Search

Reinforcement learning (RL) has proven highly effective in addressing complex decision-making and control tasks. However, in most traditional RL algorithms, the policy is typically parameterized as a diagonal Gaussian distribution with learned mean and variance, which constrains their capability to acquire complex policies. In response to this problem, we propose an online RL algorithm termed diffusion actor-critic with entropy regulator (DACER). This algorithm conceptualizes the reverse process of the diffusion model as a novel policy function and leverages the capability of the diffusion model to fit multimodal distributions, thereby enhancing the representational capacity of the policy. Since the distribution of the diffusion policy lacks an analytical expression, its entropy cannot be determined analytically. To mitigate this, we propose a method to estimate the entropy of the diffusion policy utilizing Gaussian mixture model. Building on the estimated entropy, we can learn a parameter $\alpha$ that modulates the degree of exploration and exploitation. Parameter $\alpha$ will be employed to adaptively regulate the variance of the added noise, which is applied to the action output by the diffusion model. Experimental trials on MuJoCo benchmarks and a multimodal task demonstrate that the DACER algorithm achieves state-of-the-art (SOTA) performance in most MuJoCo control tasks while exhibiting a stronger representational capacity of the diffusion policy.

Knowledge distillation (KD) has emerged as a promising technique in deep learning, typically employed to enhance a compact student network through learning from their high-performance but more complex teacher variant. When applied in the context of image super-resolution, most KD approaches are modified versions of methods developed for other computer vision tasks, which are based on training strategies with a single teacher and simple loss functions. In this paper, we propose a novel Multi-Teacher Knowledge Distillation (MTKD) framework specifically for image super-resolution. It exploits the advantages of multiple teachers by combining and enhancing the outputs of these teacher models, which then guides the learning process of the compact student network. To achieve more effective learning performance, we have also developed a new wavelet-based loss function for MTKD, which can better optimize the training process by observing differences in both the spatial and frequency domains. We fully evaluate the effectiveness of the proposed method by comparing it to five commonly used KD methods for image super-resolution based on three popular network architectures. The results show that the proposed MTKD method achieves evident improvements in super-resolution performance, up to 0.46dB (based on PSNR), over state-of-the-art KD approaches across different network structures. The source code of MTKD will be made available here for public evaluation.

Deep learning techniques have been applied in the context of image super-resolution (SR), achieving remarkable advances in terms of reconstruction performance. Existing techniques typically employ highly complex model structures which result in large model sizes and slow inference speeds. This often leads to high energy consumption and restricts their adoption for practical applications. To address this issue, this work employs a three-stage workflow for compressing deep SR models which significantly reduces their memory requirement. Restoration performance has been maintained through teacher-student knowledge distillation using a newly designed distillation loss. We have applied this approach to two popular image super-resolution networks, SwinIR and EDSR, to demonstrate its effectiveness. The resulting compact models, SwinIRmini and EDSRmini, attain an 89% and 96% reduction in both model size and floating-point operations (FLOPs) respectively, compared to their original versions. They also retain competitive super-resolution performance compared to their original models and other commonly used SR approaches. The source code and pre-trained models for these two lightweight SR approaches are released at https://pikapi22.github.io/CDISM/.

Large-scale cloud systems play a pivotal role in modern IT infrastructure. However, incidents occurring within these systems can lead to service disruptions and adversely affect user experience. To swiftly resolve such incidents, on-call engineers depend on crafting domain-specific language (DSL) queries to analyze telemetry data. However, writing these queries can be challenging and time-consuming. This paper presents a thorough empirical study on the utilization of queries of KQL, a DSL employed for incident management in a large-scale cloud management system at Microsoft. The findings obtained underscore the importance and viability of KQL queries recommendation to enhance incident management. Building upon these valuable insights, we introduce Xpert, an end-to-end machine learning framework that automates KQL recommendation process. By leveraging historical incident data and large language models, Xpert generates customized KQL queries tailored to new incidents. Furthermore, Xpert incorporates a novel performance metric called Xcore, enabling a thorough evaluation of query quality from three comprehensive perspectives. We conduct extensive evaluations of Xpert, demonstrating its effectiveness in offline settings. Notably, we deploy Xpert in the real production environment of a large-scale incident management system in Microsoft, validating its efficiency in supporting incident management. To the best of our knowledge, this paper represents the first empirical study of its kind, and Xpert stands as a pioneering DSL query recommendation framework designed for incident management., Comment: Accepted as a reseach paper at ICSE 2024

