The accelerating deployment of artificial intelligence (AI) workloads, driven by recent advancements in AI technology, has significantly increased the demand for computing resources and supporting infrastructure. However, the physical and electrical capacity of data centers cannot scale at the same pace, which introduces a new challenge: accommodating rising compute demand under stringent power and space constraints. New high performance GPUs offer better power efficiency compared to previous generations, and liquid cooling systems are significantly more efficient than traditional air cooling. These advancements create an opportunity to upgrade data centers that accommodate more AI workloads with limitations in physical infrastructure and power availability.We propose CAPLAI, Capacity-Aware PLanning for AI infrastructure, an AI-assisted stochastic optimization framework for lifecycle planning in GPU data centers. Our method adopts large language models (llMs) to generate diverse and plausible future scenarios that capture demand growth, hardware efficiency decay, electricity prices, and resale market trends. These scenarios feed into a stochastic optimization model that determines GPU purchase, retirement, and cooling infrastructure upgrades. We evaluate our framework using real-world traces and constraints derived from a large-scale AI data center at Brookhaven National Lab. Compared to conventional threshold-based heuristics, our approach increases the effective GPU computing capacity within the same power limit by up to 36 % and reduces lifecycle operating cost by up to 32 %. Results demonstrate that capacityaware, AI-guided planning significantly enhances efficiency and robustness with the escalating demand and infrastructural limits.

GPU spatial sharing among jobs is an effective approach to increase resource utilization and reduce the monetary and environmental costs of running deep learning workloads. While hardware support for GPU spatial sharing already exists, accurately predicting GPU interference between colocated workloads remains a concern. This makes it challenging to improve GPU utilization by sharing the GPU between workloads without severely impacting their performance. Existing approaches to identify and mitigate GPU interference often require extensive profiling and/or hardware modifications, making them difficult to deploy in practice.

Quantum Internet has potential capabilities far beyond the traditional Internet and is thus a promising future platform for communication and computation. Entanglement is a cornerstone of quantum mechanics and forms the basis of numerous quantum applications in the quantum Internet. While existing studies primarily focus on two-user entanglement, a plethora of applications necessitates the leap to multi-user entanglement. This paper tackles the fundamental problem of multi-user entanglement routing in the quantum Internet, aiming to entangle multiple quantum users with a high entanglement rate. We abstract the problem as a novel graph routing problem, which is not readily addressed by existing graph problem solutions due to the unique characteristics of the quantum Internet. To address this problem, we first consider a sufficient condition ensuring a feasible solution's existence and design an algorithm with the optimal solution. Given the NP-Completeness and NP- Hardness of determining a feasible solution's existence and deriving an optimal solution in general cases, respectively, we propose two heuristic algorithms to offer efficient solutions, which are shown, via extensive simulations, to outperform the existing algorithms in terms of entanglement rates.

Adjusting batch sizes and adaptively tuning other hyperparameters can significantly speed up deep neural network (DNN) training. Despite the ubiquity of heterogeneous clusters, existing adaptive DNN training techniques solely consider homogeneous environments. Optimizing distributed DNN training over heterogeneous clusters is technically challenging, and directly adapting existing techniques results in low utilization and poor performance. To solve this problem, we introduce Cannikin - a novel data-parallel distributed training system. Cannikin achieves efficient and near optimal performance by accurately modeling the optimal system performance and predicting adaptive batch size training metrics for DNNs in heterogeneous clusters. We implemented Cannikin in PyTorch and conducted experiments over 16 GPUs in Chameleon. Empirical results show that Cannikin reduces DNN training in heterogeneous clusters by up to 52% compared to the state-of-art adaptive training system and up to 85% compared to native PyTorch DistributedDataParallel.

With the proliferation of smart devices in recent years, many applications requiring high computing capability and low latency have emerged. Edge computing is one of the promising paradigms to support such applications. Due to the high volatility of edge environments, e.g., frequent movements of mobile devices, varying task sizes, and time-variant channel conditions, we have to make the offloading and resource management decisions on the fly. This paper formulates and studies the problem of online task offloading and resource management in heterogeneous mobile edge environments. The goal of the problem is to minimize the overall system latency. We prove that the problem is NP-hard. Moreover, traditional algorithms needing long decision-making times are insufficient to support applications with high volatility. This paper proposes a deep learning-assisted online algorithm that can make fast decisions. In particular, we design an offline solver for the proposed problem and use a deep neural network to emulate the solver. We conduct extensive simulations to evaluate the proposed approach. Results show that the proposed approach is around$50,000\times$and$500\times$faster than the commercial Gurobi solver for the optimal solution and the proposed offline approximation solver, respectively. Moreover, the overall latency under the proposed approach is near-optimal.

