Source-linked AI summary
Carbon-aware Resource Management for Latency-Sensitive Cloud Computing Environments: A Taxonomy and Future Directions
Tharindu B. Hewage, Shashikant Ilager, Maria Rodriguez Read, Rajkumar Buyya
TL;DR
Latency-sensitive cloud workloads constrain the trade-off between application performance and carbon efficiency, while their growth increases pressure on cloud emissions. This paper develops a taxonomy, reviews operational and embodied carbon-management research, and identifies gaps and future directions for resource management in these environments.
Problem
Latency-sensitive workloads have stringent performance requirements, making carbon reduction difficult when cloud infrastructure scales and deployment patterns are rigid.
Method
The paper proposes a taxonomy and uses it to conduct an in-depth review of carbon-aware resource management across operational and embodied carbon aspects.
Results
The review classifies existing approaches, analyzes their challenges and optimization aspects, and identifies open problems for future research.
Takeaways & Limitations
The taxonomy gives cloud operators a holistic view of carbon-optimization opportunities and gives researchers a basis for developing new techniques.
Abstract
from arXiv · showhide
Proliferation of cloud-based latency-sensitive workloads requires infrastructures tuned to their workload-specific latency constraints. Today, they shape the cloud from a generalized computing platform to diverse workload-specific cloud environments. As the demand for latency-sensitive workloads increases, cloud service providers continue to scale their infrastructure, adversely increasing the carbon footprint and challenging climate-crisis-driven net-zero emission goals. Due to performance-oriented rigid deployment patterns of latency-optimizations, reducing its carbon footprint is challenging. Therefore, efficient techniques that exploit application specific opportunities are needed in that. To this end, we present a detailed taxonomy of recent literature on carbon-aware resource management in latency-sensitive cloud computing environments. Using the taxonomy, we analyze existing works discussing their optimization aspects, identify the gaps, and highlight future research directions.
1 Introduction
Latency-sensitive cloud workloads limit the performance–carbon trade-offs available to traditional cloud optimization, while variable low-carbon energy and rigid resource allocation complicate carbon reduction. The article reviews existing techniques, classifies their design aspects, and identifies research gaps.
- Motivation: Latency-sensitive workloads cannot generally sacrifice performance for carbon efficiency, making their carbon optimization distinct from traditional hyperscale cloud optimization.The challenge arises because workload-specific latency SLOs constrain operational choices.
- Challenges: Variable availability of low-carbon energy complicates maintaining required application latency while maximizing renewable-energy integration.Renewable sources such as solar and wind can vary in availability across time and geography.
- Carbon dimensions: Carbon-aware management must address both operational carbon from energy and cooling and embodied carbon from manufacturing, supplying, recycling, and scaling IT assets.These two carbon dimensions arise from operating cloud infrastructure and deploying or maintaining its physical elements.
- Challenges: Hard latency bounds require rigid, high-performance resource allocations that restrict power management and workload redirection toward lower-carbon sites.Examples include isolated CPU cores and disabling CPU power-optimization features.
- Article contribution: The article reviews existing literature and proposes a taxonomy spanning system design, workload latency characteristics, management goals, and operational or embodied carbon.It uses the classification to analyze resource-management techniques and identify gaps and future research directions.
2 Related Surveys
Prior surveys examine carbon efficiency in cloud environments, often framing the area as green cloud computing centered on low-carbon renewable energy.
- Prior surveys: Existing surveys study cloud carbon efficiency across resource-management topics such as workload scheduling and power management.Preliminary surveys characterize carbon-efficient cloud computing as green cloud computing, emphasizing low-carbon renewable-energy integration.
3 The Taxonomy
The taxonomy divides carbon-aware resource management into operational and embodied carbon management, then organizes each by system, workload, and lifecycle considerations.
- Taxonomy structure: The taxonomy’s top-level split is operational carbon management versus embodied carbon management.Operational management is organized by workload management, control architecture, system tier, and latency tolerance; embodied management covers pre- and post-deployment stages.
3.1 Operational Carbon Management
Operational carbon management adapts workload execution, control, and infrastructure decisions to variable low-carbon energy while preserving latency constraints. The reviewed literature spans shifting, throttling, preemption, and centralized or decentralized coordination across cloud tiers.
- Energy integration: Operational carbon optimization primarily uses low-carbon energy sources, but renewable intermittency complicates resource management.Solar and wind availability can vary, requiring cloud operators to match workload execution with changing energy conditions.
- Workload management: Latency slack enables workload shifting: CDN capacity shifting with 60ms additional slack was associated with over 60% carbon-emission reduction.Other approaches exploit serverless functions outside critical paths or dynamically provision and load-balance services across regions.
- Workload management: Throttling and preemption adjust power or execution in place, using mechanisms such as DVFS, server duty cycles, power capping, and preemptive VMs.These methods trade performance, storage use, or fault tolerance against renewable-energy availability and latency requirements.
- Control architecture: Centralized control coordinates workloads, data centers, servers, and power delivery, whereas decentralized control improves autonomy and scalability but adds synchronization overhead.The literature rarely applies decentralized control because its overhead can affect latency, although blockchain-based geo-distributed management has been explored.
