A Manifesto for Future Generation Cloud Computing: Research Directions for the Next Decade

1 Introduction

• Cloud computing provides on-demand, subscription-based IT resources that have accelerated business formation, global enterprise expansion, scientific progress, and pervasive application models.Its adoption has made Cloud computing a backbone of the modern economy and an emerging fifth utility.Source: §1 · p. 2
• SaaS, PaaS, and IaaS support different abstraction and control levels, enabling flexible application access, application development, and pay-as-you-go infrastructure use.SaaS reduces licensing and maintenance costs, PaaS provides development and auto-scaling frameworks, and IaaS offers leased virtual machines and storage with flexible resource control.Source: §1 · p. 2

1.1 Motivation and Goals of the Manifesto

• Cloud technologies and services are expanding through containers, Container as a Service, specialised analytics, and serverless architectures.These developments improve physical-resource utilisation, hide hardware complexity, and help industries configure Cloud capabilities for specific business requirements.Source: §1.1 · p. 2
• Geographically distributed Clouds increasingly integrate edge resources and multiple data centres to support IoT and Fog computing while reducing latency, improving reliability, and enhancing sustainability.Fog computing moves decision making closer to data sources through resources such as base stations, gateways, switches, and routers.Source: §1.1 · p. 2
• In 2016, 70% of organisations had at least one application in the Cloud, while public Cloud services were projected to grow from $70 billion in 2015 to more than $203 billion in 2020.This expansion creates new challenges and requires re-evaluating approaches to scalability, resource management, reliability, and security.Source: §1.1 · p. 3
• The manifesto identifies open challenges, examines existing solutions and limitations, analyses emerging trends and impact areas, and proposes research directions for the next decade.Its intended impact spans the next two or three generations of utility-oriented Cloud technologies, infrastructures, and application services.Source: §1.1 · pp. 3–4

2.1 Scalability and Elasticity

• Cloud scalability and elasticity improve application performance and cost-effectiveness, but further progress requires advances in hardware, middleware, algorithms, demand prediction, and dynamic resource management.Scaling is constrained by compute, memory, and interconnect technologies, while elasticity depends on accurate workload models and applications that can dynamically adjust resources.Source: §2.1 · pp. 4, 6

2.2 Resource Management and Scheduling

• Resource management and scheduling must become more robust to inaccurate workload information while coordinating control loops, fairness, cross-layer objectives, hybrid and federated Clouds, and security.Existing policies include provisioning, migration, auto-scaling, throttling, admission control, orchestration, workflow scheduling, and multi-Cloud load balancing, but demand prediction remains brittle and ML/AI’s sufficiency is unresolved.Source: §2.2 · p. 6

2.3 Reliability

• Cloud reliability remains difficult because massive-scale, service-oriented, heterogeneous systems experience independent and correlated failures that can cascade into major service disruptions.Existing environments lack thorough service-reliability models, automatic reliability-aware management, and failure-aware provisioning policies, while reliability and energy efficiency also interact.Source: §2.3 · pp. 6–7

2.4 Sustainability

• Cloud sustainability requires jointly managing energy consumption and QoS across servers, cooling, networks, and workloads rather than focusing only on VM consolidation.Proposed approaches include ML-based task allocation, holistic SLA and energy optimisation, deep-learning scheduling, energy rationing, variable clock rates, and renewable-energy mixing.Source: §2.4 · pp. 7–8

2.5 Heterogeneity

• Cloud heterogeneity spans VM configurations, providers, and hardware architectures, yet resource and workload management remain fragmented and applications struggle to balance accelerator performance against programming ease.A general-purpose platform must integrate all three heterogeneity levels, while hardware-agnostic software that adapts to available resources remains an open challenge.Source: §2.5 · p. 8