Mobile edge computing has emerged as a prevalent computing paradigm to support applications that demand low latency and high computational capacity. Hardware reconfigurable accelerators exhibit high energy efficiency and low latency compared to general-purpose servers, making them ideal for integration into mobile edge computing systems. This paper investigates the problem of joint task offloading, access point selection, and resource allocation in heterogeneous edge environments for latency minimization. Given the heterogeneity of edge computing devices and the interdependence of the decisions required for offloading, access point selection, and resource allocation, it is challenging to optimize over them simultaneously. We decomposed the proposed problem into two disjoint subproblems and developed algorithms for each of them. The first subproblem is to jointly determine access point selection and communication resource allocation decisions, for which we have proposed an algorithm with a provable approximation ratio of$2.62/(1-8\lambda )$, where$\lambda$is a tunable parameter balancing the approximation ratio and time complexity. Additionally, we offer a faster variant of the algorithm with an approximation ratio of$(\sqrt{3}+1)^{2}$. The second subproblem is to determine offloading and computing resource allocation decisions jointly and is NP-hard, where we developed algorithms based on relaxation and rounding. We conducted comprehensive numerical simulations to evaluate the proposed algorithms, and the results demonstrated that our algorithms outperformed existing baselines and achieved near-optimal performance across various settings.

With the rapid advancement of edge computing and network function virtualization, it is promising to provide flexible and low-latency network services at the edge. However, due to the vulnerability of edge services and the volatility of edge computing system states, i.e., service request rates, failure rates, and resource prices, it is challenging to minimize the online service cost while providing the availability guarantee. This paper considers the problem of online virtual network function backup under availability constraints (OVBAC) for cost minimization in edge environments. We formulate the problem based on the characteristics of the volatility system states derived from real-world data and show the hardness of the formulated problem. We use an online backup deployment scheme named Drift-Plus-Penalty (DPP) with provable near-optimal performance for the AVBAC problem. In particular, DPP needs to solve an integer programming problem at the beginning of each time slot. We propose a dynamic programming-based algorithm that can optimally solve the problem in pseudo-polynomial time. Extensive real-world data-driven simulations demonstrate that DPP significantly outperforms popular baselines used in practice.

Quantum networks have emerged as a future platform for quantum information exchange and applications, with promising capabilities far beyond traditional communication networks. Remote quantum entanglement is an essential component of a quantum network. How to efficiently design a multi-routing entanglement protocol is a fundamental yet challenging problem. In this paper, we study a quantum entanglement routing problem to simultaneously maximize the number of quantum-user pairs and their expected throughput. Our approach is to formulate the problem as two sequential integer programming problems. We propose efficient entanglement routing algorithms for these two optimization problems and analyze their time complexity and performance bounds. Evaluation results highlight that our approach outperforms existing solutions in both the number of quantum-user pairs served and network throughput.

Mobile edge computing with the near-data processing paradigm can support applications requiring low latency and high computing capability, where energy cost is a significant part of the expenditure. This paper formulates and studies the problem of online joint task offloading and resource allocation for latency minimization subjecting to a time average energy cost constraint in mobile edge computing systems. The formulated problem has four time-variant system states, i.e., data lengths, task sizes, channel conditions, and electricity prices, which are modeled based on real-world data. At the beginning of each time slot, the system has to make five online decisions jointly: base station selection, server selection for task offloading, communication bandwidth allocation, computing resource allocation, and frequency scaling. We prove the offline version of the formulated problem is NP-hard. We design an online algorithm with a provable approximation ratio and low computational complexity for the proposed problem. In particular, it balances energy cost and latency based on the drift-plus-penalty algorithm and makes server and base station selection decisions using a game theoretic-based algorithm. We conduct extensive real-world data-driven simulations to evaluate the proposed algorithm. Simulation results show that the proposed approach outperforms popular baselines.

