arXiv CS open access

Carbon-aware Resource Management for Latency-Sensitive Cloud Computing Environments: A Taxonomy and Future Directions

2026 · arxiv_cs

arXiv CS · Papers · 2026

cloud distributed-computing parallel-computing

Conceptio Open Archive

Analyze & explore related research in Conceptio

Connect this record directly into your research workflow or AI agent. Explore citations across open repositories, query full text, and export structured proof bundles.

Explore in Search →Connect via MCP / API ↗

Carbon-aware Resource Management for Latency-Sensitive Cloud Computing Environments: A Taxonomy and Future Directions

arXiv:2609.03270v1 [cs.DC] 3 Sep 2026

THARINDU B. HEWAGE, University of Melbourne, Australia SHASHIKANT ILAGER, University of Amsterdam, Netherlands MARIA RODRIGUEZ READ, University of Melbourne, Australia RAJKUMAR BUYYA, University of Melbourne, Australia Proliferation of cloud-based latency-sensitive workloads requires infrastructures tuned to their workloadspecific 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. CCS Concepts: • General and reference → Surveys and overviews; • Computer systems organization → Distributed architectures. Additional Key Words and Phrases: cloud computing, sustainability, de-carbonization, latency-sensitive applications

Introduction

The primary difference between carbon-aware resource management for latency-sensitive workloads and carbon optimization in traditional hyper-scale cloud environments is the lack of flexibility in application workloads to compromise performance over carbon efficiency. Cloud carbon optimization involves managing cloud environments, which are complex cyber-physical systems of carbon-intensive stable energy sources, carbon-efficient variable availability energy sources, IT assets that guarantee reliable and high performance through carbon-intensive frequent component upgrades, and aging and recycled IT assets with lower carbon intensity with degraded performance and reliability. Carbon-aware resource management refers to the overall aspect of managing the resource requirements of the application workload while optimizing for carbon efficiency, emphasizing harmony between application workload performance and intricacies of cloud environment elements. With the stringent service level objective (SLOs) of latency-sensitive applications, carbon-aware resource management has fewer opportunities to compromise application workload performance, requiring the exploration of complex application patterns intertwined with deployment architectures to uncover alternative means of carbon optimization while meeting application SLOs. We identify two significant aspects in that, which must be dealt with in a manner suitable for the intricacies of latency-sensitive cloud environments. • Operational Carbon Management: As cloud environments operate, their IT assets consume electricity and produce heat as waste, which requires cooling systems to counterbalance the thermal energy. Energy consumption of IT assets and cooling systems draws power from various energy sources, where the carbon intensity of the source can vary. The direct impact of the cloud environment on carbon emissions of the energy source primarily defines its operational carbon footprint. As the carbon intensity of the energy source reduces, their supply variability increases more often, thus challenging application workload

performance. Maintaining required application latency performance while maximizing the integration of low-carbon intensity energy sources requires efficient resource management techniques. • Embodied Carbon Management: Deploying, maintaining, and scaling physical elements of cloud environments incur an indirect carbon cost associated with carbon emissions made during manufacturing, supplying, and recycling IT assets, which are embodied in them. As the demand for latency-sensitive applications increases, cloud environments must scale accordingly while maintaining the performance of IT assets through frequent hardware refreshes, leading to the accumulation of embodied carbon. Providers must employ efficient resource management techniques to slow embodied carbon accumulation while maintaining latency SLOs of application workloads by managing IT assets with heterogeneous reliability and performance characteristics. Next, we identify the challenges associated with carbon-aware resource management techniques, which are significant in latency-sensitive workloads. We analyze the stated challenges from the perspective of carbon footprint reduction and meeting latency SLOs of end users. • Diverse latency requirements of application workloads: Latency-sensitive application workloads exhibit varying degrees of latency tolerance in their service level agreements (SLOs). For instance, workloads may allow intermittent latency degradations, which end-user applications can tolerate. However, that may not be possible for time-critical applications, where SLOs expect strict bounds on the response time. Thus, these systems must apply carbon optimizations adapting to the application latency requirements. Further, applications may exhibit different SLOs within their components. Navigating through these complex latency requirements in cloud environments is challenging, where impacting applications SLOs can incur heavy penalties to the cloud operator. • Variable-availability of carbon-efficient energy: Often, energy sources with lesser carbon intensity are renewable sources with variable availability, such as solar and wind. Compared to stable sources, such as nuclear energy with low carbon intensity, variable availability renewable energy plants can be built and made operational with lesser cost and time. Additionally, energy capacity availability can spatially vary in geographically distributed cloud environments, which are increasingly prevailing due to capabilities of low latency application execution near end users. As a result, cloud environments are significantly challenged due to the intermittent availability of low-carbon energy capacities in both space and time to deliver consistent performance to meet application latency SLOs. Not integrating intermittent energy sources could significantly increase the cloud environment’s carbon footprint, whereas the opposite must ensure that application performance is intact. • Intermittent performance degradations in inter-cloud networks: Cloud environments typically consist of hyper-scale data centers. A typical deployment pattern for latency-sensitive cloud applications is to use networked data centers that are geographically spread. These inter-cloud networks communicate through Wide Area Networks (WAN), which can undergo intermittent traffic congestions, degrading the communication latency. Geo-spread availability of data centers enables cloud operators to redirect workload execution to match the dynamic availability of low-carbon energy capacity during the day. However, WAN traffic congestion challenges the opportunities to do so. Degraded network performance could increase traffic redirection delay, violating application latency SLOs. • Hard-constraints of resource allocation for latency performance: Specific latency-sensitive applications require deterministic system performance to ensure critical service delivery. For instance, real-time services must meet the hard bounds of their response times. In 2

order to do so, they demand cloud servers to be tuned for high-performance and rigid allocation of computing resources, such as isolated CPU cores for virtual machines and turning off power optimization features of the CPU. As a result, the power management capabilities of the cloud environment are limited, especially when integrating low-carbon renewable energy sources with variable availability. Further, rigid resource allocations make redirecting workloads among sites with better carbon efficiency difficult due to limited placement opportunities satisfying the same placement constraints. • Resource-constrained cloud environments: In order to satisfy the low-latency serving of application services, some latency-sensitive workloads must be deployed closer to users at the network’s edge. These are primarily metropolitan areas where building new data centers can be expensive. A typical deployment pattern is to deploy cloud servers in resourceconstrained cloud environments, such as colocated data centers, which allocate specific space and power budget limits to servers. In return, the site’s capability to utilize low-carbon energy sources or headroom with low-carbon servers is limited. For the provider, efficient use of the sites to reduce their carbon impact under resource-constrained environments becomes a complex task. • Degradation of IT assets: Due to strict performance requirements, low-latency applications may use underlying hardware in an unsustainable manner, leading to premature degradations and carbon-costly frequent hardware replacements. For the cloud operator, alleviating hardware degradation is paramount for both cost and carbon. Since most hardware degradations are slow processes, understanding application patterns that lead to degradations over time and uncovering opportunities to optimize that while maintaining short-term application latency performance can be challenging. Both aspects of resource management identified above need to be addressed, considering the discussed challenges. Researchers have experimented with various techniques to overcome these challenges and derive better system architectures, workload execution patterns, resource allocation techniques, and resource scheduling techniques. In this article, we conduct an in-depth review of the existing literature and identify a classification of the aspects that influence their decisions. In the classification, we discuss inherent challenges and concerns related to those aspects in the context of carbon optimization for latency-sensitive application workloads. Carbon optimization in cloud environments is determined by the features of the underlying system and the characteristics of the workload latency SLOs. Similarly, our classification comprises key design aspects of the systems, characteristics of the latency-sensitive workloads that the infrastructure is designed for, and the goals of the resource management. Further, using our proposed taxonomy, we summarize existing research on operational or embodied carbon aspects, system architecture designs, and resource scheduling techniques in executing latency-sensitive application workloads. Moreover, we propose ideas for future work to advance resource management in this context. Cloud operators would benefit from our classification by understanding the key focus areas of carbon optimization under latency-sensitive cloud environments and the existing approaches that have already been evaluated. Researchers studying resource management techniques can refer to existing approaches and future work ideas, which could subsequently form the basis for designing novel techniques. The classification provides a holistic overview of existing and emerging literature for carbon optimization in both embodied and operational carbon aspects. The rest of the article is organized as follows. An overview of the existing surveys and studies on carbon-aware resource management in latency-sensitive cloud environments is provided in Section 2. Section 3 presents the proposed taxonomy of resource management. Section 4 summarizes existing works on carbon-aware resource management in latency-sensitive cloud environments 3