2.6 Interconnected Clouds

• Interconnected Clouds remain limited by nonstandard interfaces, APIs, formats, authentication, and contextualisation, making portability, cross-site networking, and richer multi-provider applications difficult.Existing middleware often operates at the lowest common denominator, while open challenges include coordinating authorisation, access, billing, and Fog-computing integration.Source: §2.6 · p. 9

2.7 Empowering Resource-Constrained Devices

• Mobile Cloud computing addresses constrained devices through task delegation and code offloading, but offloading still requires adaptive decisions based on device context, network conditions, and Cloud load.Cloud-centric IoT also faces sub-second latency and energy challenges, motivating Fog computing to bring computation and supervision closer to devices.Source: §2.7 · p. 10

2.8 Security and Privacy

• Security and privacy research must preserve confidentiality when encrypted data are queried or mined across domains, while exploiting Cloud data and ML for threat analysis without exposing sensitive information.Indexes and homomorphic encryption support computation over protected data but require privacy–precision trade-offs, and cross-domain knowledge can reveal more personal information than anticipated.Source: §2.8 · pp. 10–11
• Advanced Persistent Threats threaten Cloud confidentiality, integrity, and availability, requiring coordinated technical and policy-driven defences rather than technical protection alone.Encryption, trusted environments, and one-time passwords are cited as complementary mitigation examples.Source: §2.8 · p. 12

2.9 Economics of Cloud Computing

• Cloud economics must address pricing, brokerage, SLA monitoring, migration risks, restrictive licensing, and difficult provider comparisons caused by inconsistent resource terminology.Migration can alter organisational skills and roles, while licensing models are shifting toward BYOL, credits, tokens, elastic units, and on-demand use.Source: §2.9 · pp. 12–13

2.10 Application Development and Delivery

• Cloud delivery benefits from programmable infrastructure, DevOps, and orchestration, but heterogeneous Edge–Cloud applications remain insufficiently automated and application evolution lacks holistic engineering methods.Continuous delivery accelerates releases, whereas research is needed to combine it with careful design, testing, and systematic use of performance and reliability data.Source: §2.10 · p. 13

2.11 Data Management

• Data management still lacks effective metadata services, regulatory and privacy safeguards, low-latency Edge processing, network-aware placement, unified batch-stream processing, and elastic scaling.Serverless functions increasingly off-load resource allocation and scaling to Cloud providers while relying on external state services.Source: §2.11 · p. 14

2.12 Networking

• Cloud networks must become more energy proportional and provide scalable, secure performance guarantees despite multi-tenant limits and unpredictable geographically distributed public-network conditions.Current SLA coverage largely excludes network delay and bandwidth guarantees, while VLAN scale is theoretically limited to 4,096 segments.Source: §2.12 · p. 15

2.13 Usability

• Cloud usability requires closer alignment between human-computer interaction and distributed-systems research, with capable, personal, reliable, secure, and valuable services supported by APIs and DevOps.DevOps improves productivity and can simplify migration of legacy software to multi-tenant elastic Cloud services.Source: §2.13 · pp. 15–16

3.1 Containers

• Containers provide lightweight, isolated application environments, start in less than a second, use few resources, and support more host instances than VMs.Their weaker kernel-sharing isolation drives research on Intel SGX and Unikernel, while Kubernetes, Mesos, and Swarm address cluster management.Source: §3.1 · pp. 16–17

3.2 Fog Computing

• Fog computing places compute near data sources to reduce latency and improve location-aware, mobile, wireless, real-time, and IoT services.Its major challenges are multi-party SLA management across independently owned nodes and enhanced security and privacy for interactions between users and Cloud data centres.Source: §3.2 · pp. 17–18

3.3 Big Data

• Big Data is shifting attention from data volume toward streaming-data velocity, while edge analytics address latency and bandwidth constraints from rapidly growing IoT data.Cloud Data Lakes also require provenance metadata and linked-data models so their contents do not become unusable “data graveyards”.Source: §3.3 · p. 18

