Federated learning (FL) is a training technique that enables client devices to jointly learn a shared model by aggregating locally computed models without exposing their raw data. While most of the existing work focuses on improving the FL model accuracy, in this paper, we focus on the improving the training efficiency, which is often a hurdle for adopting FL in real world applications. Specifically, we design an efficient FL framework which jointly optimizes model accuracy, processing latency and communication efficiency, all of which are primary design considerations for real implementation of FL. Inspired by the recent success of Multi Agent Reinforcement Learning (MARL) in solving complex control problems, we present FedMarl, a federated learning framework that relies on trained MARL agents to perform efficient run-time client selection. Experiments show that FedMarl can significantly improve model accuracy with much lower processing latency and communication cost.

The emergence of the Internet of Things (IoT) has led to a remarkable increase in the volume of data generated at the network edge. In order to support real-time smart IoT applications, massive amounts of data generated from edge devices need to be processed using methods such as deep neural networks (DNNs) with low latency. To improve application performance and minimize resource cost, enterprises have begun to adopt Edge computing, a computation paradigm that advocates processing input data locally at the network edge. However, as edge nodes are often resource-constrained, running data-intensive DNN inference tasks on each individual edge node often incurs high latency, which seriously limits the practicality and effectiveness of this model.

Recent studies have shown that introducing communication between agents can significantly improve overall performance in cooperative Multi-agent reinforcement learning (MARL). However, existing communication schemes often require agents to exchange an excessive number of messages at run-time under a reliable communication channel, which hinders its practicality in many real-world situations. In this paper, we present \textit{Temporal Message Control} (TMC), a simple yet effective approach for achieving succinct and robust communication in MARL. TMC applies a temporal smoothing technique to drastically reduce the amount of information exchanged between agents. Experiments show that TMC can significantly reduce inter-agent communication overhead without impacting accuracy. Furthermore, TMC demonstrates much better robustness against transmission loss than existing approaches in lossy networking environments.

Multi-agent reinforcement learning (MARL) has recently received considerable attention due to its applicability to a wide range of real-world applications. However, achieving efficient communication among agents has always been an overarching problem in MARL. In this work, we propose Variance Based Control (VBC), a simple yet efficient technique to improve communication efficiency in MARL. By limiting the variance of the exchanged messages between agents during the training phase, the noisy component in the messages can be eliminated effectively, while the useful part can be preserved and utilized by the agents for better performance. Our evaluation using multiple MARL benchmarks indicates that our method achieves $2-10\times$ lower in communication overhead than state-of-the-art MARL algorithms, while allowing agents to achieve better overall performance.

Cloud data centers today use visualization technologies to facilitate allocation of physical resources to multiple applications. As cloud data centers continue to grow in scale and complexity, effectively monitoring and identifying system anomalies is becoming a critical problem. Furthermore, due to complex dependencies between system components in a virtualized data center, a single cause of anomaly can typically trigger multiple alarms. Therefore there is also a need to efficiently analyze and identify the causes of the anomalies in a scalable and effective manner, in order to reduce the overhead of diagnosis and troubleshooting performed by the cloud operator. Motivated by these observations, we present a mechanism for automatic anomaly detection and root cause analysis in virtualized cloud data centers. We first use unsupervised learning techniques to identify abnormal system behaviors, and then propose a technique for root cause analysis with consideration to anomaly propagation among system components. Using a real virtualized cloud testbed, we show that our mechanism efficiently identifies system anomalies and accurately determines their causes.

MapReduce has become a prevalent programming model for building data processing applications in the cloud. While being widely used, existing MapReduce schedulers still suffer from an issue known as partitioning skew, where the output of map tasks is unevenly distributed among reduce tasks. Existing solutions follow a similar principle that repartitions workload among reduce tasks. However, those approaches often incur high performance overhead due to the partition size prediction and repartitioning. In this paper, we present DREAMS, a framework that provides run-time partitioning skew mitigation. Instead of repartitioning workload among reduce tasks, we cope with the partitioning skew problem by controlling the amount of resources allocated to each reduce task. Our approach completely eliminates the repartitioning overhead, yet is simple to implement. Experiments using both real and synthetic workloads running on a 21-node Hadoop cluster demonstrate that DREAMS can effectively mitigate the negative impact of partitioning skew, thereby improving the job completion time by up to a factor of$2.29$over the native Hadoop YARN. Compared to the state-of-the-art solution, DREAMS can improve the job completion time by a factor of$1.65$.

