PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey and Taxonomy of Self-Aware and Self-Adaptive Cloud Autoscaling Systems

Tao Chen, Rami Bahsoon, Xin Yao

arXiv:1609.03590v4cs.SEcs.DC

TL;DR

Cloud autoscaling must adapt to dynamic and uncertain conditions, yet SSCAS research is not mature enough for reliable exploitation. This paper surveys SSCAS and builds a taxonomy; it finds recurring approaches across awareness, architecture, QoS modeling, control granularity, and decision making, while identifying open challenges.

  • Problem

    SSCAS research is not mature enough for reliable exploitation, despite the need to adapt to dynamic and uncertain cloud conditions.

    Source: §1 · pp. 2–3

  • Method

    The paper surveys SSCAS research and organizes it in a taxonomy spanning self-* capabilities and key autoscaling design aspects.

    Source: §6 · p. 29

  • Results

    Self-configuring and self-optimizing, feedback loops, analytical and machine-learning QoS models, service/application-level control, and search-based optimization are prominent in SSCAS.

    Source: §6 · p. 29

  • Takeaways & Limitations

    The survey identifies open challenges for future SSCAS research, including knowledge representation, QoS interference, and broader real-world evaluation.

    Source: §1 · p. 3

    QoS interference is often overlooked in SSCAS QoS modeling and decision making.

    Source: §1 · p. 3

Abstract

from arXiv · show

Autoscaling system can reconfigure cloud-based services and applications, through various configurations of cloud software and provisions of hardware resources, to adapt to the changing environment at runtime. Such a behavior offers the foundation for achieving elasticity in modern cloud computing paradigm. Given the dynamic and uncertain nature of the shared cloud infrastructure, cloud autoscaling system has been engineered as one of the most complex, sophisticated and intelligent artifacts created by human, aiming to achieve self-aware, self-adaptive and dependable runtime scaling. Yet, existing Self-aware and Self-adaptive Cloud Autoscaling System (SSCAS) is not mature to a state that it can be reliably exploited in the cloud. In this article, we survey the state-of-the-art research studies on SSCAS and provide a comprehensive taxonomy for this field. We present detailed analysis of the results and provide insights on open challenges, as well as the promising directions that are worth investigated in the future work of this area of research. Our survey and taxonomy contribute to the fundamentals of engineering more intelligent autoscaling systems in the cloud.

1 INTRODUCTION

The paper frames autoscaling as a response to dynamic, uncertain cloud conditions and surveys self-aware and self-adaptive cloud autoscaling systems (SSCAS). It summarizes common approaches and identifies open challenges for future research.

1 INTRODUCTION

  • Cloud autoscaling must adapt software configurations and hardware resources under workload uncertainty and QoS interference among co-located services.Source: §1 · p. 2
  • Stimulus-, time-, and goal-awareness are the most commonly considered self-awareness levels, while self-configuring and self-optimizing are the most common self-adaptivity notions.Source: §1 · p. 2
  • Feedback-loop architectures, analytical and machine-learning QoS models, service/application-level control, and explicit optimization are prominent in SSCAS research.Many decision approaches assume a single objective or aggregate objectives with weighted sums; QoS interference is absent from many studies.Source: §1 · p. 3
  • Open challenges include representing required knowledge, addressing QoS interference, coordinating software and hardware adaptation, and broadening control granularity and objectives.The authors also call for arbitrary resource combinations and more real-world cases.Source: §1 · p. 3

2 BACKGOUND

The background describes autoscaling architecture and distinguishes self-adaptivity from self-awareness. It introduces their respective capabilities as context for SSCAS.

2.1 Autoscaling System in Cloud

  • Autoscaling separates a managing part with an engine from the manageable services and VMs or containers, using external adaptation.A typical engine uses monitoring and scaling in a feedback loop, and may apply vertical or horizontal actions.Source: §2.1 · pp. 3–4

2.2 Self-Adaptivity in Sofware Systems

  • Self-adaptivity means changing behavior in response to perceived uncertainty and system state, with self-configuring, self-healing, self-optimizing, and self-protecting as its four properties.Source: §2.2 · p. 4

2.3 Self-Awareness in Sofware Systems

  • Self-awareness concerns acquiring knowledge about the system and environment; its levels range from stimulus and interaction awareness to time, goal, and meta-self-awareness.Source: §2.3 · p. 5

3 COMPARISON TO RELATED SURVEYS

The paper distinguishes its survey from reviews of cloud resource management, infrastructure techniques, web applications, and QoS modeling. It emphasizes an application-agnostic focus on SSCAS and a finer taxonomy of its methods and self-* properties.