3.4 Serverless Computing

• Serverless computing abstracts server management through BaaS and FaaS, reducing backend code and administration while handling provisioning, workload balancing, reliability, and fault tolerance.Current FaaS models restrict execution time, statelessness, and deployment, requiring event-based and timeout-aware application designs [19].Source: §3.4 · p. 19

3.5 Software-defined Cloud Computing

• Software-defined Cloud Computing abstracts compute, storage, and network resources, with SDN separating programmable control from forwarding and NFV replacing proprietary appliances with software functions.SDN and NFV support agile services, network-energy proportionality, virtual-function consolidation, interconnected Clouds, and DDoS mitigation.Source: §3.5 · pp. 19–20

3.6 Blockchain

• Blockchain offers Cloud-hosted, cryptographically secured and auditable transaction services, but growing chains, smart contracts, analytics, and data marketplaces require scalable infrastructure and trustworthy data-quality mechanisms.Cloud can expose blockchain through PaaS and SaaS, while blockchain can support secure, verifiable data trading across industries.Source: §3.6 · pp. 20–21

3.7 Machine and Deep Learning

• Machine and deep learning can optimize Cloud resource management and provide scalable hosted AI services, enabling demand prediction, SLA estimation, placement, auto-tuning, and adaptive security.A key research challenge is creating reusable models across users and contexts despite application-specific differences.Source: §3.7 · p. 21

4 Future Research Directions

• Next-decade Cloud services must manage massive heterogeneous IoT and sensor data while adapting across hardware and optimising QoS, energy, and cost.These services must satisfy application-level quality-of-service constraints and economic restrictions.Source: §4 · p. 22
• Future Cloud research should address existing limitations while responding to opportunities and problems created by emerging trends.The manifesto presents these as key research directions for the coming decade.Source: §4 · p. 22

4.1 Scalability and Elasticity

• Scalability and elasticity challenges should be decomposed across hardware, middleware, and application levels.Source: §4.1 · p. 22
• Special-purpose Clouds should support deep learning, stream analytics, and image or video recognition through new hardware, virtualization, programming, and economic abstractions.Relevant directions include accelerators and non-traditional architectures such as neuromorphic and quantum computing, plus just-in-time compilation and composable FaaS micro-services.Source: §4.1 · p. 22
• Middleware research should improve infrastructure reuse, large-scale provisioning, geographic failover, InterCloud execution, and consistent HPC performance despite concurrent Cloud users.Scheduling should avoid execution noise by matching HPC workloads appropriately or using dedicated clusters.Source: §4.1 · p. 22
• Communication-intensive applications require high-throughput, low-latency networks and sustainable multi-client provisioning for near-real-time Big Data, HPC, and IoT control loops.These capabilities support applications such as smart cities [184] and industrial IoT.Source: §4.1 · p. 22

• Fog and Edge computing require cooperative resource management across centralised CDCs and distributed devices, accounting for location, mobility, topology dynamics, privacy, security, and multiple control loops.Container technology may help because of its small footprint and fast deployment.Source: §4.2 · p. 23
• Scheduling research must generalise multi-resource fairness and game-theoretic methods to large heterogeneous CDC–Edge environments, where mean-field games may address scalability.These methods currently target smaller clusters or Cloud economic models, while mean-field games reason about many resources, devices, and user types.Source: §4.2 · p. 23
• Serverless management must balance erratic execution time, network latency, per-access cost, QoS, privacy, security, provider revenue, and proactive provisioning across FaaS and local microservices.Providers need allocation that is neither excessive nor insufficient, while soft execution-time constraints and startup latency require proactive FaaS provisioning and potentially game-theoretic market models.Source: §4.2 · p. 23
• SDN and NFV enable joint VM–traffic consolidation and on-demand service-chain placement, but require new heuristics and algorithms for optimal orchestration, consolidation, and dynamic deployment.These mechanisms can improve energy consumption and SLA satisfaction while placing virtual network functions such as NAT, firewalls, and IDS.Source: §4.2 · p. 23
• AI, deep learning, and ML may improve VM provisioning, consolidation, and load balancing at scale, but potential loss of control and determinism makes explainable AI important for adoption [68].These techniques are motivated by growing complexity, heterogeneity, and scale in Cloud management.Source: §4.2 · p. 24
• Extreme-scale scientific workflows require fault tolerance, performance modelling, data management, efficient resource use, anomaly detection, and failure prediction, yet serverless HPC applications remain underexplored.Big Data analytics can support monitoring and workflow management, while serverless functions may enable reusable, shareable complex workflows.Source: §4.2 · p. 24