Many multicast services such as live multimedia distribution and real-time event monitoring require constructing a multicast mechanism that involves network functions (e.g. firewall, video transcoding). Network Function Virtualization (NFV) is a concept that proposes using virtualization to implement network functions on infrastructure building block (such as high volume servers, virtual machines), where software provides the functionality of existing purpose-built network equipment. We present an approach for building the multicast mechanism whereby multicast flows are processed by NFV before reaching their end users. We propose a routing algorithm and a method for building an appropriate multicast topology.

Distributed file systems such as Google File System and Hadoop Distributed File System have been used to store large volumes of data in Cloud data centers. These systems divide data sets in blocks of fixed size and replicate them over multiple machines to achieve both reliability and efficiency. Recent studies have shown that data blocks tend to have a wide disparity in data popularity. In this context, the naive block replication schemes used by these systems often cause an uneven load distribution across machines, which reduces the overall I/O throughput of the system. While many replication algorithms have been proposed, existing solutions have not carefully studied the placement of data blocks that balances the load across machines, while ensuring node and rack-level reliability requirements are satisfied. In this paper, we study the dynamic data replication problem with the goal of balancing machine load while ensuring machine and rack-level reliability requirements are met. We propose several local search algorithms that provide constant approximation guarantees, yet simple and practical for implementation. We further present Aurora, a dynamic block placement mechanism that implements these algorithms in the Hadoop Distributed File System with minimal overhead. Through experiments using workload traces from Yahoo! and Facebook, we show Aurora reduces machine load imbalance by up to 26.9% compared to existing solutions, while satisfying node and rack-level reliability requirements.

MapReduce has become a popular model for large-scale data processing in recent years. However, existing MapRe-duce schedulers still suffer from an issue known as partitioning skew, where the output of map tasks is unevenly distributed among reduce tasks. In this paper, we present DREAMS, a framework that provides run-time partitioning skew mitigation. Unlike previous approaches that try to balance the workload of reducers by repartitioning the intermediate data assigned to each reduce task, in DREAMS we cope with partitioning skew by adjusting task run-time resource allocation. We show that our approach allows DREAMS to eliminate the overhead of data repartitioning. Through experiments using both real and synthetic workloads running on a 11-node virtual virtualised Hadoop cluster, we show that DREAMS can effectively mitigate negative impact of partitioning skew, thereby improving job performance by up to 20.3%.

Real-time content delivery services such as live media streaming, news casting and real-time event subscription/publication systems have become popular in recent years. Unlike traditional content delivery applications, real-time content delivery requires live content to be processed and delivered to end users in a timely and efficient manner. Furthermore, as both content producers and consumers may change over-time, it is a challenge to provision resources for these applications to achieve high service quality while minimizing total operational costs. Fortunately, the recent development of Cloud computing and Software Defined Networking (SDN) enables efficient implementation of real-time content delivery systems. Recently, the concept of Software Defined Infrastructure aims at combining Cloud computing and SDN to provide an unified framework for application deployment and management. In this paper, we present Kaleidoscope, an architecture for real-time content delivery in Software Defined Infrastructures. Kaleidoscope leverages network virtualization, SDN-based broadcasting and dynamic cloud resource provisioning to achieve high resource efficiency and service performance. Specifically, we present a resource management scheme that controls Cloud resource allocation and network configuration at run-time in accordance with service demand. Experiments show that Kaleidoscope is able to achieve lower resource cost while providing high service quality.

MapReduce has become a popular model for data-intensive computation in recent years. By breaking down each job into small map and reduce tasks and executing them in parallel across a large number of machines, MapReduce can significantly reduce the running time of data-intensive jobs. However, despite recent efforts toward designing resource-efficient MapReduce schedulers, existing solutions that focus on scheduling at the task-level still offer sub-optimal job performance. This is because tasks can have highly varying resource requirements during their lifetime, which makes it difficult for task-level schedulers to effectively utilize available resources to reduce job execution time. To address this limitation, we introduce PRISM, a fine-grained resource-aware MapReduce scheduler that divides tasks into phases, where each phase has a constant resource usage profile, and performs scheduling at the phase level. We first demonstrate the importance of phase-level scheduling by showing the resource usage variability within the lifetime of a task using a wide-range of MapReduce jobs. We then present a phase-level scheduling algorithm that improves execution parallelism and resource utilization without introducing stragglers. In a 10-node Hadoop cluster running standard benchmarks, PRISM offers high resource utilization and provides 1.3× improvement in job running time compared to the current Hadoop schedulers.

