Generic cloud platform multi-objective optimization leveraging models@run.time

Abstract

Cloud computing promises scalable hosting by offering an elastic management of virtual machines which run on top of hardware data centers. This elastic management as a cornerstone of PaaS (Platform As A Service) has to deal with trade-offs between conflicting requirements such as cost and quality of service. Solving such trade-offs is a challenging problem. Indeed, most of PaaS providers consider only one optimization axis or ad-hoc multi-objective resolution techniques using domain specific heuristics.This paper aims at proposing a generic approach to build cloud optimization by combining modeling and search based paradigms. Our approach is two-fold: 1) To reason about a cloud environment, we use a Models@run.time approach to have an abstraction layer of a cloud configuration that supports monitoring capabilities and represents cloud intrinsic parameters like cost, load information, etc.. 2) We use a search-based algorithm to navigate through cloud candidate configuration solutions in order to solve the Cloud Multiobjective Optimization Problem (CMOP).We validate our approach based on a case study that we define with our cloud provider partner EBRC 1 as representative of a dynamic management problem of heterogeneous distributed cloud nodes. We implement a prototype of our PaaS supervision framework using Kevoree, a Models@run.time platform. The prototype 1

Self-Adaptive Trade-off Decision Making for Autoscaling Cloud-Based Services

“…The best decision is selected as indicated by the weighted-sum formulation. • MOGA -A most commonly used multi-objective genetic algorithm derived from NSGA-II, e.g., [7] [8] [17]. We have also designed MOGA to benefit from our QoS modeling [20] [21] and region controlling technique [22].…”

“…We now evaluate elasticity of the proposed approach by means of over-and under-provisioning using (17). Table 8 shows the average results for all managed service-instances and VMs on the master PM.…”

“…Similarly, [7] have also leveraged on NSGA-II but they have additionally considered the reduction of harmonic QoS objectives. [17] is another example that use NSGA-II, but the authors apply a weaker form of paretodominance, namely epsilon dominance, where it requires a static epsilon value as the minimum number of better objectives that a non-dominated decision needs to achieve. However, all these approaches apply pure pareto-dominance to evaluate the overall quality of decisions for all the objectives during the optimization, which will lead to a large set of non-pareto-dominated decisions.…”

“…In other words, the optimality and diversity of the resulted tradeoffs decisions tend to be limited and therefore causing it difficult to achieve well-compromised trade-offs. There is a limited amount of work that leverage on the notion of multiobjective optimization [7] [8] [17] and pareto-dominance [18] based sort. Most commonly, they apply Multi-Objective Genetic Algorithm (MOGA), e.g., NSGA-II [19], to search the trade-offs decisions without explicitly using weights.…”

“…Majority of the existing multi-objective optimization work [7] [8] [17] apply MOGA, most commonly the NSGA-II [19], to make decision for autoscaling (we will experimentally compare against them in Section 6). They have ignored QoS interference and thus only considered a limited number of objectives (i.e., up to 4).…”