4.3 Reliability

• Reliability research should develop hierarchical, service-oriented Reliability as a Service models with failure and workload modules supporting resilient performance and flexible provisioning.These models must address independent, temporal, and spatially correlated failures while meeting users’ performance, reliability, and QoS expectations.Source: §4.3 · p. 24
• Deep and machine learning could predict failures from massive datasets, enabling failure-aware provisioning and more intelligent, power-efficient fault tolerance for Fog and IoT applications.Replication may help at the edge, but reliability must be balanced against system power efficiency.Source: §4.3 · p. 24
• Cloud storage research must optimize replication and erasure coding across durability, availability, storage overhead, bandwidth, energy, and recovery performance.A proposed hybrid approach dynamically replicates erasure-coded data using node-failure prediction, reducing network traffic and storage overhead while improving Big Data performance.Source: §4.3 · p. 24

4.4 Sustainability

• Sustainability research primarily targets ICT energy consumption, while pollution and equipment decontamination are acknowledged but excluded from this paper.The scope therefore centers on energy awareness and energy-efficient Cloud operation.Source: §4.4 · p. 25
• Geographically distributed data centres can exploit renewable energy and natural cooling through energy- and carbon-aware provisioning, load balancing, and pricing that encourages providers to follow renewables.Dynamic coupling of energy flows with computing and communication resources can satisfy QoS and SLAs while minimizing consumption, but substantially more work is needed.Source: §4.4 · p. 25
• Locating data centres remotely may increase network energy use and violate latency, locality, legal, security, or intermittency constraints, making energy-efficient Cloud operation a major priority.Financial transactions require very short delays, some data must remain within national or EU boundaries, and renewable energy may be unavailable locally.Source: §4.4 · p. 25
• Because performance and data processing increase energy use, adaptive self-aware machine-learning techniques must manage QoS, SLAs, and energy consumption online.Continuous system measurement can support online decisions that optimize sustainability while respecting QoS and SLA requirements.Source: §4.4 · p. 25
• Fog creates a trade-off between higher energy use in diverse distributed systems and network savings from proximity, while VM consolidation, energy-aware routing, and SDN offer sustainability opportunities [89, 91, 92, 24, 94, 80, 96, 93, 79].Further work should evaluate Fog’s network-QoS benefits and develop sustainable networks without sacrificing reliability, scalability, or performance.Source: §4.4 · pp. 25–26

Unified management

• The manifesto proposes an umbrella platform integrating workload and resource management across VM, vendor, and hardware-architecture heterogeneity, including heterogeneity-aware memory management.Optimal strategies would be selected according to application requirements; hypervisor-centered memory management currently limits heterogeneity benefits, while guest operating systems may become heterogeneity-aware.Source: §4.5 · p. 26

Programming abstractions

• An open-source high-level programming language should provide abstraction and elasticity for heterogeneous Cloud applications, including those spanning Cloud and Fog resources.Current accelerator programming requires device-specific languages and low-level effort, restricting adoption beyond scientific applications.Source: §4.5 · p. 26
• Automated fine-grained task partitioning and coarse-to-fine conversion are needed to overcome OpenCL’s explicit CPU–accelerator programming and limited performance portability.TANGRAM provides portability across accelerators but requires performance models and adaptive runtimes to optimize CPU–accelerator interaction.Source: §4.5 · p. 26