Many multicast services such as live multimedia distribution and real-time event monitoring require multicast mechanisms that involve network functions (e.g., firewall and video transcoding). Network function virtualization (NFV) is a concept that proposes using virtualization to implement network functions on infrastructure building block (such as high volume servers and virtual machines), where software provides the functionality of existing purpose-built network equipment. We present an approach for building the multicast mechanism whereby multicast flows are processed by NFV before reaching their end users. We propose a routing algorithm and a method for building an appropriate multicast topology.

Cloud computing has become a cost-effective model for deploying online services in recent years. To improve the Quality-of-Service (QoS) of the provisioned services, recently a number of proposals have advocated to provision both guaranteed server and network resources in the form of Virtual Data Centers (VDCs). However, existing VDC scheduling algorithms have not fully considered the reliability aspect of the allocations in terms of (1) hardware failure characteristics on which the service is hosted, and (2) the impact of individual failures on service availability, given the dependencies among the virtual components. To address this limitation, in this paper we present a technique for computing VDC availability that considers heterogeneous hardware failure rates and dependencies among virtual components. We then propose Venice, an availability-aware VDC embedding framework for achieving high VDC availability and low operational costs. Experiments show Venice can significantly improve VDC availability while achieving higher income compared to availability-oblivious solutions.

With the proliferation of cloud computing, virtu-alization has become the cornerstone of modern data centers and an effective solution to reduce operational costs, maximize utilization and improve performance and reliability. One of the powerful features provided by virtualization is Virtual Machine (VM) migration, which facilitates moving workloads within the infrastructure to reach various performance objectives. As recent virtual resource management schemes are more reliant on this feature, a large number of VM migrations may be triggered simultaneously to optimize resource allocations. In this context, a challenging problem is to find an efficient migration plan, i.e., an optimal sequence in which migrations should be triggered in order to minimize the total migration time and impact on services. In this paper, we propose CQNCR (read as sequencer), an effective technique for determining the execution order of massive VM migrations within data centers. Specifically, given an initial and a target resource configuration, CQNCR sequences VM migrations so as to efficiently reach the final configuration with minimal time and impact on performance. Experiments show that CQNCR can significantly reduce total migration time by up to 35% and service downtime by up to 60%.

Cloud computing is a paradigm that harnesses massive resource capacity of data centers to support applications in a scalable, flexible, reliable and cost-effective manner. Despite its recent success and rapid adoption in the IT industry, recent literature has shown that effective workload management in cloud computing environments remains to be a difficult challenge. A key reason behind this difficulty is that resources and workloads in production environments are both heterogeneous and dynamic. In particular, large cloud data centers often consist of machines with heterogeneous capacities and performance characteristics. At the same time, cloud workloads often show significant diversity in terms of priority, resource requirements, arrival rate and performance objectives. Consequently, it is difficult to devise heterogeneity and dynamicity-aware management scheme that satisfy diverse application performance objectives, while reducing operational expenses such as energy consumption. This work addresses several key challenges pertaining to dynamic workload management in heterogenous Cloud environments. Specifically, we first present a scheme that place service application across geographically distributed data centers to meet service demand while minimizing total resource usage cost. Then, we design a heterogeneity-aware dynamic application provisioning technique to minimize energy consumption while satisfying performance objectives. Finally, we study the problem of MapReduce scheduling and present a novel scheme that leverages heterogenous run-time task usage characteristics. Through experiments and simulations, we show our proposed solutions can significantly reduce data center energy consumption, while achieving better application performance in terms of service response time and job completion time.