based on the taxonomy. Finally, Section 5 discusses the identified gaps in the literature, and Section 6 concludes this article. 2

Related Surveys

Literature surveys investigate existing literature for an area of interest to understand its current state of progress better. In that regard, many surveys study the carbon efficiency aspect of cloud environments. Preliminary surveys characterize carbon efficient cloud computing as green cloud computing emphasizing integrating low-carbon intensity renewable energy sources [15, 21, 26]. In that, they study diverse aspects of resource management, such as workload scheduling, power management, and electronic waste management [15, 21, 26]. However, they model for generic abstractions of application workloads, such as either batch or interactive workloads, and generic cloud data centers with mostly the same hyper-scale data center system architectures designed for those. In contrast, more recently, a branch of surveys can be seen, which study carbon efficiency in cloud environments that are designed for emerging latency-sensitive workloads [2, 4, 6, 12, 17, 20, 23, 29]. In their studies, application latency-specific requirement are studied, such as low-latency streaming applications [4, 17] and Internet of Things (IoT) applications [2, 6, 12, 23, 29]. Yet none of these studies consider the holistic view of carbon-aware resource management under latencysensitive application execution. They consider carbon efficiency as one of many aspects of resource management and often overlook both operational and embodied carbon aspects. Given that power grids continue to decarbonize by integrating renewable energy sources, embodied carbon is more significant than ever. In this article, we address emerging latency-sensitive cloud applications through a holistic study of resource management techniques for both operational and embodied carbon management. 3

The Taxonomy

At the topmost level, our taxonomy segregates carbon-aware resource management into the two primary aspects that we identified: operational and embodied carbon management. Operational carbon management comprises workload management, control architecture, system tier, and latency tolerance. Embodied carbon management comprises pre-deployment and post-deployment stages. Figure 1 illustrates the proposed taxonomy. In the following sections, we discuss each category in detail, referring to techniques explored in literature so far. 3.1

Operational Carbon Management

Most works on carbon efficiency focus on reducing carbon emissions at the operational level of the cloud environment. Here, the foundational approach is to utilize low-carbon-intensive energy sources to power the IT assets. In doing so, resource management techniques are challenged by the variable availability of the renewable energy sources present today, such as solar and wind. Operators opt for variable-availability renewable energy plants due to the lesser cost and deployment times, in contrast to low-carbon-intensive energy plants that provide relatively stable electricity generation. For instance, nuclear and hydro plants capable of providing stable, low-carbon-intensive energy can take many years to finish construction, whereas a solar power plant can be deployed in a few years. Adversely, cloud operators integrating renewable energy sources must match their variable availability with the data center power load, introducing significant challenges in ensuring the timing constraints of latency-sensitive workloads over data center load adjustments. In that, resource management techniques must address the integrated management of application latency performance and data center power load. This section identifies and briefly reviews their methods under the key resource management aspects of operational carbon management. 4

Fig. 1. The taxonomy of carbon-aware resource management in latency-sensitive cloud environments.

3.1.1 Workload Management. Workload execution yields dynamic power consumption patterns in cloud servers. In clouds, workload execution is carried out in isolated environments from the cloud operator due to privacy concerns, leaving the operator with high-level abstractions of running workloads. In return, the operator must utilize those abstractions to match the server load to the dynamic availability of renewable energy sources. In the following sections, we discuss prominent methods present in the literature in doing that, emphasizing how existing works leverage those for latency-sensitive workloads. Admission Control and Placement: Distributed system architectures of cloud environments typically consist of multiple data centers or compute clusters that provide heterogeneous performances for latency and carbon efficiency. Therefore, cloud providers have the opportunity to optimize carbon efficiency via the admission control or placement decisions made upon the arrival of workloads. In that, workloads are scheduled to execution sites, improving renewable energy utilization with minimum impact on the latency performance. Yuan et al. [35] study the problem of scheduling tasks among distributed green data centers to meet their response time constraints. Data centers in their system model integrate on-site renewable energy and periodically relay energy metrics to a centralized task scheduler, such as the electricity price of the power grid and the conversion rate of wind and solar radiation to electricity. Users’ tasks arrive at the centralized scheduler and are queued according to the application to which each task belongs. They propose a task scheduling algorithm to split the queued tasks among distributed green data centers to minimize the combined energy cost of both power grid and renewable energy while strictly meeting the task’s delay-bounded constraints. In contrast to centralized task arrival, tasks can also arrive in each compute region. Chien et al. [3] design request direction algorithms to minimize the carbon cost of latency-sensitive generative AI inference requests. Their system model consists of multiple compute regions with different carbon intensities. When a compute region receives a user request, it will be directed to the region with minimum carbon intensity, where the placement decision depends upon the calculated latency of both computation and round trip network time for each compute region. Admission control and placement also fit well with workloads with bounded response times, such as real-time services, because carbon efficiency can be estimated prior to service deployment. Kaur et al. [14] investigate scheduling containers of real-time services in the edge-cloud continuum. They schedule containers among Kubernetes clusters having on-site renewable energy integration. Their scheduling algorithm encompasses carbon footprint minimization and constraints to maintain the real-time performance of services, such as reducing network interference. Load Shifting: Load shifting involves either suspending/resuming workloads (i.e., shifting in time) or geographically migrating workloads (i.e. shifting in space). Both aim to match the abundance of renewable energy, which varies based on time or location. Load shifting with latency-sensitive workloads primarily leverages shifting in space to avoid latency violation risks of suspending workloads. Shifting in space involves chasing renewable energy availability across geographical locations, in which the network latency performance and scheduling overhead must be efficiently managed. Sajid et al. [22] proposes connecting geographically distributed data centers with high-performance networks and addressing scheduling overhead with a blockchain-based decentralized approach. Sun et al. [31] leverage space shifting but use it as a last resort to incur a minimum impact on workloads. In return, they achieve average space shifting to 0.015 times per VM per hour, minimizing the associated latency overhead. For workloads that allow a tolerable latency slack, Sukprasert et al. [30] conduct space shifting within a subset of locations that incur latency impact within the tolerable slack. Limiting the use of a subset of locations can reduce overall carbon efficiency. To mitigate that, they demonstrate complementing existing workloads with shiftable batch workloads. 6