Historical foundations

• Earlier virtualization and memory technologies, including Multics and IBM CP-67, may provide inspiration for addressing modern Cloud heterogeneity challenges.VMware later applied related virtualization technology to x86 servers, while earlier systems also supported I/O virtualization.Source: §4.5 · pp. 26–27

Emerging infrastructure

• Disaggregated data centres separate CPU, memory, and storage into networked resource blades, potentially reshaping traditional IaaS provision.Most prototypes remain proprietary and early-stage, and deployment requires substantial network-fabric development.Source: §4.5 · p. 27

• Cloud portability and interoperability remain difficult because pure standardisation is insufficient, despite containers, configuration tools, adapters, and libraries being widely used [36].These longstanding challenges remain insufficiently addressed because of their complexity.Source: §4.6 · p. 27
• InterCloud middleware should compose simple services into complex functionality without requiring a common minimum feature set, while letting users trade off cost, response time, and data sovereignty.The resulting service could be unique to one provider or assembled from multiple providers according to user constraints.Source: §4.6 · p. 27
• User-level middleware must identify candidate services and compose them without relying on Cloud-provider cooperation, given providers’ incentives to retain functionality within their own services.This requirement applies to InterCloud and hybrid Cloud environments.Source: §4.6 · p. 27
• Such middleware must hide delivery details from users and independently interpret provider APIs, determine required inputs, and verify outputs for cross-provider composition.It therefore needs capable interfaces to both users and providers, underscoring the complexity of the proposed vision.Source: §4.6 · p. 27
• Ubiquitous Cloud federation depends on vendors perceiving financial benefits, requiring novel billing, accounting, pricing, and InterCloud marketplaces.Vendor incentives are presented as a condition for widespread interconnected-Cloud adoption.Source: §4.6 · p. 27
• SDN and NFV may improve composed services through traffic prioritisation, security enforcement, and quality-of-service support across providers and data centres.SDN has already been used to optimise wide-area traffic connecting Google data centres.Source: §4.6 · p. 27

• Mobile Cloud offloading needs better multi-tenancy models to share rising costs among heterogeneous mobile users, devices, and Cloud resources.Increasing device capabilities limit suitable offloading applications while higher-capacity Cloud instances increase per-offloading costs.Source: §4.7 · p. 28
• Emerging serverless computing, CaaS, and Fog capabilities remain insufficiently explored for mobile Cloud applications and social-networking scenarios.Research should design ideal scenarios while examining how these technologies can extend mobile Cloud patterns.Source: §4.7 · p. 28
• Incentive mechanisms should encourage heterogeneous mobile Cloud offloading and participation in IoT, Fog, P2P, and consumer-as-provider models.Fog infrastructures offered by independent providers and the MQL5 Cloud Network illustrate architectures requiring appropriate rewards.Source: §4.7 · p. 28
• Container technology can provide lightweight, efficient runtimes for Edge servers, including customised Cloud Android Containers supporting offloading and applications in IoT and AI.Containers are presented as convenient application-packing tools expected to become widely deployed with Fog and Edge computing.Source: §4.7 · p. 28
• Edge analytics should process sensor data across edge devices, intermediate Fog nodes, and Cloud while optimizing QoS, QoE, placement, scheduling, and deployment cost.Frameworks such as Apache Edgent and existing scheduling and placement work motivate further research on multi-layer Fog deployment for multi-stage analytics and dataflows.Source: §4.7 · p. 28

Security and Privacy

• Security and privacy solutions must be tailored to scenario-specific trust assumptions across networks, databases, virtualization, scheduling, and sensitive-data management.Cloud providers may be treated as completely untrusted or trustworthy, requiring context-specific designs.Source: §4.8 · p. 28