3 COMPARISON TO RELATED SURVEYS

  • Unlike broad cloud resource-management and QoS surveys, this work focuses on automatically scaling software configurations and hardware provisioning to handle dynamic workloads.Source: §3 · p. 5
  • Compared with infrastructure-level autoscaling reviews and web-application surveys, this survey targets logical architecture and algorithms and is application agnostic.It relates approaches to finer-grained levels of self-adaptivity than a coarse self-adaptive/not-self-adaptive distinction.Source: §3 · pp. 5–6
  • The survey correlates approaches with self-awareness, self-adaptivity, and required knowledge, classifies studies in a taxonomy, and discusses broader open challenges.Source: §3 · p. 6

4 TAXONOMY AND SURVEY RESULTS FOR SSCAS

The survey reviews SSCAS through research questions and a taxonomy of its logical aspects. Its findings describe how self-awareness and self-adaptivity levels are represented in the reviewed studies and identify gaps in knowledge discussion.

4.1 Review Process and Research Qestions

  • The review covers 109 studies selected through database searches and inclusion and exclusion criteria, guided by five research questions about SSCAS.Source: §4.1 · p. 6

4.2 A Taxonomy of SSCAS Research

  • The taxonomy organizes SSCAS around self-awareness, self-adaptivity, architecture, QoS modeling, control granularity, and decision making.These aspects structure the survey’s analysis of the field.Source: §4.2 · pp. 7–8
  • The taxonomy treats QoS modeling as relating quality to environmental conditions and software or hardware control primitives, with model selection and QoS interference among its challenges.Source: §4.2 · pp. 7–8

4.3 The Levels of Self-Awareness and Self-Adaptivity in Cloud Autoscaling Systems

  • Stimulus-awareness appears in all reviewed studies, while interaction- and meta-awareness are less studied; self-configuring and self-optimizing predominate among adaptivity levels.No study explicitly aims for self-protecting, and only 33% discuss required architectural knowledge implicitly or explicitly.Source: §4.3 · pp. 8–9, Figure 3 · p. 9

RQ2: What are the architectural paterns used for engineering SSCAS?

The survey classifies SSCAS architectures as feedback loop, Observe-Decide-Act (ODA), or Monitor-Analyze-Plan-Execute (MAPE), finding feedback loops predominant. It describes how each pattern organizes control and adaptation.

RQ2: What are the architectural paterns used for engineering SSCAS?

4.4.1 Feedback Loop.

  • Single-loop feedback is the simplest and most common pattern, while multiple loops can separate global and local controls.Single loops may add QoS modeling or workload prediction; multiple loops can operate at different architectural levels.Source: §4.4.1 · p. 9, Architectural Pattern · p. 10

4.4.2 Observe-Decide-Act.

  • ODA extends feedback control by assigning steps to developer and runtime roles, linking human activities to adaptive behavior.MNEMOS uses a system monitor, portfolio scheduler, and VM manager as observer, decider, and executor, respectively.Source: §4.4.2 · p. 10

4.4.3 Monitor-Analyze-Plan-Execute.

  • MAPE divides decision-making into analysis and planning, and can use a knowledge component to support adaptation.Examples include single loops and nested loops that adapt software primitives before hardware primitives.Source: §4.4.3 · p. 11

4.5 QoS Modeling

The survey reviews QoS modeling approaches used in SSCAS and reports that analytical and machine-learning models predominate. It also finds limited modeling of QoS interference and dynamic changes to model inputs.

4.5 QoS Modeling

  • QoS models relate quality attributes to control and environmental primitives, but some SSCAS approaches instead use model-free methods.The survey divides analytical modeling into queuing, dependability, and black-box models.Source: §4.5 · p. 11, §4.5.1 · p. 13
  • Analytical models (43%) and machine-learning models (38%) predominate; QoS interference is modeled in 13% of studies, and fully dynamic models in 7%.Semi-dynamic models account for 54% and static models for 39%.Source: Figure 5 · p. 11, §4.5 · p. 12
  • The most explicitly modeled output count is three and the most explicitly considered input count is four; 65% of studies tested only certain inputs or outputs.Source: §4.5 · p. 13

