Palladio Optimization Suite: QoS optimization for component-based Cloud applications
- 1. FP7-ICT-2011-8-318484www.modaclouds.eu www.modaclouds.eu Palladio Optimization Suite: QoS optimization for component-based Cloud applications Michele Ciavotta, Danilo Ardagna Politecnico di Milano, Dipartimento di Elettronica, Informazione e Bioingegneria Anne Koziolek Karlsruhe Institute of Technology, Institute for Program Structures and Data Organization
- 2. 2 www.modaclouds.eu FP7-ICT-2011-8-318484 Introduction • Modern software applications have evolved in terms of size and scope • Design an application in the best possible way is crucial • It is often a manual process = arduous and time-consuming • Explore the space of design alternatives and (cloud) services • Assess the QoS of several Software Architectures (SAs) • Specific models and tools have been created to predict the QoS • Analytical and Simulation based solvers • Not user friendly
- 3. 3 www.modaclouds.eu FP7-ICT-2011-8-318484 Palladio Optimization Suite • A collection of complementary plugins (at the moment 2) • Running atop Palladio Bench (graphical interface and M2M transf.) • Automatic exploration of the space of possible architectures • Advanced exploration paradigms • Evolutionary Algorithm • Local Search
- 4. 4 www.modaclouds.eu FP7-ICT-2011-8-318484 • Multi-objective optimization of component- based applications • A Pareto front of Nondominated solutions • Evolutionary algorithm to explore the architectural space • An initial population from a candidate solution defined by the user in Palladio Component Model (PCM) format • The individuals are then modified along degrees of freedom. • Server Farm Configuration • Component Selection • Component Allocation PerOpteryx SPACE4Clouds • Optimization of Cloud architectures • Hybrid approach: Mathematical model + Local search based engine • Cost Optimization under QoS and Architectural constraints • LINE Solver • Percentiles are supported • Random environments • Optimization over a 24-hour time horizon. • Variable workload • Elasticity
- 5. 5 www.modaclouds.eu FP7-ICT-2011-8-318484 Objectives • Minimize • Application Cost (one hour) • Response time (2 components) • Maximize • Throughput (2 components) Possible decisions: • Allocation of 9 software components • Different VM types per component groups Workflow - Phase 1: PerOpteryx Solver • SimuCom (Simulation) Solution: • Hourly cost: 1.29 $ • Avg. Resp. Time: 0.30 s • Avg. Throughput: 9.9 req./s • Number of tiers: 5 Workflow – Phase 1
- 6. 6 www.modaclouds.eu FP7-ICT-2011-8-318484 Workflow – Phase 2 Workflow – Phase 2: Space4Cloud Possible decisions: • Type of VM for each application tier • Number of VM for each our of the day Objective • Minimization of the daily cost under variable workload Constraints • Average response time < 0.6 s • 95-th percentile < 1 s Solvers • MILP solver (Cplex, CBC, etc) • LINE solver (for LQN models)
- 7. 7 www.modaclouds.eu FP7-ICT-2011-8-318484 Conclusions • Suite for multi-attribute QoS optimization of component based cloud applications • Combination of an evolutionary optimization with a local-search-based approach • Time-varying workload and the distinctive traits of the cloud are considered Future work • More work on the integration of the two tools • Validation of the results on industrial settings • Extension to data-intensive applications
Editor's Notes
- #3: Modern applications have greatly changed, in size and scope. In many cases they are web-based with several of services and microservices deployed in different region with many databases and cache systems. Design them in the best possible way is crucial because an error made in the early stages of design can cost a lot amount of money later. Software design is somehow a kind of art. There are pattern to follows but they cannot guarantee important QoS characteristics. Everything has to be done manually, many design alternatives and services must be considered and the quality of Service has to be predicted. Yes, but how? It is true that there are models and tools to assess the quality of a software platform (QN, Petri nets) that can be solved both analytically or by simulation. This approach is not user friendly and goes beyond the normal skills of a software architect.
- #4: The palladio Optimization suite is a collection of complementary plugins that have in common the idea of optimize the architecture of an application at desing time. At the moment there are 2 plugins in the collection, namely PerOpteryx and Space4Clouds. Both run atop Palladio bench, exploiting its metamodels, the grafical interface and the M2M transformation to queuing models. Both plugins deal with the exploration of the space of possible architectures for a software architecture, but while PO is related to in-house application, S4C refers only to Cloud application with IaaS and PaaS services Both implements Advanced exploration paradigms in particular Genetic algorithms for peropteryx and Local search based appaorch is implemented withing S4C
- #5: PO is tool of Multi-objective optimization of component-based applications It works with population of candidate solutions and returns a Pareto front of Nondominated solutions is returned Implements the NSGA-II evolutionary algorithm to explore the architectural space An initial population from a candidate solution defined by the user in Palladio Component Model (PCM) format The individuals are then modified along degrees of freedom. Server Farm Configuration Component Selection Component Allocation Design-time QoS assessment and optimization of Cloud applications Hybrid approach: Mathematical model + Local search based engine Cost Optimization under QoS and Architectural constraints LINE Solver LQNS solver Percentiles are supported Random environments Optimization over a 24-hour time horizon. Variable workload Elasticity
- #7: The running costs is 47.5 $