Data Protection and Sharing

• Research must develop scalable, efficient protection for encrypted and distributed data, supporting rich queries, selective multi-user sharing, integrity guarantees, query privacy, and computation integrity.Solutions should avoid information leakage, support concurrent access and multiple providers, and let users evaluate integrity guarantees against cost and supported computations.Source: §4.8 · pp. 28–29
• Provider-selection frameworks should let users express diverse requirements and preferences while verifying that Cloud services comply with signed contracts.Existing approaches often consider only limited requirements such as cost and performance or rely on predefined indicators.Source: §4.8 · p. 29

Emerging Environments

• Fog Computing and Big Data increase privacy, trust, inference, provenance, and protection challenges through decentralized nodes, fragmented information, data variety, and greater risk of sensitive-data inference.Blockchain can provide immutable, verifiable, and traceable provenance, but blockchain data cannot be changed or deleted, creating additional privacy concerns.Source: §4.8 · p. 29

Infrastructure Security

• Infrastructure research must secure multi-tenant virtualization, resource placement, virtual-machine allocation, and legitimate-request identification against DoS and other cyber-attacks.Cloud systems should distinguish coordinated attacks from legitimate traffic, dropping excess load or avoiding wasteful provisioning and energy use.Source: §4.8 · pp. 29–30

Serverless computing

• Serverless economics depends on execution behavior and workload type, with costs shaped by transaction rates, concurrency, backend performance, and execution time.Container deployments’ lower overheads support real-time workloads, while serverless functions charge only when called and typically process data streams.Source: §4.9 · p. 30

Fog and Edge economics

• Fog and Edge resources could reduce Cloud operating costs, but Micro Data Centre business models must address incentives, contracts, federation, discovery, security, and profitable Cloud outsourcing.Proposed models include dynamic provider selection, pre-agreed performance-based contracts, federated workload sharing, and Cloud-to-MDC exchange when QoS targets cannot be met.Source: §4.9 · p. 30

Uncertainty-aware architecture and markets

• Cloud architectures must balance dependable, efficient provision with economic value under uncertainty through abstractions and competitive or cooperative strategies involving tenants and providers.Future market models should seek continuous “win-win” outcomes for providers, users, and intermediaries while preserving qualities such as performance, availability, reliability, security, and compliance.Source: §4.9 · pp. 30–31

Cloud migration economics

• Migrating IT systems and departments to the Cloud raises unresolved questions about risk tolerance, business continuity, trust, regulation, data placement, and revenue models across storage and processing actors.Research must determine which data stays near users and which moves to data centres, considering cybersecurity expertise, access rights, GDPR, the US CLOUD Act, and Edge analytics.Source: §4.9 · p. 31

Cloud marketplace

• The Cloud marketplace is expanding beyond dominant providers, with Edge computing creating opportunities for telco alliances that support real-time stream processing and Edge analytics.Cloud Harmony reported over 90 providers, while autonomous vehicles and smart-city sensing are increasing data volumes and demand for mobile infrastructure partnerships.Source: §4.9 · p. 31

Application development and delivery

• Continuous delivery needs methods, tools, and organisational processes to monitor and evolve application design and quality while supporting IoT, Fog/Edge, Big Data, and serverless technologies.Current agile and continuous delivery practices can reduce design-time reasoning about SLA compliance, business alignment, and value-driven design, risking unsuitable early architectures.Source: §4.10 · p. 31
• Infrastructure-as-code adoption requires software engineering methods and quality tools for writing, debugging, and evolving code across multiple programming and scripting languages.Continuous delivery tools such as Chef, Ansible, and Puppet are expected to increasingly automate infrastructure creation, configuration, and management.Source: §4.10 · p. 31
• Cloud-native architectures and patterns should support human-oriented application definition, runtime context changes, SLA adaptation, and decomposition across microservices, containers, serverless, and FaaS.Decomposition choices affect application security, performance, reliability, and operational costs, while emerging architectures simplify automatic scaling, parallelism, and distribution.Source: §4.10 · p. 31
• Integration and control patterns must combine external and serverless services, bridge Edge cyber-physical systems with Cloudlets and swarmlets, and enable runtime switching through model-driven meta-controls.These meta-controls can support reasoning, what-if analysis, monitoring-feedback analysis, and correct enactment of adaptation decisions.Source: §4.10 · p. 32
• Architectures combining multiple paradigms need workload formalisms that propagate performance, reliability, and security requirements while characterising processed-data properties, and orchestration must assess resource and quality trade-offs.Trade-offs span on-demand, reserved, spot, and burstable resources and the security, privacy, performance, and cost effects of FaaS-based designs.Source: §4.10 · p. 32
• Future-generation Clouds require translucent programming models and APIs for modular, probabilistic, non-deterministic, adaptive, and decentralised applications, including policy control and transparent migration between edge resources and CDCs.Functional concerns can be specified, tested, and verified modularly, while non-functional concerns such as fault tolerance can be delegated to middleware or controlled declaratively.Source: §4.10 · p. 32