Existing learning-based autonomous driving (AD) systems face challenges in comprehending high-level information, generalizing to rare events, and providing interpretability. To address these problems, this work employs Large Language Models (LLMs) as a decision-making component for complex AD scenarios that require human commonsense understanding. We devise cognitive pathways to enable comprehensive reasoning with LLMs, and develop algorithms for translating LLM decisions into actionable driving commands. Through this approach, LLM decisions are seamlessly integrated with low-level controllers by guided parameter matrix adaptation. Extensive experiments demonstrate that our proposed method not only consistently surpasses baseline approaches in single-vehicle tasks, but also helps handle complex driving behaviors even multi-vehicle coordination, thanks to the commonsense reasoning capabilities of LLMs. This paper presents an initial step toward leveraging LLMs as effective decision-makers for intricate AD scenarios in terms of safety, efficiency, generalizability, and interoperability. We aspire for it to serve as inspiration for future research in this field. Project page: https://sites.google.com/view/llm-mpc

Motion prediction is crucial for autonomous vehicles to operate safely in complex traffic environments. Extracting effective spatiotemporal relationships among traffic elements is key to accurate forecasting. Inspired by the successful practice of pretrained large language models, this paper presents SEPT, a modeling framework that leverages self-supervised learning to develop powerful spatiotemporal understanding for complex traffic scenes. Specifically, our approach involves three masking-reconstruction modeling tasks on scene inputs including agents' trajectories and road network, pretraining the scene encoder to capture kinematics within trajectory, spatial structure of road network, and interactions among roads and agents. The pretrained encoder is then finetuned on the downstream forecasting task. Extensive experiments demonstrate that SEPT, without elaborate architectural design or manual feature engineering, achieves state-of-the-art performance on the Argoverse 1 and Argoverse 2 motion forecasting benchmarks, outperforming previous methods on all main metrics by a large margin.

This is a companion report for the OOPSLA 2023 paper of the same title, presenting a detailed end-to-end account of the $\lambda^*_{\mathsf{G}}$ graph IR, at a level of detail beyond a regular conference paper. Our first concern is adequacy and soundness of $\lambda^*_{\mathsf{G}}$, which we derive from a direct-style imperative functional language (a variant of Bao et al.'s $\lambda^*$-calculus with reachability types and a simple effect system) by a series of type-preserving translations into a calculus in monadic normalform (MNF). Static reachability types and effects entirely inform $\lambda^*_{\mathsf{G}}$'s dependency synthesis. We argue for its adequacy by proving its functional properties along with dependency safety via progress and preservation lemmas with respect to a notion of call-by-value (CBV) reduction that checks the observed order of effects. Our second concern is establishing the correctness of $\lambda^*_{\mathsf{G}}$'s equational rules that drive compiler optimizations (e.g., DCE, $\lambda$-hoisting, etc.), by proving contextual equivalence using logical relations. A key insight is that the functional properties of dependency synthesis permit a logical relation on $\lambda^*_{\mathsf{G}}$ in MNF in terms of previously developed logical relations for the direct-style $\lambda^*$-calculus. Finally, we also include a longer version of the conference paper's section on code generation and code motion for $\lambda^*_{\mathsf{G}}$ as implemented in Scala~LMS., Comment: arXiv admin note: text overlap with arXiv:2309.05885

Resource-constrained end devices can offload computation to backend clouds. The stochastic wireless channel that an end device is connected to can introduce bursty computation demand to the cloud. Specifically, under good channel conditions, a device can transmit more data to the cloud, which consequently yields higher instantaneous computation demand. Conversely, poor channel conditions can result in lower instantaneous demand. The performance indicator for such a mobile cloud computing system is the average of the response time, which is the time span from the arrival of the computation demand at the backend cloud instance to the completion of its execution. The question we target in this paper is how resources should be provisioned for the backend cloud instance to address this bursty computation demand and guarantee a desired quality-of-service (QoS), namely, a user-specified average response time. To answer this question, we model the mobile cloud computing system as two tandem queues. We analyze this queueing network using the fluid flow analysis framework, and derive the analytical relationship between the required resource capacity at the backend cloud instance and the desired QoS, given the workload generation process at the end device and the wireless channel conditions. Having obtained the required resource capacity for a desired QoS, we then determine whether it is economical to provision this resource capacity by subscribing to the traditional static instance or the recently introduced burstable instance offered by public cloud providers. Finally, trace-driven simulations validate our theoretical results.

In this paper, we consider a parking lot that manages the charging processes of its parked electric vehicles (EVs). Upon arrival, each EV requests a certain amount of energy. This request should be fulfilled before the EV’s departure. It is of critical importance to coordinate the EVs’ charging rates to smooth out the load profile of the parking lot because inappropriate charging rates can lead to sharp spikes and fluctuations on the load profile, imposing negative effects on the power grid. Meanwhile, empirical studies show that many parking lots exhibit statistical patterns on EV dynamics. For example, the bulk of EVs arrive during rush hours. Therefore, in this paper, we incorporate such patterns into charging rate coordination. Although the statistical patterns can be summarized from historical data, they are difficult to be analytically modeled. As a result, we adopt a model-free deep reinforcement learning approach. We also take the latest continuous charging rate control technology into consideration. The decision variables are thus continuous and a policy gradient algorithm is needed to perform reinforcement learning. Technically, we first formulate the problem as a Markov decision process (MDP) with unknown state transition probabilities. To further derive a deep policy gradient algorithm, the challenge lies in the inconsistent and state-dependent action space of the MDP model, due to the constraint to satisfy EVs’ energy demands before their scheduled departure. To tackle the challenge, we design a customized model for neural network training by extending the action space to be consistent and state-independent, and revise the reward function to penalize the neural network output if it is beyond the action space of the original MDP model. With this customized model, we then develop a deep policy gradient algorithm based on the proximal policy gradient framework. Numerical results show that our algorithm outperforms the benchmarks. IEEE