Generating a long-distance quantum entanglement is one of the most essential functions of a quantum network to support quantum communication and computing applications. The successful entanglement rate during a probabilistic entanglement process decreases dramatically with distance, and swapping is a widely-applied quantum technique to address this issue. Most existing entanglement routing protocols use a classic entanglement-swapping method based on Bell State measurements that can only fuse two successful entanglement links. This paper appeals to a more general and efficient swapping method, namely n-fusion based on Greenberger-Horne-Zeilinger measurements that can fuse n successful entanglement links, to maximize the entanglement rate for multiple quantum-user pairs over a quantum network. We propose efficient entanglement routing algorithms that utilize the properties of n-fusion for quantum networks with general topologies. Evaluation results highlight that our proposed algorithm under n-fusion can greatly improve the network performance compared with existing ones.

For many application domains, the integration of machine learning (ML) models into decision making is hindered by the poor explainability and theoretical guarantees of black box models. Although the emerging area of algorithms with predictions offers a way to leverage ML while enjoying worst-case guarantees, existing work usually assumes access to only one predictor. We demonstrate how to more effectively utilize historical datasets and application domain knowledge by intentionally using predictors of different quantities. By leveraging the heterogeneity in our predictors, we are able to achieve improved performance, explainability and computational efficiency over predictor-agnostic methods. Theoretical results are supplemented by large-scale empirical evaluations with production data demonstrating the success of our methods on optimization problems occurring in large distributed computing systems.

Mobile edge computing is becoming one of the ubiquitous computing paradigms to support applications requiring low latency and high computing capability. FPGA-based reconfigurable accelerators have high energy efficiency and low latency compared to general-purpose servers. Therefore, it is natural to incorporate reconfigurable accelerators in mobile edge computing systems. This paper formulates and studies the problem of joint task offloading, access point selection, and resource allocation in heterogeneous edge environments for latency minimization. Due to the heterogeneity in edge computing devices and the coupling between offloading, access point selection, and resource allocation decisions, it is challenging to optimize over them simultaneously. We decomposed the proposed problem into two disjoint subproblems and developed algorithms for them. The first subproblem is to jointly determine offloading and computing resource allocation decisions and is NP-hard, where we developed an algorithm based on semidefinite relaxation. The second subproblem is to jointly determine access point selection and communication resource allocation decisions, where we proposed an algorithm with a provable approximation ratio of 2.62. We conducted extensive numerical simulations to evaluate the proposed algorithms. Results highlighted that the proposed algorithms outperformed baselines and were near-optimal over a wide range of settings.

Introducing the emerging serverless paradigm into edge computing could avoid over- and under-provisioning of limited edge resources and make complex edge resource management transparent to application developers, which largely facilitates the cost-effectiveness, portability, and short time-to-market of edge applications. However, the computation/data dispersion and device/network heterogeneity of edge environments prevent current serverless computing platforms from acclimating to the network edge. In this paper, we address such challenges by formulating a container placement and data flow routing problem, which fully considers the heterogeneity of edge networks and the overhead of operating serverless platforms on resource-limited edge servers. We design an online algorithm to solve the problem. We further show its local optimum for each arriving container and prove its theoretical guarantee to the optimal offline solution. We also conduct extensive simulations based on practical experiment results to show the advantages of the proposed algorithm over existing baselines.

Edge caching is an effective way to reduce congestion and latency in 5G networks. Non-volatile memory (NVM) devices are developing fast, with the potential of fast access, and higher endurance versus traditional storage devices, to further boost mobile data offloading efficiency in 5G networks. This paper studies how to effectively use the two-layer storage system (NVM-enhanced) in 5G edge caching. We first model an edge caching optimization problem with NVM storage devices included and develop a parallel distributed algorithm with guaranteed convergence in joint caching and routing decisions. A fully decentralized algorithm for scenarios without any coordination is further developed which also guarantees the convergence. Real-world trace-driven simulations and experiments over a small-scale system demonstrate that NVM significantly boosts the performance of edge caching and the proposed algorithms outperform the existing ones.