Murillo et al. [19] apply exploitation of tolerable latency slack in content delivery networks (CDNs) and show that in CDNs, increasing the latency slack to 60ms can lead to over 60% reduction of carbon emissions. To do so, they complement space shifting with a second form of shifting, called capacity shifting, which moves CDN capacity to greener regions at a coarser time scale of days or weeks. Another opportunity is the workload portability present within applications. Gsteiger et al. [8] exploit space-shifting opportunities within serverless applications. They leverage functions not in the application workflow’s critical path and use them to perform space shifting. As a result, end-to-end application latency is left intact. Space-shifting opportunities are also present in specific application load-balancing scenarios. Souza et al. [28] exploit space-shifting opportunities in distributed web services. They conduct dynamic server provisioning and load balancing for dynamic carbon intensities and network latency constraints across geo-distributed regions. Throttling: Throttling is an in-place technique to dynamically adjust the server performance without shifting the running load. Using throttling, the cloud operator can degrade the servers’ performance to reduce its power draw, matching the variable availability of renewable energy. In return, servers take a prolonged time to execute workload instructions, impacting workload service quality such as increasing the response latency in latency-sensitive applications. Therefore, throttling must be carefully orchestrated so that its impact on the application does not violate service level objectives (SLOs). Li et al. [16] use a data center power delivery architecture that provides access to both renewable energy sources and energy storage and propose a CPU hardware controller called Chameleon. Chameleon switches between two modes of power management in matching renewable energy dynamics: an energy mode that throttles server performance using dynamic voltage and frequency scaling and a performance mode that uses storage energy for power deficiencies. When applied, DVFS degrades CPU performance, affecting running workloads. The authors employ a reinforcement learning-based controlling mechanism to efficiently manage mode switching to mitigate that. Jahanshahi et al. [11] propose a data center power shaping framework that leverages throttling through DVFS to match the dynamics of renewable energy. They co-locate latency-critical workloads and latency-tolerant complementary workloads. Then, CPU cores are grouped into three categories: working, offset, and free. Latency-critical workloads are then pinned to working cores with minimal throttling possibilities, and the remaining cores are allocated for complementary workloads. With that, they leverage data center-level power shaping signals to manage core groups efficiently, balancing throttling and workload performance. Sharma et al. [25] propose an application-independent power management mechanism called Blink. Blink activates and deactivates servers to yield a duty cycle and, as a result, can control the average power consumption of servers. Applications that can leverage Blink architecture modify their design to adopt that. The duty cycle in Blink throttles servers, which is specified through a blinking policy. Blinking policy balances power management and application performance depending on the application. Experiments with latency-sensitive applications such as Memcached show blinking can be adopted with only a modest throttling overhead. Govindan et al. [7] experiment battery energy storage to utilize renewable energy and use application throttling to sustain battery longevity via lesser frequent discharge cycles. In combining both, they aim to balance between application throttling and premature battery failures. Agarwal et al. [1] propose application throttling through host virtualization to match renewable energy variations. They dynamically change the number of physical cores mapped to virtual machines while keeping the number of virtual cores intact. In return, application interruptions are avoided, yet CPU performance can be throttled. Their approach applies to applications that can tolerate a specific degree of degradation of latency performance. Souza et al. [27] propose a virtualized data center-level energy system to allow applications to control the energy sources according to their workload patterns. In return, applications can 7

better adapt to renewable energy dynamics by exploiting application-specific usage patterns. Here, applications are provided with APIs to throttle their application containers through power capping to control their power draw, which is then exercised to match their latency SLOs better. Preemption: Isolated execution environments in clouds typically do not reveal underlying infrastructure dynamics to users. For example, a virtual machine’s service level objectives (SLOs) are to provide a dedicated and stable physical machine, regardless of underlying resource allocation dynamics. Adversely, cloud operators lose the opportunity to offload infrastructure dynamics to end users where applicable, such as absorbing variable availability of renewable energy. However, recent cloud offerings, such as preemptible VMs [5, 24], attempt to narrow that by providing virtual machines that reflect infrastructure-level dynamics to users. Preemptible VMs unlock the opportunity to relay variable availability of renewable energy to the application layer in specific latency-sensitive cloud environments, given that the application layer provides fault tolerance. Preemption allows reducing server load upon loss of renewable energy capacity through fault tolerance. Sun et al. [31] leverage preemptive VMs to delay drawing power from the carbon-intensive grids in geographically distributed networks of modular data centers integrating renewable energy. Their application layer allows users to define fault tolerance by deploying both preemptive and regular VMs. In case of a loss of energy capacity, they first attempt to reduce the data center power draw via VM migrations. If that is insufficient, they opt to shut down preemptive VMs to further reduce the data center power draw rather than sourcing energy from the carbon-intensive power grid. In doing so, an up-time threshold is maintained for preemptive VMs, and VM shutdowns are conducted to adhere to that. 3.1.2 Control Architecture. Integrated management of workloads and variable availability of renewable energy sources requires coordination between different resource management elements such as workloads, data centers and servers, and power delivery. In that regard, there are two prominent control architectures: centralized and decentralized. In this section, we discuss how related works apply their chosen control architecture, adhering to the latency constraints of their cloud environment design. Centralized: A centralized architecture enables a single point of control, providing an ease of management to the cloud operator. For instance, it allows operators to monitor the cloud environment and execute scheduling logic centrally, yielding better maintenance of scheduling logic, such as extending scheduling logic for future requirements. Further, system elements such as servers and data centers in centralized control are synchronized by design, eliminating the need to perform complex state synchronizations. Yuan et al. [35] leverage centralized control to schedule tasks for applications replicated across multiple data centers that integrate on-site renewable energy. They employ per-application queues and execute scheduling logic to optimize green energy while meeting delay-bound constraints. Jahanshahi et al. [11] design PowerMorph: a power reshaping framework to support frequency regulation requirements of the power grids that can integrate renewable energy sources. PowerMorph reshapes the power consumption of data centers that execute latency-critical applications. A critical challenge in their design is that frequency regulation bids happen every hour, yet the data center should follow the regulation signal every two seconds, where the latter does not provide enough opportunity to perform cluster-level optimization. To overcome that, they offload it to the server level and utilize centralized control at the data center level to conduct regulation provisions. Sharma et al. [25] propose an application-independent centralized control plane that manages cluster power. Through the APIs, latency-sensitive applications interact with the control plane to regulate their power consumption through opportunities present in specific application patterns. 8