We consider end-to-end live video streaming in mobile edge networks with fluctuating but predictable throughput. We employ scalable video coding (SVC) to adapt the bitrate of each frame to the throughput evolution in a timely manner. When streaming SVC-encoded live video, the video source should dynamically derive and configure a set of coding parameters for each upcoming chunk, which consists of a number of consecutive frames. The coding parameters quantify the throughput evolution of an upcoming chunk by providing a few bitrate options for the frames within this chunk to encode with. Unlike the majority of existing works that simply treat the coding parameters as given (i.e., taking them for granted), in this paper, we derive the coding parameters from a comprehensive mathematical modeling and optimization perspective. Our approach consists of an optimization problem that captures the fundamental trade-off in determining the coding parameters and a customized search algorithm based on the analysis of the optimization problem's solution structure. Numerical results produced by the NS-3 network simulator show that our approach outperforms the state-of-the-art benchmarks. © 1967-2012 IEEE.

This paper proposes a tandem fluid queue model for a mobile computing system with computation offloading and analytically derives its key quality-of-service (QoS) metric. Based on the performance model, we further evaluate the economic benefits of burstable instances, a new type of cloud instances that are recently introduced to the market, and make suggestions on whether burstable instances should be used for mobile computing to save users’ costs under different wireless channel conditions and QoS requirements.

In this paper, we consider a parking lot that manages the charging processes of its parked electric vehicles (EVs). Upon arrival, each EV requests a certain amount of energy. This request should be fulfilled before the EV’s departure. It is of critical importance to coordinate the EVs’ charging rates to smooth out the load profile of the parking lot because inappropriate charging rates can lead to sharp spikes and fluctuations on the load profile, imposing negative effects on the power grid. Meanwhile, empirical studies show that many parking lots exhibit statistical patterns on EV dynamics. For example, the bulk of EVs arrive during rush hours. Therefore, in this paper, we incorporate such patterns into charging rate coordination. Although the statistical patterns can be summarized from historical data, they are difficult to be analytically modeled. As a result, we adopt a model-free deep reinforcement learning approach. We also take the latest continuous charging rate control technology into consideration. The decision variables are thus continuous and a policy gradient algorithm is needed to perform reinforcement learning. Technically, we first formulate the problem as a Markov decision process (MDP) with unknown state transition probabilities. To further derive a deep policy gradient algorithm, the challenge lies in the inconsistent and state-dependent action space of the MDP model, due to the constraint to satisfy EVs’ energy demands before their scheduled departure. To tackle the challenge, we design a customized model for neural network training by extending the action space to be consistent and state-independent, and revise the reward function to penalize the neural network output if it is beyond the action space of the original MDP model. With this customized model, we then develop a deep policy gradient algorithm based on the proximal policy gradient framework. Numerical results show that our algorithm outperforms the benchmarks. IEEE

In this paper, we consider a parking lot that manages the charging processes of its parked electric vehicles (EVs). Upon arrival, each EV requests a certain amount of energy. This request should be fulfilled before the EV’s departure. It is of critical importance to coordinate the EVs’ charging rates to smooth out the load profile of the parking lot because inappropriate charging rates can lead to sharp spikes and fluctuations on the load profile, imposing negative effects on the power grid. Meanwhile, empirical studies show that many parking lots exhibit statistical patterns on EV dynamics. For example, the bulk of EVs arrive during rush hours. Therefore, in this paper, we incorporate such patterns into charging rate coordination. Although the statistical patterns can be summarized from historical data, they are difficult to be analytically modeled. As a result, we adopt a model-free deep reinforcement learning approach. We also take the latest continuous charging rate control technology into consideration. The decision variables are thus continuous and a policy gradient algorithm is needed to perform reinforcement learning. Technically, we first formulate the problem as a Markov decision process (MDP) with unknown state transition probabilities. To further derive a deep policy gradient algorithm, the challenge lies in the inconsistent and state-dependent action space of the MDP model, due to the constraint to satisfy EVs’ energy demands before their scheduled departure. To tackle the challenge, we design a customized model for neural network training by extending the action space to be consistent and state-independent, and revise the reward function to penalize the neural network output if it is beyond the action space of the original MDP model. With this customized model, we then develop a deep policy gradient algorithm based on the proximal policy gradient framework. Numerical results show that our algorithm outperforms the benchmarks. IEEE

This paper proposes a tandem fluid queue model for a mobile computing system with computation offloading and analytically derives its key quality-of-service (QoS) metric. Based on the performance model, we further evaluate the economic benefits of burstable instances, a new type of cloud instances that are recently introduced to the market, and make suggestions on whether burstable instances should be used for mobile computing to save users’ costs under different wireless channel conditions and QoS requirements.