Caching popular contents at the network edge is promising to reduce the retrieval latency, the network congestion, and the number of requests to the remote content provider during peak hours. In general, edge caching resource is costly and highly limited. Nevertheless, it is possible to provide cost-effective caching services using unreliable resources, which are resources reserved for other applications but have not been fully used or resources on vulnerable servers. In this paper, we consider the problem of caching popular contents over unreliable resources as a less expensive solution to limited edge caching capacity. In particular, to address the unreliability of edge resources, erasure coding is leveraged to increase the availability of cached contents. We formulate the problem as a discrete optimization problem and prove it is NP-hard. We start with two special cases of the problem and provide optimal algorithms for them. We then design an algorithm for the general version of the proposed problem and provide a provable performance guarantee. Real-world data-driven simulations demonstrate that the proposed algorithms significantly outperform popular baselines, and the rewards for the general version of the problem are near-optimal.

The fast development of mobile edge computing (MEC) and service virtualization brings new opportunities to the deployment of interactive applications, e.g., VR education, stream gaming, autopilot assistance, at the network edge for better performance. Ensuring quality of experience (QoE) for such services often requires the satisfaction of multiple quality of service (QoS) factors, e.g., short delay, high throughput rate, low packet loss. Nevertheless, existing mobile edge networks often fail to meet these requirements due to the mobility of end users and the volatility of network conditions. In this paper, we propose a novel scheme that both reduces delay and adjusts data throughput rate for QoE enhancement. We design an online service placement and throughput rate adjustment (SPTA) algorithm which coordinately migrates virtual services while tuning their data throughput rates based on real-time bandwidth fluctuation. By implementing a small-scale prototype supporting stream gaming at the edge, we show the necessity and feasibility of our work. Based on data from the experiments, we conduct real-world trace driven simulations to further demonstrate the advantages of our scheme over existing baselines.

Quantum networks are considered as a promising future platform for quantum information exchange and quantum applications, which have capabilities far beyond the traditional communication networks. Remote quantum entanglement is an essential component of a quantum network. How to efficiently design a multi-routing entanglement protocol is a fundamental yet challenging problem. In this paper, we study a quantum entanglement routing problem to simultaneously maximize the number of quantum-user pairs and their expected throughput. Our approach is to formulate the problem as two sequential integer programming steps. We propose efficient entanglement routing algorithms for the two integer programming steps and analyze their time complexity and performance bounds. Results of evaluation highlight that our approach outperforms existing solutions in both served quantum-user pairs numbers and the network expected throughput.

The emerging network function virtualization is migrating traditional middleboxes, e.g., firewalls, load balancers, proxies, from dedicated hardware to virtual network functions (VNFs) running on commercial servers defined as network points of presence (N-PoPs). VNFs further chain up for more complex network services called service function chains (SFCs). SFCs introduce new flexibility and scalability which greatly reduce expenses and rolling out time of network services. However, chasing the lowest cost may lead to congestion on popular N-PoPs and links, thus resulting in performance degradation or violation of service-level agreements. To address this problem, we propose a novel scheme that reduces the operating cost and controls network congestion at the same time. It does so by placing VNFs and routing flows among them jointly. Given the problem is NP-hard, we design an approximation algorithm named candidate path selection (CPS) with a theoretical performance guarantee. We then consider cases when SFC demands fluctuate frequently. We propose an online candidate path selection (OCPS) algorithm to handle such cases considering the VNF migration cost. OCPS is designed to preserve good performance under various migration costs and prediction errors. Extensive simulation results highlight that CPS and OCPS algorithms perform better than baselines and comparably to the optimal solution.

Emerging virtual network functions (VNFs) bring new opportunities to network services on the edge within customers’ premises. Network services are realized by chained up VNFs, which are called service function chains (SFCs). These services are deployed on commercial edge servers for higher flexibility and scalability. Despite such promises, it is still unclear how to provide highly available and cost-effective SFCs under edge resource limitations and time-varying VNF failures. In this paper, we propose a novel Reliability-aware Adaptive Deployment scheme named RAD to efficiently place and back up SFCs over both the edge and the cloud. Specifically, RAD first deploys SFCs to fully utilize edge resources. It then uses both static backups and dynamic ones created on the fly to guarantee the availability under the resource limitation of edge networks. RAD does not assume failure rates of VNFs but instead strives to find the sweet spot between the desired availability of SFCs and the backup cost. Theoretical performance bounds, extensive simulations, and small-scale experiments highlight that RAD provides significantly higher availability with lower backup costs compared with existing baselines.