Metadata and provenance

• Cloud data platforms need first-class metadata and provenance to support data discovery, quality, auditing, and governance, beyond basic storage versioning.Data lakes commonly lack source and provenance tracking, while scientific repositories demonstrate the value of contextual metadata and processing histories.Source: §4.11 · p. 32

Edge and distributed processing

• Fog-layer data services and joint optimisation of placement, routing, elasticity, and bandwidth are needed to reduce latency for IoT and 5G workloads.Existing CDNs mainly serve slow-changing data and do not support writing data closer to the edge; stream-processing placement further complicates latency guarantees.Source: §4.11 · pp. 32–33

Security and privacy

• Hybrid cloud architectures can protect proprietary data under enterprise supervision while retaining public-cloud benefits, but provider-access auditing remains unresolved.Private data centres near public cloud data centres can use dedicated high-bandwidth networks to manage sensitive data.Source: §4.11 · p. 33

Edge and distributed processing

• Distributed cloud data management requires high-level abstractions and platform bindings that place processing near data while minimising network usage, latency, and elasticity costs.Current solutions are often limited to a single cluster or data centre, motivating deployment and management across Internet-edge resources.Source: §4.11 · p. 33

Emerging workloads

• IoT, deep learning, and blockchain require specialised services for streaming and trusted edge data, rapid model switching, compact training datasets, and scalable provenance tracking.Deep-learning workloads especially stress storage and access for video and image datasets, while blockchain can provide an implicit audit trail.Source: §4.11 · p. 33

• SDN enables more efficient traffic engineering through global network visibility, dynamic flow scheduling, and management based on utilisation and flow sizes, with future methods targeting delay, bandwidth, jitter, and Cloud SLAs.Novel SDN traffic-engineering methods must exploit global network and flow characteristics to meet application QoS requirements and enforce committed SLAs.Source: §4.12 · pp. 33–34
• SDN can introduce Cloud security services for policy enforcement, attack detection, monitoring, and anomaly detection, but its controllers, control-plane links, and forwarding devices create new or amplified threats requiring research.Potential targets include the SDN controller and inter-SDN communications, while compromised forwarding devices may provide adversaries with greater capabilities.Source: §4.12 · p. 34
• Programmable SDN is expected to simplify intra-data-centre networking and reduce provider capital and operational expenditure, while interconnected-Cloud networking and SDN over public channels remain insufficiently investigated.The Cloud community has prioritised intra-data-centre networking, but efficient solutions for interconnected environments are also highly demanded.Source: §4.12 · p. 34
• NFV research must improve service-chain orchestration through QoS-aware VNF placement, consolidation, migration, and adaptive auto-scaling that responds to performance variation and satisfies application SLAs.Scaling should coordinate multiple VNFs across service chains rather than treating each network service independently, while traffic engineering can reduce network communication cost.Source: §4.12 · p. 34
• IoT and 5G will sharply increase network capacity demands, requiring expanded bandwidth, faster broadband, optical networking, and stronger integration with Edge and Fog computing.Edge and Fog computing alone cannot satisfy all networking demand generated by billions of devices and expanding mobile connectivity.Source: §4.12 · p. 34
• Edge and Fog networking still requires research on topology-aware application placement, dynamic fault detection, and network slicing management beyond existing 5G resource-management and discovery studies.Current 5G integration discussions in Edge and Fog computing remain narrow in scope.Source: §4.12 · p. 34