We consider end-to-end live video streaming in mobile edge networks with fluctuating but predictable throughput. We employ scalable video coding (SVC) to adapt the bitrate of each frame to the throughput evolution in a timely manner. When streaming SVC-encoded live video, the video source should dynamically derive and configure a set of coding parameters for each upcoming chunk, which consists of a number of consecutive frames. The coding parameters quantify the throughput evolution of an upcoming chunk by providing a few bitrate options for the frames within this chunk to encode with. Unlike the majority of existing works that simply treat the coding parameters as given (i.e., taking them for granted), in this paper, we derive the coding parameters from a comprehensive mathematical modeling and optimization perspective. Our approach consists of an optimization problem that captures the fundamental trade-off in determining the coding parameters and a customized search algorithm based on the analysis of the optimization problem's solution structure. Numerical results produced by the NS-3 network simulator show that our approach outperforms the state-of-the-art benchmarks. © 1967-2012 IEEE.

This paper proposes a tandem fluid queue model for a mobile computing system with computation offloading and analytically derives its key quality-of-service (QoS) metric. Based on the performance model, we further evaluate the economic benefits of burstable instances, a new type of cloud instances that are recently introduced to the market, and make suggestions on whether burstable instances should be used for mobile computing to save users’ costs under different wireless channel conditions and QoS requirements.

We consider end-to-end live video streaming in mobile edge networks with fluctuating but predictable throughput. We employ scalable video coding (SVC) to adapt the bitrate of each frame to the throughput evolution in a timely manner. When streaming SVC-encoded live video, the video source should dynamically derive and configure a set of coding parameters for each upcoming chunk, which consists of a number of consecutive frames. The coding parameters quantify the throughput evolution of an upcoming chunk by providing a few bitrate options for the frames within this chunk to encode with. Unlike the majority of existing works that simply treat the coding parameters as given (i.e., taking them for granted), in this paper, we derive the coding parameters from a comprehensive mathematical modeling and optimization perspective. Our approach consists of an optimization problem that captures the fundamental trade-off in determining the coding parameters and a customized search algorithm based on the analysis of the optimization problem's solution structure. Numerical results produced by the NS-3 network simulator show that our approach outperforms the state-of-the-art benchmarks. © 1967-2012 IEEE.

Leading cloud providers recently introduced a new instance type named burstable instances to better match the time-varying workloads of tenants and further reduce their costs. In the research community, however, little has been done to understand burstable instances from a theoretical perspective. This paper presents the first unified framework to model, analyze, and optimize the operation of burstable instances. Specifically, we model the resource provisioning of burstable instances, identify key performance metrics, and derive the analytical performance given the resource provisioning decisions. We then characterize the equilibrium behind tenants' responses to the prices offered for different burstable instance service classes, taking into account the impact of tenants' actions on the performance achieved by each service class. In addition, we investigate how a cloud provider can leverage knowledge of this equilibrium to find the prices that maximize its total revenue. Finally, we validate our framework on real traces and demonstrate its usage to price burstable offerings in a public cloud. IEEE

This paper presents learning on the fly (LoF), a two-stage online framework to predict the achievable application-layer throughput in the downlink data communication from an unmanned aerial vehicle (UAV) to a ground access point. LoF is based on a recurrent neural network (RNN). While the UAV is flying, LoF trains the RNN with constantly observed throughput and in the meantime, makes throughput predictions for the near future. Both the training and prediction can concurrently run on a non-GPU device at the network edge (e.g., on the UAV). To this end, we design LoF with a lightweight RNN architecture and a customized training process by weighted sampling on a sliding window. We implement LoF using PyTorch. Numerical results show that LoF is able to achieve an average prediction accuracy of 87.65%, outperforming existing approaches in the literature.

Recent studies have reported that big data analytics clusters, such as Hadoop, create substantial power peaks, which bring instability and inflexibility issues to the power grid. Substantial power peaks also lead to high penalty charges from electric utility companies, accounting for more than 30% of the electricity bill for a cluster operator according to empirical studies. Therefore, in this paper, we present an optimal framework that schedules computing jobs in large-scale data analytics clusters to mitigate power peaks. The scheduling model captures important properties of modern distributed data analytics clusters, including bundled resource provisioning and job-to-task decomposition with distributed processing. The scheduling problem is formulated as a nonlinear integer program. Its solution is derived by decomposing it into two classes of sub-problems and solving each class with an exact and efficient solution method. As a direct application, we detail the implementation of our proposed scheduling framework on a Hadoop cluster, and demonstrate its efficacy by extensive trace-driven simulations based on the CloudSim simulator.