Wireless data collection requires a sequence of resource provisioning decisions due to the limited battery capacity of wireless sensors. The corresponding online resource provisioning problem is challenging. Recently, many prediction methods have been proposed that can be used to benefit the performance of various systems through their incorporation. Therefore, in this article, we focus on online resource provisioning problems with short-term predictions motivated by the wireless data collection problem. Specifically, we design separate online algorithms for systems in which the state evolves in either a stationary manner or an arbitrarily determined manner and prove their performance bounds where their bounds improve as the amount of available predictions increases. Additionally, we design a meta-algorithm that can choose which online algorithm to implement at each point in time, depending on the recent behavior of the system environment. The practical performances of the proposed algorithms are corroborated in trace-driven numerical simulations of data collection of shared bikes. Additionally, we show that the performance of our meta-algorithm in various system environments can be better than that of the single best algorithm chosen in hindsight.

Cloud computing is becoming one of the ubiquitous computing paradigms for enterprises and organizations in recent years. Due to the volatility of system states such as cloud resource price and workload demand, it is challenging to provision cloud resources efficiently. This paper studies online cloud resource provisioning problems under cost budget where no accurate or distributional future information is available. We develop an algorithmic framework and design online algorithms based on the framework. We prove the competitive ratio of the proposed algorithms. We further show the proposed algorithms have better performance than a prominent existing algorithm named CR-Pursuit. While prior works on the problem require the objective functions to be concave, the proposed algorithms work for non-convex and non-concave objective functions. We conduct real-world trace-driven simulations. Results highlight the proposed algorithms outperform baselines significantly over a wide range of settings.

The increasing size of Deep Neural Networks (DNNs) has necessitated the use of multiple GPUs to host a single DNN model, a practice commonly referred to as model parallelism. The key challenge for model parallelism is to efficiently and effectively partition the DNN model across GPUs to avoid communication overheads while maximizing the GPU utilization, with the end-goal of minimizing the training time of DNN models. Existing approaches either take a long time(hours or even days) to find an effective partition or settle for sub-optimal partitioning, invariably increasing the end-to-end training effort. In this paper, we design and implement Pesto, a fast and near-optimal model placement technique for automatically partitioning arbitrary DNNs across multiple GPUs. The key idea in Pesto is to jointly optimize the model placement and scheduling at the fine-grained operation level to minimize inter-GPU communication while maximizing the opportunity to parallelize the model across GPUs. By carefully formulating the problem as an integer program, Pesto can provide the optimal placement and scheduling. We implement Pesto in TensorFlow and show that Pesto can reduce model training time by up to 31% compared to state-of-the-art approaches, across several large DNN models.

Programmable packet scheduling enables scheduling algorithms to be programmed into the data plane without changing the hardware. Existing proposals either have no hardware implementations for switch ASICs or require multiple strict-priority queues.

Modern deep learning frameworks support a variety of hardware, including CPU, GPU, and other accelerators, to perform computation. In this paper, we study how to schedule jobs over such interchangeable resources - each with a different rate of computation - to optimize performance while providing fairness among users in a shared cluster. We demonstrate theoretically and empirically that existing solutions and their straightforward modifications perform poorly in the presence of interchangeable resources, which motivates the design and implementation of AlloX. At its core, AlloX transforms the scheduling problem into a min-cost bipartite matching problem and provides dynamic fair allocation over time. We theoretically prove its optimality in an ideal, offline setting and show empirically that it works well in the online scenario by incorporating with Kubernetes. Evaluations on a small-scale CPU-GPU hybrid cluster and large-scale simulations highlight that AlloX can reduce the average job completion time significantly (by up to 95% when the system load is high) while providing fairness and preventing starvation.

The fast development of virtual network functions (VNFs) brings new opportunities to network service deployment on edge networks. For complicated services, VNFs can chain up to form service function chains (SFCs). Despite the promises, it is still not clear how to backup VNFs to minimize the cost while meeting the SFC availability requirements in an online manner. In this paper, we propose a novel self-adapting scheme named SAB to efficiently backup VNFs over both the edge and the cloud. Specifically, SAB uses both static backups and dynamic ones created on the fly to accommodate the resource limitation of edge networks. For each VNF backup, SAB determines whether to place it on the edge or the cloud, and if on the edge, which edge server to use for load balancing. SAB does not assume failure rates of VNFs but instead strives to find the sweet point between the desired availability of SFCs and the backup cost. Both theoretical performance bounds and extensive simulation results highlight that SAB provides significantly higher availability with lower backup cost compared with existing baselines.