In return, the cluster can integrate renewable energy sources while offloading load adjustment to applications. Govindan et al. [7] introduce energy buffers using batteries to integrate renewable energy in data centers. They use a centralized peak power budget enforcer component to leverage a hybrid approach combining batteries and server throttling. The hybrid approach performs better in dynamically adapting workload service level agreements, such as latency constraints. Sun et al. [31] use a centralized VM scheduler to obtain the global perspective of a modular data center network that integrates on-site renewable energy. It then uses that to find the best fit for VMs to achieve optimized decisions across a mix of modular data centers, priority VMs, and non-priority VMs. In return, the least interruptions are made with priority VMs, such as VMs executing latency-sensitive workloads. Murillo et al. [19] conduct load shifting in content delivery networks deployed on geographically distributed edge data centers that integrate renewable energy. They implement a centralized management component that provides an information service for weather and energy costs, capacity, and demand. As a result, the central component makes informed decisions about spatial load balancing and capacity shifting, considering trade-offs between latency, carbon, and cost. Gsteiger et al. [8] leverage centralized carbon forecasting, pricing, and transmission latency to offload serverless workflows across geo-distributed regions with different carbon intensities. Through centralized management, they identify and deploy serverless functions that have the potential to offload while keeping latency-sensitive functions intact. Similarly, Souza et al. [28] conduct centralized carbon forecasting, load balancing, and provisioning for distributed web services to minimize emissions while reducing the latency caused by load balancing. Kaur et al. [14] facilitate real-time services across edge nodes integrating on-site renewable energy. In that, a centralized controller runs scheduling logic to schedule applications to meet energy and performance obligations. Decentralized: In decentralized control, cloud environments achieve improved scalability and autonomy of individual elements. Further, it eliminates single point of failures in the system design. Nevertheless, decentralized control can introduce additional synchronization overheads between individual system elements. As a result, decentralized control is rarely applied in the literature. However, given that the carbon efficiency of renewable energy sources improves as they spread spatially, decentralized control can deliver scalable cloud environments that provide low-latency application performances, while effectively utilizing geographically spread low-carbon intensity energy sources. Sajid et al. [22] propose a blockchain-based decentralized workload distribution and management model for geo-distributed data centers. It optimizes variability in renewable energy generation as a cost optimization problem and uses a blockchain model to employ a decentralized control architecture. Their work is designed for low-latency networks. Nevertheless, their approach can increase both request scheduling times and processing overheads, affecting application latency performances. To mitigate that, they design their management technique to optimize for both. Li et al. [16] integrate renewable energy in a decentralized manner by allocating individual power budgets to servers and introducing each with a power management microcontroller. In return, each server aims to maximize renewable energy usage individually. Server performance, such as workload latency, is maintained through the power profiles of the microcontroller, which switches between an energy-oriented profile or a performance profile to maintain adequate workload performance while utilizing both renewable energy and battery-stored energy. Agarwal et al. [1] similarly integrate renewable energy. They drive the availability of CPU cores to match renewable energy availability individually at the server level in a decentralized manner. They modify the virtualization layer to efficiently manage dynamic CPU core availability so that virtual machines can continue executing, eliminating service interruptions.

3.1.3 System Tier. Carbon-aware resource management techniques can be applied at multiple system tiers of the cloud environments. Given the diverse latency SLOs of latency-sensitive applications, each tier offers different optimization opportunities often specific to application dynamics. In this section, we discuss relevant works in the literature that apply operational carbon optimization at various system tiers. Server Level: Server-level integration of renewable energy commonly involves improving the power delivery architecture of the data center to supply a particular renewable energy allocation to each server. Then, server-level workload management ensures the server load adheres to the energy allocation. As a result, power management can be conducted without migrating workloads between servers, eliminating associated temporary service blackouts that impact application latency performance. Li et al. [16] leverage server-level energy allocation and introduce a micro-controller in each server to match the energy dynamics by throttling the server performance when needed, given that the workload throughput constraints allow that. Otherwise, battery-stored energy is used to match energy availability. Similarly, Agarwal et al. [1] use a server-level renewable energy budget. They employ dynamic core availability for running workloads to throttle at the server level when needed. In return, workload migration and the associated interruptions are avoided. Data Center Level: Data center-level integration of renewable energy allows resource management techniques to absorb the variable availability of renewable energy by adjusting the power consumption of servers as a whole. In return, resource management techniques can explore utilization patterns across the data center server to adjust its power consumption while minimizing the application latency performance. Sharma et al. [25] employ an application-independent power management framework that allows defining policies to engage servers in the data center to yield a duty cycle of activation and deactivation, thus reducing the average power draw of the data center to match renewable energy availability. They demonstrate that specific distributed latency-sensitive applications integrated into the framework through their APIs can adapt their application patterns to deliver acceptable performances. Govindan et al. [7] apply energy buffers in the data center power delivery with UPS batteries to absorb renewable energy variations. They use server throttling at the data center level to optimize battery life and application performance, such as workload latency. Souza et al. [27] leverage data center-level renewable energy integration to provide virtualized energy systems to applications. In return, individual applications executing in the data center manage grid energy, renewable energy, and batteries through virtualization, employing application-specific carbon budgeting policies that allow meeting their latency constraints. Data center and Server Levels: Given application characteristics, integrated management at both data center and server levels can yield better integration of renewable energy. In that, serverlevel integration can enable workload management, avoiding workload migrations, while data center-level management can complement that across the servers. Jahanshahi et al. [11] design a power reshaping framework for data centers catering to energy capacity signals from the power grid, such as the varying capacity of renewable energy integration. They execute both latency-sensitive and complementary workloads in servers. Due to short intervals of power shaping and long intervals of power biding with the grid, they employ server-level power management via workload throttling and core allocation among latency-sensitive and complementary workloads while conducting data center-level power biding at longer intervals. Inter-cloud Level: Integrating renewable energy at the inter-cloud level allows renewable energy availability to be utilized in different geographical locations. Here, the challenging aspect is maintaining the application performance of latency-sensitive workloads over the communication overhead of the network that connects geographical locations. In that, resource management 10

techniques aim to tackle the variable availability of renewable energy by directing workloads to locations with sufficient energy availability while adhering to latency constraints. Yuan et al. [35] conduct spatiotemporal task scheduling across geographically distributed green data centers. Their task scheduling problem applies a delay-bound constraint to maintain workload latency performance, and scheduling algorithms are designed to adhere to that. Sajid et al. [22] design their work for high-performance inter-cloud networks. Hence, they optimize application latency performance by minimizing both processing overhead and inter-cloud workload migration delays. Sun et al. [31] co-locate complementary workloads with latency-sensitive workloads in geo-distributed clouds, and aim to reduce the workload migration frequency of latency-sensitive workloads to minimize their latency impact. Sukprasert et al. [30] explore inter-cloud workload execution for latency-sensitive interactive workloads to harness renewable energy. They migrate interactive workloads to greener locations if latency constraints allow it. Murillo et al. [19] apply inter-cloud renewable energy harnessing to content delivery networks. They combine request latency-aware workload shifting with VM capacity shifting while maximizing renewable energy utilization within latency boundaries. Gsteiger et al. [8] utilize inter-cloud renewable energy availability for serverless applications. They segregate functions from serverless application workflows that can be offloaded without increasing end-to-end latency, and use that to utilize energy availability across clouds. Souza et al. [28] harness renewable energy across clouds for distributed web services. They provision resources based on carbon forecasting and conduct load balancing across clouds within the latency constraints. Chien et al. [3] investigate inter-cloud serving of generative AI requests. Directing requests to locations with minimum carbon intensity shows that renewable energy can be utilized without significantly impacting request latency. Kaur et al. [14] leverage a multi-cluster deployment across different locations to execute real-time services. They propose a controller to utilize renewable energy across locations while maintaining adequate application performance. 3.1.4 Latency Tolerance. Application workloads in latency-sensitive cloud environments exhibit various service level objectives (SLOs) in their latency performance. We classify them into two primary categories: Low-latency and Bounded-latency. Low-latency applications can tolerate latency performance degradations to a specific level, where resource management techniques aim to improve their service quality by minimizing such degradations. Bounded-latency applications, such as real-time applications, have bounded latency responses. For those, resource management techniques must strictly meet the response time boundaries. In this section, we discuss how relevant works achieve that in utilizing the variable-available renewable energy sources. Low-latency: Low latency applications that tolerate certain levels of latency degradation provide better flexibility in adapting to performance optimization of variable-available renewable energy integrations. As a result, most related works exploit that in two primary aspects: server performance throttling to match renewable energy availability and workload migration across geographical locations for greener energy availability. Li et al. [16] leverage dynamic throttling of server performance to engage an energy-efficient power profile that harnesses renewable energy. They employ techniques to dynamically switch between a high-performance power profile to minimize the impact of workload latency from server performance throttling. Jahanshahi et al. [11] use Dynamic Voltage Frequency Scaling (DVFS) to throttle CPU core performance, matching renewable energy availability. They minimize latency increases for latency-critical workloads by throttling cores allocated to complementary workloads first. Similarly, Govindan et al. [7] use DVFS-based throttling to optimize between latency performance and longevity of battery energy storage, and Agarwal et al. [1] use throttling through shrinking CPU core availability to exploit latency performance for renewable energy 11