Recent studies have reported that big data analytics clusters, such as Hadoop, create substantial power peaks, which bring instability and inflexibility issues to the power grid. Substantial power peaks also lead to high penalty charges from electric utility companies, accounting for more than 30% of the electricity bill for a cluster operator according to empirical studies. Therefore, in this paper, we present an optimal framework that schedules computing jobs in large-scale data analytics clusters to mitigate power peaks. The scheduling model captures important properties of modern distributed data analytics clusters, including bundled resource provisioning and job-to-task decomposition with distributed processing. The scheduling problem is formulated as a nonlinear integer program. Its solution is derived by decomposing it into two classes of sub-problems and solving each class with an exact and efficient solution method. As a direct application, we detail the implementation of our proposed scheduling framework on a Hadoop cluster, and demonstrate its efficacy by extensive trace-driven simulations based on the CloudSim simulator.

This paper presents learning on the fly (LoF), a two-stage online framework to predict the achievable application-layer throughput in the downlink data communication from an unmanned aerial vehicle (UAV) to a ground access point. LoF is based on a recurrent neural network (RNN). While the UAV is flying, LoF trains the RNN with constantly observed throughput and in the meantime, makes throughput predictions for the near future. Both the training and prediction can concurrently run on a non-GPU device at the network edge (e.g., on the UAV). To this end, we design LoF with a lightweight RNN architecture and a customized training process by weighted sampling on a sliding window. We implement LoF using PyTorch. Numerical results show that LoF is able to achieve an average prediction accuracy of 87.65%, outperforming existing approaches in the literature.

Leading cloud providers recently introduced a new instance type named burstable instances to better match the time-varying workloads of tenants and further reduce their costs. In the research community, however, little has been done to understand burstable instances from a theoretical perspective. This paper presents the first unified framework to model, analyze, and optimize the operation of burstable instances. Specifically, we model the resource provisioning of burstable instances, identify key performance metrics, and derive the analytical performance given the resource provisioning decisions. We then characterize the equilibrium behind tenants' responses to the prices offered for different burstable instance service classes, taking into account the impact of tenants' actions on the performance achieved by each service class. In addition, we investigate how a cloud provider can leverage knowledge of this equilibrium to find the prices that maximize its total revenue. Finally, we validate our framework on real traces and demonstrate its usage to price burstable offerings in a public cloud. IEEE

Recent studies have reported that big data analytics clusters, such as Hadoop, create substantial power peaks, which bring instability and inflexibility issues to the power grid. Substantial power peaks also lead to high penalty charges from electric utility companies, accounting for more than 30% of the electricity bill for a cluster operator according to empirical studies. Therefore, in this paper, we present an optimal framework that schedules computing jobs in large-scale data analytics clusters to mitigate power peaks. The scheduling model captures important properties of modern distributed data analytics clusters, including bundled resource provisioning and job-to-task decomposition with distributed processing. The scheduling problem is formulated as a nonlinear integer program. Its solution is derived by decomposing it into two classes of sub-problems and solving each class with an exact and efficient solution method. As a direct application, we detail the implementation of our proposed scheduling framework on a Hadoop cluster, and demonstrate its efficacy by extensive trace-driven simulations based on the CloudSim simulator.

This paper presents learning on the fly (LoF), a two-stage online framework to predict the achievable application-layer throughput in the downlink data communication from an unmanned aerial vehicle (UAV) to a ground access point. LoF is based on a recurrent neural network (RNN). While the UAV is flying, LoF trains the RNN with constantly observed throughput and in the meantime, makes throughput predictions for the near future. Both the training and prediction can concurrently run on a non-GPU device at the network edge (e.g., on the UAV). To this end, we design LoF with a lightweight RNN architecture and a customized training process by weighted sampling on a sliding window. We implement LoF using PyTorch. Numerical results show that LoF is able to achieve an average prediction accuracy of 87.65%, outperforming existing approaches in the literature.

Leading cloud providers recently introduced a new instance type named burstable instances to better match the time-varying workloads of tenants and further reduce their costs. In the research community, however, little has been done to understand burstable instances from a theoretical perspective. This paper presents the first unified framework to model, analyze, and optimize the operation of burstable instances. Specifically, we model the resource provisioning of burstable instances, identify key performance metrics, and derive the analytical performance given the resource provisioning decisions. We then characterize the equilibrium behind tenants' responses to the prices offered for different burstable instance service classes, taking into account the impact of tenants' actions on the performance achieved by each service class. In addition, we investigate how a cloud provider can leverage knowledge of this equilibrium to find the prices that maximize its total revenue. Finally, we validate our framework on real traces and demonstrate its usage to price burstable offerings in a public cloud. IEEE