The fast development of virtual network functions (VNFs) brings new challenges to providing reliability. The widely adopted approach of deploying backups incurs financial costs and environmental impacts. On the other hand, the recent trend of incorporating renewable energy into computing systems provides great potentials, yet the volatility of renewable energy generation presents significant operational challenges. In this paper, we optimize availability of VNFs under a limited backup budget and renewable energy using a dynamic strategy GVB. GVB applies a novel online algorithm to solve the VNF reliability optimization problem with non-stationary energy generation and VNF failures. Both theoretical bound and extensive simulation results highlight that GVB provides higher reliability compared with existing baselines.

Edge caching is an effective approach to improve the quality of service for mobile users and therefore a critical component for 5G networks. Despite the importance, it is not clear how to determine which contents to cache and how to the serve requests in 5G networks to minimize the total operational cost in a distributed and online manner, especially when some mobile users can be served by multiple small base stations. In this paper, we formulate an optimization problem to jointly decide the caching policy and the routing decision. There are two challenges: the need for distributed control and the lack of future information. We therefore develop an online distributed algorithm with provable performance guarantees in terms of convergence and competitive ratio compared to the offline optimal solution. Numerical simulations based on real-world traces highlight the significant performance improvement compared to existing baselines.

We investigate the problem of stochastic network optimization in presence of non-stationarity and estimations of average states in the future. Specifically, we first prove that the widely-used Drift and Penalty Algorithm in the Lyapunov optimization framework works well for non-stationary systems with periodical states. However, when the system is not periodical, non-stationarity may lead to severe performance degradation, which motivates the design of a novel, online algorithm named DPNP that incorporates the estimations of average future states into the stochastic optimization framework for decision making. DPNP is an online algorithm that requires zero a-prior distributional information about estimation errors. DPNP not only has near-optimal theoretical performance guarantees, but also outperforms existing Drift and Penalty Algorithm in numerical simulations. The improvement of DPNP highlights the importance of combining historic and future state estimations in non-stationary stochastic network optimization.

This paper considers how to cache popular contents and load balancing in 5G networks to minimize the total operating cost. Specifically, popular contents requested by mobile users (MUs) are cached in small base stations (SBSs) to serve them with better quality and lower cost because the SBSs are often much closer to MUs than the base station (BS). Due to limited caching capacity and bandwidth of SBSs, the caching policy and load balancing algorithm need to be carefully designed jointly and dynamically over time. In this paper, we formulate the joint content placement and load balancing by an online optimization problem. This problem is challenging because of the integer constraint in content placement and the lack of future information. We tackle the challenges in two progressive steps. First, we propose a primal-dual algorithm to solve the problem efficiently and prove it always achieves the optimal cost assuming all system information is available. Then we integrate promising online optimization algorithms with the proposed primal-dual algorithm so that only limited short-term predictions are needed. Theoretical performance bounds are also derived. We conduct extensive numerical simulations to evaluate the performance of proposed algorithms. Results highlight that the proposed online algorithms can reduce the system cost significantly (by as much as 27%) compared to the existing solutions and perform similarly to the offline optimal solution.

Emerging virtual network functions (VNFs) introduce new flexibility and scalability into traditional middlebox. Specifically, middleboxes are virtualized as software-based platforms running on commodity servers known as network points of presence (N-PoPs). Traditional network services are therefore realized by chained VNFs, i.e., service function chains (SFCs), running on potentially multiple N-PoPs. SFCs can be flexibly placed and routed to reduce operating cost. However, excessively pursuing low cost may incur congestion on some popular N-PoPs and links, which results in performance degradation or even violation of the service level of agreements.