harnessing. Sharma et al. [25] use application-independent blinking of server activation to match renewable energy availability, and offload optimizing latency performance over throttling to the application. A similar approach can be seen with Souza et al. [27], where a virtualized energy system allows applications to manage to throttle their containers according to workload patterns while minimizing the latency performance impact over variable availability of renewable energy. Besides server performance throttling, many works exploit low latency for renewable harnessing across geographical locations. Sajid et al. [22] exploits latency flexibilities in workloads with intermittent interruptions of workload availability to shift and execute them in greener data centers. They propose an optimized scheduling approach to minimize the interruption durations. A similar workload migration approach is used by Sun et al. [31] for modular data centers. They prioritize migrating VMs with fewer VM states to reduce the impact of migration overhead on latency performance. Murillo et al. [19] explore adjusting response latency time in content delivery networks to better harness renewable energy across edge data centers. They show that increasing latency within safe limits can reduce carbon emissions up to 35.5%. Gsteiger et al. [8] leverage low latency in serverless applications for carbon optimization by segregating functions in application workflows, and offloading those with lesser latency impact to greener locations. Souza et al. [28] increase request latency in distributed web services for renewable energy harvesting across locations by combining load balancing and resource provisioning. Chien et al. [3] apply renewable energy harnessing across geo-spread locations to serve generative AI requests, and show low-latency characteristics of requests allow operating within safe limits. Bounded-latency: Exploiting applications with bounded latency constraints for renewable energy integration is less common. This is due to the rigid nature of latency upper bounds, which do not provide much room for resource management techniques to absorb renewable energy variations. Yuan et al. [35] leverage a scheduling algorithm that strictly guarantees the task’s delay-bound constraints while maximizing the usage of renewable energy across distributed green data centers. Kaur et al. [14] explore real-time services with well-defined response time boundaries for harvesting renewable energy across Kubernetes clusters with on-site renewable energy integration. They propose a controller to optimize green energy utilization and performance impacts from application interference in container scheduling. 3.2

Embodied Carbon Management

Management of embodied carbon is broad, spanning across the lifecycle of IT assets. We classify those into two primary stages: Pre-deployment and Post-deployment. We then discuss related resource management techniques in planning, designing, managing, and recycling IT assets in cloud environments, referring to the primary stages that we identified, while narrowing our focus for those considering the impact to application latency performance. 3.2.1 Pre-deployment. Since embodied carbon management is conducted through the management of IT assets, the pre-deployment stage of cloud environments approaches that at the infrastructure design. A well-designed infrastructure for carbon efficiency reduces the need to optimize again once the infrastructure is available to execute workloads. Here, we focus on two primary aspects: infrastructure design and low-carbon component use, emphasizing the latency performance of running workloads as a design goal. Infrastructure Design: Cloud environments can exhibit heterogeneous designs depending on the workload performance requirements. Each design can offer specific opportunities to improve embodied carbon efficiency. Moreover, specific tools in estimating embodied carbon footprints can lead to better designs. Finally, techniques that maintain adequate application latency performance with a reduced hardware footprint can also reduce the embodied carbon footprint. 12

Sun et al. [31] investigate geographically distributed data center networks capable of serving low-latency workloads. They use modular data centers that integrate renewable energy sources (rMDC) in their design. rMDC incurs a lesser embodied carbon footprint than traditional data centers. Their work enables colocating rMDC with more stable energy sources, reducing the number of servers required to utilize peak energy spikes, thus further reducing the embodied carbon. Ji et al. [13] design a carbon estimation tool for servers with accelerators, such as GPU and FPGA, that serves latency-sensitive workloads. They advance accurate embodied carbon estimation and enable operators to derive improved carbon efficiency in their designs. Tannu et al. [32] explore embodied carbon intensity in storage mediums. Their work provides insightful guidance on selecting carbon-efficient storage mediums like HDD or SSD. SSDs provide better performance for workload latency than HDDs, yet their embodied carbon efficiency can differ. Therefore, carbon insights help operators make design decisions to balance workload performance and embodied carbon efficiency. Gupta et al. [9] aim to shrink the overall hardware footprint of the cloud environment using resource oversubscription, reducing underutilization and embodied carbon footprint. They design a dynamic system to observe resource underutilization and present a dynamic resource leasing platform. Through that, their system can execute latency-sensitive workloads with capacity availability service level agreements (SLOs) over the oversubscribed infrastructure. Results show that infrastructure shrinking as much as 25% is achievable. Low-carbon Component Use: Opting for components with lower embodied carbon footprints enables operators to optimize the carbon efficiency of their cloud environment design. However, the associated performance bottlenecks must be identified, and suitable techniques must be employed to mitigate those. Zhong et al. [37] propose a tiered memory system that reduces embodied carbon footprint. Instead of local DRAM, they use a hardware-managed tiered memory system for low-carbon CXL-based memory. Although the embodied carbon footprint of CXL is lower, it can incur higher latencies than local DRAM. They introduce a software stack to manage that. The combined hardware-managed tiering system and the software stack provide memory performance closer to local DRAM. 3.2.2 Post-deployment. Once deployed, a cloud environment’s embodied carbon footprint can be optimized by managing installed IT assets. Most works aim to extend an asset’s operating life or reuse older components. At the post-deployment stage, the embodied carbon footprint is already acquired. Thus, extending the asset’s lifetime or reusing older components allows operators to amortize the acquired carbon further. However, using aged components can increase component failure risks in the cloud environment. Resource management techniques must mitigate that by exploring opportunities in hardware-software settings while maintaining adequate workload latency performances. Lifetime Extension: Resource management techniques that optimize embodied carbon footprint through the asset’s lifetime extension exploit resource usage patterns specific to the hardware in focus. They manipulate resource usage patterns, and through that, they improve asset longevity. The resulting lifetime extension allows for further amortization of the asset’s embodied carbon footprint. Zhao et al. [36] aims to reduce uneven core wear-off in multi-core CPUs from executing workloads with core affinity, such as real-time workloads requiring deterministic performance. In return, premature failures of specific cores can be avoided for extended CPU life. They provide a performance metric independent of CPU micro-architectural characteristics to measure uneven CPU core usage, which can be used at the cloud resource management layer. Leveraging the proposed metric, the workload scheduler can shift workloads between CPU cores based on the CPU stress. Wang et al. [34] employ an aging-aware workload scheduler to maintain workload performance 13