Most bacteria cells divide by binary fission which is part of a bacteria cell cycle and requires tight regulations and precise coordination. Fast separation of Staphylococcus Aureus (S. Aureus) daughter cells, named as popping event, has been observed in recent experiments. The popping event was proposed to be driven by mechanical crack propagation in the peripheral ring which connected two daughter cells before their separation. It has also been shown that after the fast separation, a small portion of the peripheral ring was left as a hinge. In the article, we develop a fracture mechanics model for the crack growth in the peripheral ring during S. Aureus daughter cell separation. In particular, using finite element analysis, we calculate the energy release rate associated with the crack growth in the peripheral ring, when daughter cells are inflated by a uniform turgor pressure inside. Our results show that with a fixed inflation of daughter cells, the energy release rate depends on the crack length non-monotonically. The energy release rate reaches a maximum value for a crack of an intermediate length. The non-monotonic relationship between the energy release rate and crack length clearly indicates that the crack propagation in the peripheral ring can be unstable. The computed energy release rate as a function of crack length can also be used to explain the existence of a small portion of peripheral ring remained as hinge after the popping event. ©2018 Springer-Verlag GmbH Germany, part of Springer Nature., Natural Science Foundation of China (11572179), Natural Science Foundation of China (11432008)

Leading cloud providers recently introduced a new instance type named burstable instances to better match the time-varying workloads of tenants and further reduce their costs. In the research community, however, little has been done to understand burstable instances from a theoretical perspective. This paper presents the first unified framework to model, analyze, and optimize the operation of burstable instances. Specifically, we model the resource provisioning of burstable instances in different service classes, identify key performance metrics, and derive the performance given the resource provisioning decisions. We then characterize the equilibrium behind tenants' responses to the prices offered for different burstable instance service classes, taking into account the impact of tenants' actions on the performance achieved by each service class. In addition, we investigate how a cloud provider can leverage the knowledge of this equilibrium to find the prices that maximize its total revenue. Finally, we validate our framework on real traces and demonstrate its usage to price a public cloud. © 2019 IEEE.

Migrating applications to clouds has been a trend over the past decade or so. From a cloud operator's perspective, existing works have focused on how resources should be provisioned to users with a guaranteed level of quality-of-service (QoS). However, most of these works are for generic applications, where the distinct features of specific applications are not considered. In this thesis, we take advantage of applications' distinct features in the resource provisioning procedure to improve the total revenue of the operator and provide better QoS to users. We first focus on two types of applications that are widely deployed in clouds, namely, applications with bursty workloads and stream data analytics. In the fiirst technical chapter, we investigate how burstable instances, an instance type that was recently introduced by leading cloud operators, can match the time-varying workloads of applications. We present the first unified framework to model, analyze, and optimize the operation of burstable instances. In the second technical chapter, we consider a heterogeneous cloud-based cluster for stream data analytics, which is shared by multiple analytics jobs. An efficient resource allocation scheme is proposed to achieve max-min fairness in the utilities of the throughput for the jobs. Finally, we move to Internet of Things (IoT) applications in the third technical chapter. In view of the stringent communication delay requirements between the computation facilities and end users from many IoT applications, fog computing has recently been proposed, where computation tasks can be offloaded extensively along the cloud-to-things continuum. In this chapter, we derive an offloading scheme for a heterogeneous fog computing network shared by multiple tasks that have heterogeneous delay requirements, where lexicographic max-min fairness is enforced in the offloading procedure.

Leading cloud providers recently introduced a new instance type named burstable instances to better match the time-varying workloads of tenants and further reduce their costs. In the research community, however, little has been done to understand burstable instances from a theoretical perspective. This paper presents the first unified framework to model, analyze, and optimize the operation of burstable instances. Specifically, we model the resource provisioning of burstable instances in different service classes, identify key performance metrics, and derive the performance given the resource provisioning decisions. We then characterize the equilibrium behind tenants' responses to the prices offered for different burstable instance service classes, taking into account the impact of tenants' actions on the performance achieved by each service class. In addition, we investigate how a cloud provider can leverage the knowledge of this equilibrium to find the prices that maximize its total revenue. Finally, we validate our framework on real traces and demonstrate its usage to price a public cloud. © 2019 IEEE.

Distributed stream big data analytics platforms have emerged to tackle the continuously generated data streams. In stream big data analytics, the data processing workflow is abstracted as a directed graph referred to as a topology. Data are read from the storage and processed tuple by tuple, and these processing results are updated dynamically. The performance of a topology is evaluated by its throughput. This paper proposes an efficient resource allocation scheme for a heterogeneous shared stream big data analytics cluster shared by multiple topologies, in order to achieve max-min fairness in the utilities of the throughput for all the topologies. We first formulate a novel model resource allocation problem, which is a mixed 0-1 integer program. The NP-hardness of the problem is rigorously proven. To tackle this problem, we transform the non-convex constraint to several linear constraints using linearization and reformulation techniques. Based on the analysis of the problem-specific structure and characteristics, we propose an approach that iteratively solves the continuous problem with a fixed set of discrete variables optimally, and updates the discrete variables heuristically. Simulations show that our proposed resource allocation scheme remarkably improves the max-min fairness in utilities of the topology throughput, and is low in computational complexity.