4.13 Usability

• Usability improvements should provide spending estimates, layered Cloud visualizations, easier API management, and better support for data-intensive applications’ expected QoS.These measures target workload- and resource-related cost uncertainty, infrastructure and application understanding, varied expertise, and user experience.Source: §4.13 · p. 35
• Advisory systems should reduce users’ overload from diverse Cloud resources and services by recommending efficient allocation, transfer, visualization, and infrastructure choices.Examples include deciding whether to transfer or visualize data remotely, allocate or delete resources, or replace virtual machines with baremetal machines.Source: §4.13 · p. 35
• Evaluation is difficult because Cloud users differ in profiles, needs, skills, and layers, requiring scalable qualitative studies alongside traditional quantitative experiments.Relevant studies must investigate how heterogeneous users interact with IaaS, PaaS, and SaaS technologies.Source: §4.13 · p. 35
• Cloud technologies should include mechanisms that facilitate or automatically reconfigure systems to match individual user needs and preferences rather than assuming uniform skills or requirements.This direction follows from the heterogeneity of Cloud users and their differing needs across Cloud layers.Source: §4.13 · p. 35

IaaS

• IaaS research should support heterogeneous and non-traditional hardware, disaggregated data centres, sustainable infrastructure, and improved scalability, resource management, scheduling, networking, and interconnected Clouds.Proposed directions include CPUs, GPUs, TPUs, neuromorphic and quantum architectures, renewable energy, IoT-enabled cooling, and stand-alone resource blades for QoS-aware provisioning.Source: §4.14 · p. 35

PaaS

• PaaS should advance scalable elastic programming abstractions, heterogeneous-resource management, energy efficiency, portability, reliability, performance, serverless QoS and economics, data analytics, security, and edge analytics for IoT and smart cities.Machine learning, deep learning, and AI applications are expected to help address complexity, heterogeneity, scale, and load balancing in PaaS applications.Source: §4.14 · pp. 35, 37

SaaS

• SaaS research should improve application development, delivery, usability, monitoring, evolution, standards, DevOps, and abstractions supporting IoT, Edge computing, Big Data, and serverless applications.Recommended directions also include Cloud design patterns, workload and workflow formalisms, and incentive mechanisms addressing mobile Cloud adaptability and Fog architectures.Source: §4.14 · p. 37

Cross-cutting directions

• Future Cloud architectures and market models should embrace uncertainty while producing continuous win-win outcomes for providers, users, and intermediaries through ROI and SLA satisfaction.The discussion frames these outcomes as a target for all participants in future-generation Cloud ecosystems.Source: §4.14 · p. 37

Background

• Cloud computing has become a backbone of the modern economy through anytime, anywhere subscription services using pay-as-you-go delivery.The paradigm has supported faster business creation, global enterprise expansion, scientific progress, and new pervasive-computing models.Source: §5 · p. 37

Challenges

• The manifesto identifies major unresolved Cloud computing challenges, including heterogeneous IoT data streams and energy-, cost-, device-, QoS-, and security-aware services.These requirements demand new approaches and re-evaluation of existing models for scalability, resource provisioning, and security.Source: §5 · p. 37

Synthesis

• It consolidates the state of the art, limitations, emerging trends, and impact areas that shape future Cloud computing challenges.The manifesto brings these developments together to frame the field’s open issues.Source: §5 · p. 37

Research directions

• It offers comprehensive research directions for the next decade and calls the community to address the identified open issues.The directions are presented as promising for Cloud computing’s technical and economic future.Source: §5 · p. 37