among servers with heterogeneous aging characteristics. They identify older servers that could maintain sufficient performance under specific conditions, such as during low load times, and use that knowledge in the workload scheduler. Tannu et al. [32] reduce SSD wear-off in storage server fleets. They employ dynamic data redirection to locations with lower write intensity, evens out SSD writing across the server fleet, reducing their aging rate and allowing extended embodied carbon amortization. Similarly, Gupta et al. [10] reduce the SSD aging rate by increasing over-provisioning in their cloud environment setting to reduce the write amplification factor. McAllister et al. [18] exploit premature failure risks in flash-based cache. Instead of a legacy logical block addressable device interface, they propose fairyWREN, a flash cache designed for write read erase interface (WREN). WREN allows application control over data placement and garbage collection, which fairyWREN uses to reduce writes via caching policies. fairyWREN has a better read latency at peak load, improving the latency-sensitive performance of workloads and extending flash storage lifetime for improved embodied carbon amortization. Component Reuse: Used components can be employed in data centers to reduce acquired embodied carbon in newer components, given that the workload performance, such as latency requirements, is maintained. Used components have already amortized their initial embodied carbon footprint to a certain degree; thus, using them reduces the overall embodied carbon footprint of the cloud environment. Wang et al. [33] propose a framework to evaluate carbon savings at scale with used components, such as used memory and SSD. It enables providers to evaluate the performance at scale with used components, such as tail latency and low load latency. In return, operators can make an informed decision on using used components in their cloud environment. Tannu et al. [32] propose re-purposing used flash devices. They focus on Multi-level cell (MLC) devices, which store multiple bits within each cell, thus providing higher capacities. However, MLC can rapidly wear out compared to single-cell devices (SLC). The work proposes a strategy to transform used MLC into low-capacity SLC devices, enabling a second life to amortize embodied carbon. Chien et al. [3] explore embodied carbon optimization for the geo-distributed serving of low-latency generative AI requests, where used servers provide headroom in each location. The work evaluates the amount of headroom needed for effective carbon improvements. Gupta et al. [10] explore reusing general-purpose hardware instead of employing specialized accelerators. Their work shows a promising balance between component reuse and workload performance. 4

Classification of Resource Management Techniques Using Taxonomy

Table 1 and 2 outline key works on carbon-aware resource management in latency-sensitive cloud environments related to the proposed taxonomy, where Table 1 focus on operational carbon management and Table 2 focus on embodied carbon management. The works we present here propose novel resource management techniques exploring one or more resource management aspects that we have identified.

Throttling

Load Shifting

Admission Control and Placement Admission Control and Placement Throttling

[11]

[22]

[14]

[16]

[35]

Admission Control and Placement Throttling

[3]

Server Load

Task

Container

Request

Task

Request

VM, Container

Yes

[27]

Request

Load Shifting

[28]

[8]

[19]

[31]

Workload Management Load Matching Granularity Mixed Criticality Load Shifting, VM Yes Preemption Load Shifting Server Load Load Shifting Function Yes

Work

Data Center-level, Server-level Inter-Cloud

Inter-Cloud

System Tier

Inter-Cloud

Decentralized Server-level

Centralized

Data Center-level, Server-level Decentralized Inter-Cloud

Centralized

Control Kind

Server Fleets

Networked Clouds

Networked Clouds Networked Clouds

Server Fleets

Throughput Servers

Green DCs

Real-Time Services

Networked Distributed Clouds Web Services Server Virtualized Fleets or Containerized Apps Networked Gen. AI Clouds Serving

Networked Modular DCs Clouds Networked CDNs Clouds Networked Serverless Clouds

Environment Topology

Response Latency

Downtime, Criticality Response Latency End-toend Latency Response Latency Response Latency

Lowlatency

Boundedlatency

Response Latency

LowMigration latency Latency Bounded- Response latency Latency

Lowlatency

Lowlatency Lowlatency

Lowlatency Lowlatency Lowlatency

Latency SLO Tolerance Optimize

Approach

Designing, Planning Carbon Estimation Resource Oversubscription

Work

[31]

[34]

[36]

[37]

[32]

[9]

Component Selection, Workload Scheduling Component Selection Workload Scheduling Workload Scheduling

Throttling

[1]

[13]

Throttling

[7]

Server Load VM

Server Load

Data Center-level

Data Center-level Decentralized Server-level

Centralized

Centralized Server Fleets Server Fleets

Server Fleets

Virtualized Apps

Interrupttolerable Apps -

CXL Memory

Carbon vs Performance

Pre-deployment Infrastructure Low-carbon Design Components Energy Modular DC Stability Carbon vs Performance Size Reduction -

Server Aging

CPU Aging

Re-purpose Flash Devices

SSD Aging

Post-deployment Extended Life Component Reuse -

Lowlatency Lowlatency

Lowlatency Response Latency Response Latency

Response Latency

Servers

CPU Wear-off

Memory

Storage Medium, Write Intensity

Utilization

Accelerators

Memory Access Scheduling Overhead Age vs Performance

Accelerated Computing Capacityavailability SLOs Storage Access

Optimize Hardware App Latency Aspect Datacenter Size Network Delay

Table 1. Classification of Resource Management Techniques of Operational Carbon Reduction

Throttling

[25]

[3] -

Flash Device Aging -

SSD Aging

Used Components Used Servers

Generalpurpose H/W for Accelerators Storage Write Intensity Reusing at Scale Data Center Headroom

Storage Write Intensity, Accelerators

Table 2. Classification of Resource Management Techniques for Embodied Carbon Reduction

Workload Scheduling Carbon Estimation Workload Scheduling

[18]

[33]

Resource Over provisioning

[10]

Age vs Performance Age vs Performance

Storage Access, Generalpurpose H/W vs Performance Storage Access

Research Gaps and Future Directions

The in-depth review of carbon-aware resource management in latency-sensitive cloud environments we conducted highlights open problems with great potential for exploration. This section discusses those areas in detail, alongside the broader categories we have identified, for operational and embodied carbon efficiency. In return, we lay the groundwork for research and development work for the future. 5.1

Cloud Environment Characteristics

Emerging low-latency applications such as the Internet of Things (IoT) execute in cloud deployments that rent space and power with distributed resource-constrained environments such as colocation data centers. As tenants, clouds subscribe to limited power budgets from the colocation provider, which may integrate on-site renewable energy in its shared power delivery to improve carbon efficiency. As a result, tenant clouds are challenged with better utilization of renewable energy through shared power delivery, not just locally but across distributed deployments of similar clouds that provide varying renewable energy availabilities depending on the time of the day and the location. Most low-carbon intensive renewable energy integration solutions for cloud environments use data center-level allocation of renewable energy. Although that allows management of workloads across servers or even between clouds to match the renewable energy capacity intermittencies, server-level or core-level renewable energy allocations can benefit applications with strict latency SLOs preventing workload migrations. However, dynamic resource usage in clouds may allocate workloads unevenly across servers, which may yield underutilization of renewable energy with server or core-level energy allocations. Exploiting opportunities in resource management to improve such is an interesting open problem. Increasingly popular generative AI-based applications demand low-latency responses, which require cloud servers with accelerators for faster model computation. Although existing works segregating generative AI model computation workflows explore better energy efficiency through older servers, exploring embodied carbon impact in that has not been thoroughly exploited. For instance, defining latency SLOs for generative AI model inference requests can better inform the request scheduler to balance performance and carbon efficiency among older and newer generation servers, allowing cloud environments to deploy older generation servers or used servers to reduce their embodied carbon footprint. 5.2

Application Characteristics and Latency SLOs

Cloud providers increasingly offer infrastructure-as-a-service (IaaS) solutions that reveal application criticality to the cloud provider, such as evictable virtual machines. Leveraging such with application fault-tolerance for renewable energy intermittency is still an open challenge. For example, IaaS can be utilized to deploy application-specific middleware solutions serving latency-sensitive workloads that can tolerate intermittent VM failures. In that, exploiting application-specific latency service level objectives (SLOs) to integrate renewable energy with VM inevitability as a load-matching technique requires novel resource management techniques. Many renewable energy integration techniques for latency-sensitive cloud environments leverage low-latency applications. However, as the cloud paradigm continues to penetrate a wide range of use cases, applications with bounded latency SLOs, such as real-time services, correspond to a significant portion of the operational carbon footprint of cloud computing. Therefore, exploring potential opportunities to absorb the variable availability of renewable energy sources with bounded latency applications becomes paramount in reducing that. 18