Offloading computation tasks from resource-poor end devices to powerful backend clouds has become a prevalent solution thanks to the rapid development of cloud computing. However, modern Internet of Things applications, such as augmented reality and real-time monitoring, bring stringent delay requirements to the computation tasks in the deviceto-computing-facility communications. To better accommodate the delay requirements of the computation tasks, the recently proposed fog computing concept suggests that these computation tasks can be extensively offloaded to the distributed computation facilities along the cloud-to-things continuum. These computation facilities, including central clouds and the computation facilities standing at the network edge, jointly form an overlay network, named a fog network, to provide fog computing services for end devices. This paper targets a practical and efficient scheme to schedule tasks with heterogeneous delay sensitivities in a shared fog network. A mathematical model is first constructed to capture the major characteristics of a fog network. The model enforces lexicographic max-min fairness, an enhanced metric compared to conventional max-min fairness. The task offloading problem is modeled as an integer non-linear program. An efficient and exact solution method is proposed based on problem-specific analysis. Finally, synthesized-trace-driven simulations demonstrate the efficacy of our proposed offloading scheme.

Higher-order correlated excitonic states arise from the mutual interactions of excitons, which generally requires a significant exciton density and therefore high excitation levels. Here, we report the emergence of two biexcitons species, one neutral and one charged, in monolayer tungsten diselenide under moderate continuous-wave excitation. The efficient formation of biexcitons is facilitated by the long lifetime of the dark exciton state associated with a spin-forbidden transition, as well as improved sample quality from encapsulation between hexagonal boron nitride layers. From studies of the polarization and magnetic field dependence of the neutral biexciton, we conclude that this species is composed of a bright and a dark excitons residing in opposite valleys in momentum space. Our observations demonstrate that the distinctive features associated with biexciton states can be accessed at low light intensities and excitation densities.

Distributed stream big data analytics platforms have emerged to tackle the continuously generated data streams. In stream big data analytics, the data processing workflow is abstracted as a directed graph referred to as a topology. Data are read from the storage and processed tuple by tuple, and these processing results are updated dynamically. The performance of a topology is evaluated by its throughput. This paper proposes an efficient resource allocation scheme for a heterogeneous shared stream big data analytics cluster shared by multiple topologies, in order to achieve max-min fairness in the utilities of the throughput for all the topologies. We first formulate a novel model resource allocation problem, which is a mixed 0-1 integer program. The NP-hardness of the problem is rigorously proven. To tackle this problem, we transform the non-convex constraint to several linear constraints using linearization and reformulation techniques. Based on the analysis of the problem-specific structure and characteristics, we propose an approach that iteratively solves the continuous problem with a fixed set of discrete variables optimally, and updates the discrete variables heuristically. Simulations show that our proposed resource allocation scheme remarkably improves the max-min fairness in utilities of the topology throughput, and is low in computational complexity.

Offloading computation tasks from resource-poor end devices to powerful backend clouds has become a prevalent solution thanks to the rapid development of cloud computing. However, modern Internet of Things applications, such as augmented reality and real-time monitoring, bring stringent delay requirements to the computation tasks in the deviceto-computing-facility communications. To better accommodate the delay requirements of the computation tasks, the recently proposed fog computing concept suggests that these computation tasks can be extensively offloaded to the distributed computation facilities along the cloud-to-things continuum. These computation facilities, including central clouds and the computation facilities standing at the network edge, jointly form an overlay network, named a fog network, to provide fog computing services for end devices. This paper targets a practical and efficient scheme to schedule tasks with heterogeneous delay sensitivities in a shared fog network. A mathematical model is first constructed to capture the major characteristics of a fog network. The model enforces lexicographic max-min fairness, an enhanced metric compared to conventional max-min fairness. The task offloading problem is modeled as an integer non-linear program. An efficient and exact solution method is proposed based on problem-specific analysis. Finally, synthesized-trace-driven simulations demonstrate the efficacy of our proposed offloading scheme.

Higher-order correlated excitonic states arise from the mutual interactions of excitons, which generally requires a significant exciton density and therefore high excitation levels. Here, we report the emergence of two biexcitons species, one neutral and one charged, in monolayer tungsten diselenide under moderate continuous-wave excitation. The efficient formation of biexcitons is facilitated by the long lifetime of the dark exciton state associated with a spin-forbidden transition, as well as improved sample quality from encapsulation between hexagonal boron nitride layers. From studies of the polarization and magnetic field dependence of the neutral biexciton, we conclude that this species is composed of a bright and a dark excitons residing in opposite valleys in momentum space. Our observations demonstrate that the distinctive features associated with biexciton states can be accessed at low light intensities and excitation densities.