We consider online convex optimization (OCO) problems with switching costs and noisy predictions. While the design of online algorithms for OCO problems has received considerable attention, the design of algorithms in the context of noisy predictions is largely open. To this point, two promising algorithms have been proposed: Receding Horizon Control (RHC) and Averaging Fixed Horizon Control (AFHC). The comparison of these policies is largely open. AFHC has been shown to provide better worst-case performance, while RHC outperforms AFHC in many realistic settings. In this paper, we introduce a new class of policies, Committed Horizon Control (CHC), that generalizes both RHC and AFHC. We provide average-case analysis and concentration results for CHC policies, yielding the first analysis of RHC for OCO problems with noisy predictions. Further, we provide explicit results characterizing the optimal CHC policy as a function of properties of the prediction noise, e.g., variance and correlation structure. Our results provide a characterization of when AFHC outperforms RHC and vice versa, as well as when other CHC policies outperform both RHC and AFHC.

Energy expenditure has become a significant fraction of data center operating costs. Recently, “geographical load balancing” has been proposed to reduce energy cost by exploiting the electricity price differences across regions. However, this reduction of cost can paradoxically increase total energy use. We explore whether the geographical diversity of Internet-scale systems can also provide environmental gains. Specifically, we explore whether geographical load balancing can encourage use of “green” renewable energy and reduce use of “brown” fossil fuel energy. We make two contributions. First, we derive three distributed algorithms for achieving optimal geographical load balancing. Second, we show that if the price of electricity is proportional to the instantaneous fraction of the total energy that is brown, then geographical load balancing significantly reduces brown energy use. However, the benefits depend strongly on dynamic energy pricing and the form of pricing used.

Demand response is crucial for the incorporation of renewable energy into the grid. In this paper, we focus on a particularly promising industry for demand response: data centers. We use simulations to show that, not only are data centers large loads, but they can provide as much (or possibly more) flexibility as large-scale storage if given the proper incentives. However, due to the market power most data centers maintain, it is difficult to design programs that are efficient for data center demand response. To that end, we propose that prediction-based pricing is an appealing market design, and show that it outperforms more traditional supply function bidding mechanisms in situations where market power is an issue. However, prediction-based pricing may be inefficient when predictions are inaccurate, and so we provide analytic, worst-case bounds on the impact of prediction error on the efficiency of prediction-based pricing. These bounds hold even when network constraints are considered, and highlight that prediction-based pricing is surprisingly robust to prediction error.

Demand response is a crucial aspect of the future smart grid. It has the potential to provide significant peak demand reduction and to ease the incorporation of renewable energy into the grid. Data centers' participation in demand response is becoming increasingly important given the high and increasing energy consumption and the flexibility in demand management in data centers compared to conventional industrial facilities. In this extended abstract we briefly describe recent work in our full paper on two demand response schemes to reduce a data center's peak loads and energy expenditure: workload shifting and the use of local power generations. In our full paper, we conduct a detailed characterization study of coincident peak data over two decades from Fort Collins Utilities, Colorado and then develop two algorithms for data centers by combining workload scheduling and local power generation to avoid the coincident peak and reduce the energy expenditure. The first algorithm optimizes the expected cost and the second one provides a good worst-case guarantee for any coincident peak pattern. We evaluate these algorithms via numerical simulations based on real world traces from production systems. The results show that using workload shifting in combination with local generation can provide significant cost savings (up to 40% in the Fort Collins Utilities' case) compared to either alone.

Recently, the demand for data center computing has surged, increasing the total energy footprint of data centers worldwide. Data centers typically comprise three subsystems: IT equipment provides services to customers; power infrastructure supports the IT and cooling equipment; and the cooling infrastructure removes heat generated by these subsystems. This work presents a novel approach to model the energy flows in a data center and optimize its operation. Traditionally, supply-side constraints such as energy or cooling availability were treated independently from IT workload management. This work reduces electricity cost and environmental impact using a holistic approach that integrates renewable supply, dynamic pricing, and cooling supply including chiller and outside air cooling, with IT workload planning to improve the overall sustainability of data center operations. Specifically, we first predict renewable energy as well as IT demand. Then we use these predictions to generate an IT workload management plan that schedules IT workload and allocates IT resources within a data center according to time varying power supply and cooling efficiency. We have implemented and evaluated our approach using traces from real data centers and production systems. The results demonstrate that our approach can reduce both the recurring power costs and the use of non-renewable energy by as much as 60% compared to existing techniques, while still meeting the Service Level Agreements.

Energy expenditure has become a significant fraction of data center operating costs. Recently, "geographical load balancing" has been suggested to reduce energy cost by exploiting the electricity price differences across regions. However, this reduction of cost can paradoxically increase total energy use.