Another aspect of bounded latency applications is that they often require servers tuned to high-performance power profiles, which can yield server components to stress over longer periods. As a result, server components can undergo premature failures, forcing the cloud provider to engage in frequent server replacements and increasing the cloud environment’s embodied carbon footprint. In this context, sustainable usage of cloud resources while maintaining an adequate server performance to meet application latency SLOs becomes critical in managing bounded latency applications. Opportunities in server longevity for embodied carbon reduction can be seen in increasingly prevailing generative AI inference clusters, which offload most of the computation stress to accelerators such as GPUs. As a result, resource usage patterns in such clusters can lead to underutilization of resources such as CPUs. Improving resource management techniques to identify underutilization patterns and leverage that to reduce computation stress in server components can improve their longevity, leading to amortizing embodied carbon over the improved lifetime. 5.3

Resource Management Approaches

Due to low-latency performance in executing applications at the network’s edge, geographically distributed cloud environments are becoming increasingly popular. Nevertheless, carbon optimization in those deployments today mostly leverages centralized control, mainly for the ease of maintenance and management, yielding bottlenecks such as limited scaling capabilities. As a solution, decentralized control can be implemented. However, decentralized resource management requires efficient synchronization between clouds with minimum latency overhead. Exploring application-specific opportunities, such as executing with mixed latency SLOs of both low-latency and bounded components, can lead to better implementation of decentralized control. Computing models such as serverless computing allow the resource management layer to make granular management decisions. For instance, instead of application-level scheduling, the resource management layer can schedule functions that may collectively conduct an application workflow. As a result, additional opportunities are available to make granular changes to the server power draw, which can match intermittent renewable energy availability. Exploiting that for various renewable energy allocation methods, such as server-level or core-level allocations, and various latency SLOs is an interesting research problem. Another aspect is the management of thermal load on server components to improve its longevity. Inefficient resource usage patterns could stress server components over time, creating thermal hotspots and leading to premature component failures. For instance, uneven CPU stress of infrastructure tasks such as virtualization and workloads tasks have been identified to create thermal hotspots in CPU cores. Further exploring it for bounded latency applications that may incur similar stress due to their dedicated allocation of CPU cores and exploiting scheduling those with infrastructure tasks and low-stress application tasks to improve CPU longevity for servers executing mixed latency SLO applications can yield better management of their embodied carbon footprint. 6

Summary

In this article, we presented a detailed review of the aspect of carbon-aware resource management, focusing on latency-sensitive cloud environments. We proposed a taxonomy for a holistic view of carbon-aware resource management in both operational and embodied carbon efficiency. We discussed both aspects of operational and embodied carbon management and analyzed the existing works using the taxonomy. Our taxonomy presents an in-depth view of both operational and embodied carbon aspects for cloud operators to identify carbon optimization opportunities to meet their short-term and long-term carbon efficiency goals. Further, it provides the groundwork for researchers to understand existing works in carbon-aware resource management to investigate and 19

build upon their work. Finally, we provide a gap analysis highlighting the identified challenges, emphasizing the great potential for future work. References [1] Anup Agarwal, Shadi Noghabi, Inigo Goiri, Srinivasan Seshan, and Anirudh Badam. 2023. Unlocking unallocated cloud capacity for long, uninterruptible workloads. In Proceedings of 20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23). 457–478. [2] Carlos Barrios and Mohan Kumar. 2023. Service Caching and Computation Reuse Strategies at the Edge: A Survey. ACM Comput. Surv. 56, 2, Article 43 (Sept. 2023), 38 pages. doi:10.1145/3609504 [3] Andrew A Chien, Liuzixuan Lin, Hai Nguyen, Varsha Rao, Tristan Sharma, and Rajini Wijayawardana. 2023. Reducing the Carbon Impact of Generative AI Inference (today and in 2035). In Proceedings of the 2nd Workshop on Sustainable Computer Systems. Article 11, 7 pages. [4] Reza Farahani, Zoha Azimi, Christian Timmerer, and Radu Prodan. 2024. Towards AI-Assisted Sustainable Adaptive Video Streaming Systems: Tutorial and Survey. arXiv preprint arXiv:2406.02302 (2024). [5] Google. 2022. Rethinking your VM strategy with Spot VMs. https://cloud.google.com/blog/topics/cost-management/ rethinking-your-vm-strategy-spot-vms [6] Mohammad Goudarzi, Marimuthu Palaniswami, and Rajkumar Buyya. 2022. Scheduling IoT Applications in Edge and Fog Computing Environments: A Taxonomy and Future Directions. ACM Comput. Surv. 55, 7, Article 152 (Dec. 2022), 41 pages. doi:10.1145/3544836 [7] Sriram Govindan, Anand Sivasubramaniam, and Bhuvan Urgaonkar. 2011. Benefits and limitations of tapping into stored energy for datacenters. 39 (June 2011), 341–352. [8] Viktor Urban Gsteiger, Pin Hong (Daniel) Long, Yiran (Jerry) Sun, Parshan Javanrood, and Mohammad Shahrad. 2024. Caribou: Fine-Grained Geospatial Shifting of Serverless Applications for Sustainability. In Proceedings of the ACM SIGOPS 30th Symposium on Operating Systems Principles. 403–420. [9] Nishant Gupta, Iyswarya Narayanan, Shivam Handa, Sayak Chakraborti, Pankit Thapar, Baohua Shan, Ariel Rao, Yuanlai Liu, Pengyuan Wang, Yuqing Wu, Qingyi Gao, Chris Chao-Chun Cheng, Sihan You, Louis Huang, Jingyuan Fan, Kenny Yu, Kevin Lin, Tengfei Mu, Parth Malani, Haiying Wang, Trey Lu, and Peter Zhang. 2024. Dynamic Idle Resource Leasing To Safely Oversubscribe Capacity At Meta. In Proceedings of the 2024 ACM Symposium on Cloud Computing. 792–810. [10] Udit Gupta, Mariam Elgamal, Gage Hills, Gu-Yeon Wei, Hsien-Hsin S. Lee, David Brooks, and Carole-Jean Wu. 2022. ACT: designing sustainable computer systems with an architectural carbon modeling tool. In Proceedings of the 49th Annual International Symposium on Computer Architecture. 784–799. [11] Ali Jahanshahi, Nanpeng Yu, and Daniel Wong. 2022. PowerMorph: QoS-Aware Server Power Reshaping for Data Center Regulation Service. ACM Trans. Archit. Code Optim. 19, Article 36 (2022). [12] Rathinaraja Jeyaraj, Anandkumar Balasubramaniam, Ajay Kumara M.A., Nadra Guizani, and Anand Paul. 2023. Resource Management in Cloud and Cloud-influenced Technologies for Internet of Things Applications. ACM Comput. Surv. 55, 12, Article 242 (March 2023), 37 pages. doi:10.1145/3571729 [13] Shixin Ji, Zhuoping Yang, Xingzhen Chen, Stephen Cahoon, Jingtong Hu, Yiyu Shi, Alex K. Jones, and Peipei Zhou. 2024. SCARIF: Towards Carbon Modeling of Cloud Servers with Accelerators. In Proceedings of the 2024 IEEE Computer Society Annual Symposium on VLSI (ISVLSI). 496–501. [14] Kuljeet Kaur, Sahil Garg, Georges Kaddoum, Syed Hassan Ahmed, and Mohammed Atiquzzaman. 2020. KEIDS: Kubernetes-Based Energy and Interference Driven Scheduler for Industrial IoT in Edge-Cloud Ecosystem. IEEE Internet of Things Journal 7, 5 (2020), 4228–4237. doi:10.1109/JIOT.2019.2939534 [15] Fanxin Kong and Xue Liu. 2014. A Survey on Green-Energy-Aware Power Management for Datacenters. 47, 2, Article 30 (Nov. 2014), 38 pages. https://doi.org/10.1145/2642708 [16] Chao Li, Xian Li, Rui Wang, Tao Li, Nilanjan Goswami, and Depei Qian. 2013. Chameleon: Adapting throughput server to time-varying green power budget using online learning. In Proceedings of the International Symposium on Low Power Electronics and Design (ISLPED). 100–105. [17] Xunyun Liu and Rajkumar Buyya. 2020. Resource Management and Scheduling in Distributed Stream Processing Systems: A Taxonomy, Review, and Future Directions. ACM Comput. Surv. 53, 3, Article 50 (May 2020), 41 pages. [18] Sara McAllister, Yucong "Sherry" Wang, Benjamin Berg, Daniel S. Berger, George Amvrosiadis, Nathan Beckmann, and Gregory R. Ganger. 2024. FairyWREN: A Sustainable Cache for Emerging Write-Read-Erase Flash Interfaces. In Proceedings of the 18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24). 745–764. [19] Jorge Murillo, Walid A. Hanafy, David Irwin, Ramesh Sitaraman, and Prashant Shenoy. 2024. CDN-Shifter: Leveraging Spatial Workload Shifting to Decarbonize Content Delivery Networks. In Proceedings of the 2024 ACM Symposium on Cloud Computing. 505–521. 20