Distributed stream big data analytics platforms have emerged to tackle the continuously generated data streams. In stream big data analytics, the data processing workflow is abstracted as a directed graph referred to as a topology. Data are read from the storage and processed tuple by tuple, and these processing results are updated dynamically. The performance of a topology is evaluated by its throughput. This paper proposes an efficient resource allocation scheme for a heterogeneous shared stream big data analytics cluster shared by multiple topologies, in order to achieve max-min fairness in the utilities of the throughput for all the topologies. We first formulate a novel model resource allocation problem, which is a mixed 0-1 integer program. The NP-hardness of the problem is rigorously proven. To tackle this problem, we transform the non-convex constraint to several linear constraints using linearization and reformulation techniques. Based on the analysis of the problem-specific structure and characteristics, we propose an approach that iteratively solves the continuous problem with a fixed set of discrete variables optimally, and updates the discrete variables heuristically. Simulations show that our proposed resource allocation scheme remarkably improves the max-min fairness in utilities of the topology throughput, and is low in computational complexity.

Fog computing, complementary to cloud computing, has recently emerged as a new paradigm that extends the computing infrastructure from the center to the edge of the network. This article explores the design of a fog computing orchestration framework to support IoT applications. In particular, we focus on how the widely adopted cloud computing orchestration framework can be customized to fog computing systems. We first identify the major challenges in this procedure that arise due to the distinct features of fog computing. Then we discuss the necessary adaptations of the orchestration framework to accommodate these challenges.

Offloading computation tasks from resource-poor end devices to powerful backend clouds has become a prevalent solution thanks to the rapid development of cloud computing. However, modern Internet of Things applications, such as augmented reality and real-time monitoring, bring stringent delay requirements to the computation tasks in the deviceto-computing-facility communications. To better accommodate the delay requirements of the computation tasks, the recently proposed fog computing concept suggests that these computation tasks can be extensively offloaded to the distributed computation facilities along the cloud-to-things continuum. These computation facilities, including central clouds and the computation facilities standing at the network edge, jointly form an overlay network, named a fog network, to provide fog computing services for end devices. This paper targets a practical and efficient scheme to schedule tasks with heterogeneous delay sensitivities in a shared fog network. A mathematical model is first constructed to capture the major characteristics of a fog network. The model enforces lexicographic max-min fairness, an enhanced metric compared to conventional max-min fairness. The task offloading problem is modeled as an integer non-linear program. An efficient and exact solution method is proposed based on problem-specific analysis. Finally, synthesized-trace-driven simulations demonstrate the efficacy of our proposed offloading scheme.

Fog computing, complementary to cloud computing, has recently emerged as a new paradigm that extends the computing infrastructure from the center to the edge of the network. This article explores the design of a fog computing orchestration framework to support IoT applications. In particular, we focus on how the widely adopted cloud computing orchestration framework can be customized to fog computing systems. We first identify the major challenges in this procedure that arise due to the distinct features of fog computing. Then we discuss the necessary adaptations of the orchestration framework to accommodate these challenges.

Fog computing, complementary to cloud computing, has recently emerged as a new paradigm that extends the computing infrastructure from the center to the edge of the network. This article explores the design of a fog computing orchestration framework to support IoT applications. In particular, we focus on how the widely adopted cloud computing orchestration framework can be customized to fog computing systems. We first identify the major challenges in this procedure that arise due to the distinct features of fog computing. Then we discuss the necessary adaptations of the orchestration framework to accommodate these challenges.

This paper presents a procedural generation method that creates visually attractive levels for the Angry Birds game. Besides being an immensely popular mobile game, Angry Birds has recently become a test bed for various artificial intelligence technologies. We propose a new approach for procedurally generating Angry Birds levels using Chinese style and Japanese style building structures. A conducted experiment confirms the effectiveness of our approach with statistical significance.

The configuration of a Hadoop cluster is significantly important to its performance, because an improper configuration can greatly deteriorate the job execution performance. Unfortunately, systematic guidelines on how to configure a Hadoop cluster are still missing. In this paper, we undertake an empirical study on key operations and mechanisms of Hadoop job execution, including the task assignment strategy and speculative execution. Based on the experiments, we provide suggestions on the system configuration, particularly on the matching between the hardware resource partitioning scheme and the job splitting granularity.

The configuration of a Hadoop cluster is significantly important to its performance, because an improper configuration can greatly deteriorate the job execution performance. Unfortunately, systematic guidelines on how to configure a Hadoop cluster are still missing. In this paper, we undertake an empirical study on key operations and mechanisms of Hadoop job execution, including the task assignment strategy and speculative execution. Based on the experiments, we provide suggestions on the system configuration, particularly on the matching between the hardware resource partitioning scheme and the job splitting granularity.