4.5.1 Analytical Modeling.

  • Analytical approaches include single or multiple queues, layered queuing networks, dependability models, and empirical black-box models.Queueing examples relate response time or throughput to resources and workload; layered models can represent request dependencies.Source: §4.5.1 · pp. 13–14

4.5.2 Simulation Based Modeling.

  • Simulation-based QoS models are generally built offline and can require domain experts to set up, analyze, and profile simulation data.The survey describes tools for modeling applications, resource allocation, and cloud-scale autoscaling.Source: §4.5.2 · p. 14

4.5.3 Machine Learning Based Modeling.

  • Machine-learning approaches include linear and nonlinear models; nonlinear methods capture complex correlations but can incur greater overhead.The survey also describes ensemble approaches and comparisons of learning algorithms; studies capturing QoS interference in linear models are limited and focus on VM level.Source: §4.5.3 · p. 15

4.5.4 Hybrid Modeling.

  • Hybrid models combine analytical methods with learning, such as using a Kalman filter to estimate time-varying coefficients in a layered queueing model.Another example combines a queue for workload-to-response-time modeling with ARMA for response-time-to-CPU modeling.Source: §4.5.4 · pp. 15–16

4.5.5 Dynamic Primitives Selection.

  • Primitive selection is usually manual and offline, while some approaches select inputs dynamically using relevance and redundancy criteria.Workload and demand models commonly use historical data in a single dimension, limiting their ability to reason about adaptation effects and trade-offs.Source: §4.5.5 · p. 16, §4.5.6 · p. 16

RQ4: What is the granularity of control in SSCAS?

The survey categorizes SSCAS control by the level at which decisions are made, from services and applications to the cloud. Service/application control is most common, with a trade-off between finer-grained control overhead and global optimization.

RQ4: What is the granularity of control in SSCAS?

4.6.1 Controlling at Service and Application Level.

  • Service/application control is the finest granularity and can target an application, an application tier, or a cloud-based service.Studies often control QoS and/or cost for each application in isolation.Source: §4.6.1 · p. 17

4.6.2 Controlling at Virtual Machine and Container Level.

  • VM-level control makes decisions for each VM, including studies that treat an application tier and its VM as equivalent.PM-level control manages QoS interference from co-hosted VMs by controlling them collectively.Source: §4.6.2 · pp. 17–18, §4.6.3 · p. 18

4.6.4 Controlling at Cloud Level.

  • At cloud level, global controllers manage objectives such as utility, profit, availability, cost, or QoS across the cloud.Some studies instead use decentralized control and consensus to manage cloud-wide QoS and power consumption.Source: §4.6.4 · p. 18
  • Hybrid control combines levels to balance local objectives and global benefit; one approach changes the active granularity at runtime according to objective dependency.Source: §4.6.5 · p. 18

RQ5: What are the approaches used for decision making in SSCAS?

The survey compares three decision-making families for SSCAS and finds search-based optimization most prevalent, while identifying limited treatment of objective trade-offs and QoS interference. It also reviews scaling experiments and implementation dependencies.

RQ5: What are the approaches used for decision making in SSCAS?

  • 63% of studies do not explicitly trade off objectives, 29% use weighted sums, and only 16% explicitly consider QoS interference in decision making.Search-based optimization receives more attention than rule-based and control-theoretic approaches, which have similar popularity.Source: RQ5: What are the approaches used for… · pp. 18–19, Figure 7 · p. 19
  • 78% of studies claim to handle arbitrary objectives, but only 9% consider software–hardware interplay; 34% assume fixed bundles and 66% consider both scaling directions.CPU and memory are the most commonly tuned resources, and horizontal scaling is more popular than vertical scaling.Source: RQ5: What are the approaches used for… · p. 19

4.7.1 Rule Based Control.

  • Rule-based control maps manually specified conditions to scaling actions, with trade-offs often handled implicitly through those mappings.Some systems let users modify rules at runtime or define them at different cloud service levels.Source: §4.7.1 · p. 19

4.7.2 Control Theoretic Approach.

  • Control-theoretic approaches include standard controllers and extended designs that combine controllers with optimization or machine learning.The surveyed fuzzy-control examples explored CPU tuning alone and ignored QoS interference despite claiming broader hardware applicability.Source: §4.7.2 · pp. 19–20