[20] Rui Queiroz, Tiago Cruz, Jérôme Mendes, Pedro Sousa, and Paulo Simões. 2023. Container-based Virtualization for Real-time Industrial Systems—A Systematic Review. ACM Comput. Surv. 56, 3, Article 59 (Oct. 2023), 38 pages. doi:10.1145/3617591 [21] Laura-Diana Radu. 2017. Green Cloud Computing: A Literature Survey. Symmetry 9, 12 (2017). doi:10.3390/sym9120295 [22] Sara Sajid, Muhammad Jawad, Kanza Hamid, Muhammad U.S. Khan, Sahibzada M. Ali, Assad Abbas, and Samee U. Khan. 2021. Blockchain-based decentralized workload and energy management of geo-distributed data centers. Sustainable Computing: Informatics and Systems 29 (2021), 100–461. [23] Farah Aït Salaht, Frédéric Desprez, and Adrien Lebre. 2020. An Overview of Service Placement Problem in Fog and Edge Computing. ACM Comput. Surv. 53, 3, Article 65 (June 2020), 35 pages. doi:10.1145/3391196 [24] Varun Shandilya. 2020. Announcing the general availability of Azure Spot Virtual Machines. https://azure.microsoft. com/en-us/blog/announcing-the-general-availability-of-azure-spot-virtual-machines/ [25] Navin Sharma, Sean Barker, David Irwin, and Prashant Shenoy. 2011. Blink: managing server clusters on intermittent power. In Proceedings of the Sixteenth International Conference on Architectural Support for Programming Languages and Operating Systems. 185–198. [26] Junaid Shuja, Abdullah Gani, Shahaboddin Shamshirband, Raja Wasim Ahmad, and Kashif Bilal. 2016. Sustainable Cloud Data Centers: A survey of enabling techniques and technologies. Renewable and Sustainable Energy Reviews 62 (2016), 195–214. doi:10.1016/j.rser.2016.04.034 [27] Abel Souza, Noman Bashir, Jorge Murillo, Walid Hanafy, Qianlin Liang, David Irwin, and Prashant Shenoy. 2023. Ecovisor: A Virtual Energy System for Carbon-Efficient Applications. In Proceedings of the 28th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2. 252–265. [28] Abel Souza, Shruti Jasoria, Basundhara Chakrabarty, Alexander Bridgwater, Axel Lundberg, Filip Skogh, Ahmed AliEldin, David Irwin, and Prashant Shenoy. 2024. CASPER: Carbon-Aware Scheduling and Provisioning for Distributed Web Services. In Proceedings of the 14th International Green and Sustainable Computing Conference. 67–73. [29] Paulo Souza, Tiago Ferreto, and Rodrigo Calheiros. 2024. Maintenance Operations on Cloud, Edge, and IoT Environments: Taxonomy, Survey, and Research Challenges. ACM Comput. Surv. 56, 10, Article 256 (June 2024), 38 pages. [30] Thanathorn Sukprasert, Abel Souza, Noman Bashir, David Irwin, and Prashant Shenoy. 2024. On the Limitations of Carbon-Aware Temporal and Spatial Workload Shifting in the Cloud. In Proceedings of the Nineteenth European Conference on Computer Systems. 924–941. [31] Jinghan Sun, Zibo Gong, Anup Agarwal, Shadi Noghabi, Ranveer Chandra, Marc Snir, and Jian Huang. 2024. Exploring the Efficiency of Renewable Energy-based Modular Data Centers at Scale. In Proceedings of the 2024 ACM Symposium on Cloud Computing. 552–569. [32] Swamit Tannu and Prashant J. Nair. 2023. The Dirty Secret of SSDs: Embodied Carbon. SIGENERGY Energy Inform. Rev. 3 (2023), 4–9. [33] Jaylen Wang, Daniel S. Berger, Fiodar Kazhamiaka, Celine Irvene, Chaojie Zhang, Esha Choukse, Kali Frost, Rodrigo Fonseca, Brijesh Warrier, Chetan Bansal, Jonathan Stern, Ricardo Bianchini, and Akshitha Sriraman. 2024. Designing Cloud Servers for Lower Carbon. In Proceedings of the 2024 ACM/IEEE 51st Annual International Symposium on Computer Architecture (ISCA). 452–470. [34] Jaylen Wang, Udit Gupta, and Akshitha Sriraman. 2023. Peeling Back the Carbon Curtain: Carbon Optimization Challenges in Cloud Computing. In Proceedings of the 2nd Workshop on Sustainable Computer Systems. Article 8, 7 pages. [35] Haitao Yuan, Jing Bi, and MengChu Zhou. 2019. Spatiotemporal Task Scheduling for Heterogeneous Delay-Tolerant Applications in Distributed Green Data Centers. IEEE Transactions on Automation Science and Engineering 16 (2019), 1686–1697. [36] Jiechen Zhao, Katie Lim, Thomas Anderson, and Natalie Enright Jerger. 2023. The Case of Unsustainable CPU Affinity. In Proceedings of the 2nd Workshop on Sustainable Computer Systems. Article 1, 7 pages. [37] Yuhong Zhong, Daniel S. Berger, Carl Waldspurger, Ryan Wee, Ishwar Agarwal, Rajat Agarwal, Frank Hady, Karthik Kumar, Mark D. Hill, Mosharaf Chowdhury, and Asaf Cidon. 2024. Managing Memory Tiers with CXL in Virtualized Environments. In Proceedings of the 18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24). 37–56.

Carbon-aware Resource Management for Latency-Sensitive Cloud Computing Environments: A Taxonomy and Future Directions

Analyze & explore related research in Conceptio

Related documents