intense data interaction in the same Cloud 13 

Evaluation 

Evaluation Prototype (ViePEP) [sch13] 15 Process Manager Resource Manager Reasoner Service 1 Service 2 Service 5 Service 3 Service 4 Service 6 ViePEP 

Evaluation ‣10 process models out of SAP R/3 with different complexity ‣10 service models with different CPU usage, makespan and outgoing data transfer ‣2 SLA scenarios with different maximal durations ‣X-VM per service in an ad-hoc manner as baseline [fri14] 16 

Process request pattern 17 # of Invocations 0 3 5 8 10 Minutes 1 5 10 15 20 Burst arrival Pyramid arrival 

Evaluation Results 18 Cost 0 750 1500 2250 3000 Baseline Optimized Resource Leasing Cost Data Transfer Cost SLA Violation Cost Burst arrival 

Evaluation Results 18 Cost 0 750 1500 2250 3000 Baseline Optimized Resource Leasing Cost Data Transfer Cost SLA Violation Cost Burst arrival - 92 % data transfer cost - 61 % total cost 

Evaluation Results 18 Cost 0 750 1500 2250 3000 Baseline Optimized Resource Leasing Cost Data Transfer Cost SLA Violation Cost Burst arrival Cost 0 1250 2500 3750 5000 Baseline Optimized Pyramid arrival - 92 % data transfer cost - 61 % total cost 

Evaluation Results 18 Cost 0 750 1500 2250 3000 Baseline Optimized Resource Leasing Cost Data Transfer Cost SLA Violation Cost Burst arrival Cost 0 1250 2500 3750 5000 Baseline Optimized Pyramid arrival - 85 % data transfer cost - 44 % total cost - 92 % data transfer cost - 61 % total cost 

Conclusion 

Contributions ‣Demonstrate that considering data transfer costs can decrease the total execution cost ‣Provide a model to optimize elastic process execution across more than one Cloud ‣Provide a model to collocate process steps with intense data communication within one Cloud ‣Provide a model to prioritize one cloud instance while maintaining a high Quality of Service 20 

Future Work ‣Consider privacy restrictions for service executions ‣Support more complex system architectures and data communication pattern ‣Consider stateful services ‣Evaluate different resource pricing and SLAs 21 

Q & A Christoph Hochreiner [email protected] 

References [dus11] S. Dustdar, Y. Guo, B. Satzger and Truong, Hong-Linh. Principles of Elastic Processes. IEEE Internet Computing 15, 5 (2011). [fri14] S. G. Frincu and J. Gossa. On the Efficiency of Several VM Provisioning Strategies for Workflows with Multi-threaded Tasks on Clouds. IEEE Computing, vol. 96 (2014). [hoe15] P. Hoenisch, D. Schuller, S. Schulte, C. Hochreiner and S. Dustdar. Optimization of Complex Elastic Processes. IEEE Transactions on Services Computing (2015). [sch13] S. Schulte, P. Hoenisch, S. Venugopal and S. Dustdar. Introducing the Vienna Platform for Elastic Processes. ICSOC Workshops (2013). [wes07] M. Weske. Business process management: concepts, languages, architectures. Springer Heidelberg Dordrecht London New York (2007). 23