4.7.3 Search-Based Optimization.

  • Search-based optimization includes implicit methods without QoS models and explicit methods that use models to guide search and reason about trade-offs.Explicit formulations include single-objective, weighted-sum, and Pareto-based optimization; implicit methods include reinforcement learning and demand prediction.Source: §4.7.3 · pp. 21–22
  • Single-objective search can optimize one objective while treating other requirements as constraints, whereas Pareto methods handle multiple objectives without assigning weights.Surveyed techniques include integer linear programming, weighted-sum search, metaheuristics, and Pareto-based algorithms.Source: §4.7.3 · pp. 22–23

4.8 Experimental Evaluation on SSCAS

  • Experiments use simulators, custom private clouds, or public clouds, with benchmarks ranging from simple web hosting to multi-tier software and synthetic or recorded workloads.Simulators support controlled experiments but may not fully capture realistic environments; the infrastructure and application scenario affect scaling implementation.Source: §4.8 · pp. 23–24, §4.9 · p. 24, Table 9 · p. 24

5 REFLECTIONS AND OPEN CHALLENGES FOR SSCAS RESEARCH

The survey finds recurring gaps in awareness, architecture, modeling, control granularity, and decision making. Its discussion highlights trade-offs and areas where current SSCAS studies remain limited.

5.1.1 The Levels of Self-Awareness and Self-Adaptivity in Cloud Autoscaling Systems. Stimulus-

  • All 109 studies considered stimulus-awareness, while time- and goal-awareness each appeared in 52%; interaction- and meta-awareness appeared in 13% and 7%.Self-healing appeared in 5% of studies and only one work targeted self-protection; 67% did not specify required knowledge at the architecture level.Source: §5.1.1 · p. 24

5.1.2 Architectural Patern.

  • Feedback architectures accounted for 82% of studies, compared with 15% for MAPE and 3% for ODA; feedback loops can leave required knowledge coarse-grained.MAPE supports separation of concepts and sequential interactions, while the surveyed patterns still lack fine-grained representation of required knowledge.Source: §5.1.2 · p. 25

5.1.3 QoS Modeling.

  • Analytical and machine-learning models appeared in 43% and 38% of studies, respectively, with different trade-offs in assumptions, runtime efficiency, adaptability, overhead, and accuracy.Simulation is static and assumption-dependent; hybrid models may combine strengths of different approaches.Source: §5.1.3 · p. 25
  • Only 13% of studies modeled QoS interference, and dynamic input-feature selection appeared in 7%, despite dynamic or semi-dynamic model structure in 61%.The survey reports that interference increases model dimensionality, while changing input features may matter most for high-dimensional models.Source: §5.1.3 · p. 25

5.1.4 Granularity of Control.

  • Service/application-level control was the most popular granularity at 45%, but finer control can trade global optimality for scalability.Only 8% of studies considered multiple levels, which may offer a trade-off but raises questions about which levels to select.Source: §5.1.4 · p. 26
  • Most studies controlled each application, while fixed control granularity can trade global optimality against scalability and overhead.The survey notes that multiple or dynamically adjusted granularities may better address dynamic and uncertain conditions.Source: §5.1.4 · p. 26

5.1.5 Decision Making.

  • 63% of studies did not explicitly address objective trade-offs, and 84% omitted QoS interference from decision making.The survey attributes the latter omission to increased dimensionality and complexity when interference is included.Source: §5.1.5 · p. 26
  • Although 78% of studies claimed to handle arbitrary objectives, only 9% considered software–hardware interplay and 34% used fixed decision bundles.Fixed bundles reduce the search space but may rule out good solutions; 66% considered both vertical and horizontal scaling, with horizontal scaling more popular.Source: §5.1.5 · pp. 26–27

5.2 Open Problems and Challenges for SSCAS research

The survey identifies unresolved design and evaluation challenges for SSCAS, spanning knowledge representation, architecture, QoS interference, decision spaces, and real-world validation.