- System tiers: Operational carbon optimization can target multiple cloud tiers, with system-tier opportunities shaped by application latency SLOs.The taxonomy and reviewed work include server-level resource and energy-management decisions alongside broader workload and infrastructure coordination.
3.2 Embodied Carbon Management
The taxonomy classifies embodied carbon management across pre-deployment infrastructure design and low-carbon component use, and post-deployment asset management, while retaining application-latency performance as a design concern.
- Embodied Carbon Management: Embodied carbon management is divided into pre-deployment and post-deployment stages covering infrastructure planning, component use, asset management, and recycling.The review narrows these techniques to cases considering application latency performance.
- Pre-deployment: Pre-deployment management emphasizes infrastructure design and low-carbon components so latency-sensitive workloads can run efficiently without repeated infrastructure optimization.The taxonomy treats infrastructure design as the primary entry point for pre-deployment carbon efficiency.
- Pre-deployment: Renewable modular data centers can reduce embodied carbon and, when colocated with stable energy sources, reduce servers needed for peak energy spikes.The cited work uses geographically distributed modular data centers for low-latency workloads.
- Pre-deployment: CXL-based tiered memory lowers embodied carbon relative to local DRAM, while a software stack manages its higher latency to approach local-DRAM performance.The design combines hardware-managed tiering with software support.
- Post-deployment: Post-deployment techniques extend embodied-carbon amortization by lengthening asset life, reusing components, and balancing wear, but aged hardware can increase failure risk.Examples include aging-aware scheduling, SSD write redirection, and evaluation of used memory and SSDs.
4 Classification of Resource Management Techniques Using Taxonomy
The taxonomy organizes operational and embodied carbon-reduction techniques by management scope, workload control, system topology, latency tolerance, and the resource or asset being optimized.
- Taxonomy Classification: The classification tables separate operational carbon reduction from embodied carbon reduction and map works to the taxonomy’s resource-management aspects.Table 1 covers operational carbon management, while Table 2 covers embodied carbon management.
- Operational Carbon Reduction: Operational techniques span load shifting, throttling, preemption, and workload control across centralized or decentralized inter-cloud, datacenter, and server-level systems.The classification also records workload granularity such as requests, functions, containers, tasks, and VMs.
- Taxonomy Dimensions: The taxonomy records whether techniques target components, servers, or broader cloud environments and whether management is centralized or decentralized.These categories expose the operational scope and control structure of the reviewed approaches.
- Operational Carbon Reduction: Operational methods are distinguished by latency tolerance, including bounded, mixed, and other workload-specific constraints, and by whether they optimize servers, datacenters, networks, or accelerators.These dimensions capture how carbon management is coupled to latency-sensitive execution.
- Embodied Carbon Reduction: Embodied-carbon techniques cover pre-deployment infrastructure choices and post-deployment strategies such as extended component life, aging-aware scheduling, CXL memory, and SSD reuse.The classification includes CPU wear-off, server aging, memory selection, and SSD aging or repurposing.
5 Research Gaps and Future Directions
The review identifies future opportunities for carbon-aware management that preserve latency guarantees while exploiting renewable-energy variability, application-specific tolerance, finer-grained control, and longer-lived hardware.
- Cloud Environment Characteristics: Server- or core-level renewable-energy allocation could support strict latency SLOs that prevent workload migration, but uneven workload placement may underuse available renewable energy.The review identifies this tension as an open resource-management problem.
- Application Characteristics: Future work should examine embodied-carbon effects in generative-AI inference clusters, including latency-aware scheduling and underutilized CPUs that could be used to reduce component stress.Accelerator-heavy clusters can leave other server resources underutilized.
- Cloud Environment Characteristics: Application-specific middleware and fault tolerance could help latency-sensitive workloads use infrastructure that is evictable or exposed to renewable-energy intermittency.The opportunity depends on workloads that can tolerate intermittent VM failures.
- Application Characteristics: Bounded-latency applications remain an important opportunity for absorbing variable renewable energy, despite their strict performance requirements and high-performance server profiles.These applications contribute substantially to operational carbon concerns and can stress server components over time.
- Resource Management Approaches: Decentralized control may address centralized-control scaling bottlenecks in distributed clouds, while mixed latency SLOs and serverless execution offer finer-grained management opportunities.The review also proposes studying thermal-load scheduling to improve component longevity in mixed-SLO workloads.
6 Summary
The article reviews carbon-aware resource management for latency-sensitive cloud environments and proposes a taxonomy spanning operational and embodied carbon efficiency. It uses that taxonomy to analyze existing work and identify challenges and future research opportunities.
- Summary: The article presents a detailed review focused on carbon-aware resource management in latency-sensitive cloud environments.
- Summary: Its taxonomy provides a holistic view of operational and embodied carbon management for identifying short- and long-term optimization opportunities.The framework is intended for cloud operators managing both immediate and lifecycle carbon goals.
- Summary: The review analyzes existing works with the taxonomy and provides a gap analysis emphasizing substantial potential for future research.The gap analysis focuses on identified challenges and directions for building on prior work.