5.2 Open Problems and Challenges for SSCAS research

  • Only 33% of studies discuss required architectural knowledge, leaving knowledge representation a central challenge for capturing more advanced awareness levels.The survey calls for systematic ways to distinguish knowledge levels and map them into SSCAS architectures.Source: §5.2 · p. 27
  • Multiple loops may reduce design coupling and support localized adaptation, but their number and placement across abstraction levels remain open questions.The survey suggests relating loop design to the knowledge levels required by the system.Source: §5.2 · p. 27
  • Fewer than 16% of studies account for QoS interference, whose inclusion raises dimensionality and complexity in both modeling and decision making.The survey calls for approaches that reduce dimensionality or mitigate its effects.Source: §5.2 · p. 27
  • SSCAS needs holistic approaches that account for interplay between software configurations and hardware resources, which can jointly affect service QoS.The survey notes that most existing work focuses on hardware resources at the IaaS level.Source: §5.2 · p. 27
  • Considering interference and software configurations can expand the QoS-model input space, motivating generic, efficient dynamic feature selection.The challenge is to assess feature combinations for model accuracy without unreasonable overhead.Source: §5.2 · pp. 27–28
  • Fixed resource bundles can constrain consumer-relevant decisions, while arbitrary configurations may create millions of alternatives for search.The survey calls for efficient methods to find optimal or near-optimal decisions in the enlarged decision space.Source: §5.2 · p. 28
  • Single-objective and weighted-sum formulations can limit trade-off exploration, leaving balanced decisions with good convergence and diversity as an open challenge.The survey points to stochastic optimization, particularly nature-inspired algorithms, as a potentially promising direction.Source: §5.2 · p. 28
  • Large-scale practical SSCAS cases are absent from many studies, although the survey describes them as necessary to verify systems’ potential, effectiveness, and impacts.Source: §5.2 · p. 28

5.3 Current Industrial Situation of SSCAS

Industrial autoscaling has relied on simple rule- and policy-based approaches, but increasing cloud complexity makes these approaches difficult to specify and adapt. Advanced SSCAS remains insufficiently mature for reliable adoption, though Aneka’s work has been incorporated into Microsoft Azure.

5.3 Current Industrial Situation of SSCAS

  • Rule- and policy-based autoscaling leaves consumers to specify rules, which can work initially but becomes difficult as cloud control primitives and decision choices grow.These approaches require application and domain knowledge to map conditions to actions, and cannot adapt to dynamically changing workloads or application states.Source: §5.3 · pp. 28–29
  • SSCAS is not yet mature enough for reliable adoption, although Aneka’s search-based optimization work has been incorporated into Microsoft Azure.The authors present Aneka as an example of research addressing industrial challenges and informing a commercial cloud offering.Source: §5.3 · p. 29

5.4 Discussion on the Pricing for Cloud Autoscaling

Advanced autoscaling can add computational costs, making pricing part of the discussion; the authors describe two ways to charge for SSCAS and note potential returns from more accurate autoscaling.

5.4 Discussion on the Pricing for Cloud Autoscaling

  • Advanced autoscaling may require additional computational resources, and the authors argue that it cannot be considered separately from pricing.They identify both cost and revenue as drivers of industrial demand for more advanced solutions.Source: §5.4 · p. 29
  • More accurate autoscaling may improve utilization and SLA guarantees, with the authors reporting that added resource costs are marginal compared with savings in cited work [122][72][48][49].The passage links improved utilization and SLA guarantees with provider reputation and consumer attraction.Source: §5.4 · p. 29
  • The authors describe charging SSCAS either as a separately metered optional service or by bundling its charge into the normal price per time unit.The examples are charging computation as an instance or adding an SSCAS fee to the hourly instance price.Source: §5.4 · p. 29

6 CONCLUSION

The survey and taxonomy synthesize SSCAS research findings and identify open challenges. The authors hope this synthesis will motivate further work on intelligent cloud autoscaling and its interactions with other cloud-computing problems.

6 CONCLUSION

  • Stimulus-, time-, and goal-awareness are the most widely considered knowledge levels, while self-configuring and self-optimizing are the most popular self-adaptivity notions in SSCAS.These are the survey’s reported patterns across self-awareness and self-adaptivity.Source: §6 · p. 29
  • Feedback loops, analytical and machine-learning models, service/application-level control, and search-based optimization are prominent SSCAS architecture, modeling, control, and decision approaches.The survey identifies feedback loops as the most commonly exploited architecture, service/application as the most applied control granularity, and search-based optimization as the most common decision approach.Source: §6 · p. 29
  • The authors identify explicit knowledge representation, multiple-loop architectures, QoS interference, and software–hardware interplay as open challenges for SSCAS.They also call for dynamic feature selection in QoS modeling.Source: §6 · p. 30
  • Further challenges include more flexible control granularity, relaxed resource-bundle assumptions, and explicit handling of trade-offs between conflicting objectives.The authors also identify a need for more real-world SSCAS cases and scenarios.Source: §6 · p. 30
  • The authors hope the survey and taxonomy will motivate research on more intelligent cloud autoscaling systems and their interactions with other cloud-computing problems.Source: §6 · p. 30
Loading 1609.03590v4…