Data centers consume tremendous amounts of energy in terms of power distribution and cooling. Dynamic capacity provisioning is a promising approach for reducing energy consumption by dynamically adjusting the number of active machines to match resource demands. However, despite extensive studies of the problem, existing solutions have not fully considered the heterogeneity of both workload and machine hardware found in production environments. In particular, production data centers often comprise heterogeneous machines with different capacities and energy consumption characteristics. Meanwhile, the production cloud workloads typically consist of diverse applications with different priorities, performance and resource requirements. Failure to consider the heterogeneity of both machines and workloads will lead to both sub-optimal energy-savings and long scheduling delays, due to incompatibility between workload requirements and the resources offered by the provisioned machines. To address this limitation, we present Harmony, a Heterogeneity-Aware dynamic capacity provisioning scheme for cloud data centers. Specifically, we first use the K-means clustering algorithm to divide workload into distinct task classes with similar characteristics in terms of resource and performance requirements. Then we present a technique that dynamically adjusting the number of machines to minimize total energy consumption and scheduling delay. Simulations using traces from a Google's compute cluster demonstrate Harmony can reduce energy by 28 percent compared to heterogeneity-oblivious solutions.

Software Defined Networking (SDN) has emerged as a new paradigm that offers the programmability required to dynamically configure and control a network. A traditional SDN implementation relies on a logically centralized controller that runs the control plane. However, in a large-scale WAN deployment, this rudimentary centralized approach has several limitations related to performance and scalability. To address these issues, recent proposals have advocated deploying multiple controllers that work cooperatively to control a network. Nonetheless, this approach drags in an interesting problem, which we call the Dynamic Controller Provisioning Problem (DCPP). DCPP dynamically adapts the number of controllers and their locations with changing network conditions, in order to minimize flow setup time and communication overhead. In this paper, we propose a framework for deploying multiple controllers within an WAN. Our framework dynamically adjusts the number of active controllers and delegates each controller with a subset of Openflow switches according to network dynamics while ensuring minimal flow setup time and communication overhead. To this end, we formulate the optimal controller provisioning problem as an Integer Linear Program (ILP) and propose two heuristics to solve it. Simulation results show that our solution minimizes flow setup time while incurring very low communication overhead.

Data centers today consume tremendous amount of energy in terms of power distribution and cooling. Dynamic capacity provisioning is a promising approach for reducing energy consumption by dynamically adjusting the number of active machines to match resource demands. However, despite extensive studies of the problem, existing solutions for dynamic capacity provisioning have not fully considered the heterogeneity of both workload and machine hardware found in production environments. In particular, production data centers often comprise several generations of machines with different capacities, capabilities and energy consumption characteristics. Meanwhile, the workloads running in these data centers typically consist of a wide variety of applications with different priorities, performance objectives and resource requirements. Failure to consider heterogenous characteristics will lead to both sub-optimal energy-savings and long scheduling delays, due to incompatibility between workload requirements and the resources offered by the provisioned machines. To address this limitation, in this paper we present HARMONY, a Heterogeneity-Aware Resource Management System for dynamic capacity provisioning in cloud computing environments. Specifically, we first use the K-means clustering algorithm to divide the workload into distinct task classes with similar characteristics in terms of resource and performance requirements. Then we present a novel technique for dynamically adjusting the number of machines of each type to minimize total energy consumption and performance penalty in terms of scheduling delay. Through simulations using real traces from Google's compute clusters, we found that our approach can improve data center energy efficiency by up to 28% compared to heterogeneity-oblivious solutions.

Large-scale online service providers have been increasingly relying on geographically distributed cloud infrastructures for service hosting and delivery. In this context, a key challenge faced by service providers is to determine the locations where service applications should be placed such that the hosting cost is minimized while key performance requirements (e.g., response time) are ensured. Furthermore, the dynamic nature of both demand pattern and infrastructure cost favors a dynamic solution to this problem. Currently most of the existing solutions for service placement have either ignored dynamics, or provided solutions inadequate to achieve this objective. In this paper, we present a framework for dynamic service placement problems based on control- and game-theoretic models. In particular, we present a solution that optimizes the hosting cost dynamically over time according to both demand and resource price fluctuations. We further consider the case where multiple service providers compete for resources in a dynamic manner. This paper extends our previous work [1] by analyzing the outcome of the competition in terms of both price of stability and price of anarchy. Our analysis suggests that in an uncoordinated scenario where service providers behave in a selfish manner, the resulting Nash equilibrium can be arbitrarily worse than the optimal centralized solution in terms of social welfare. Based on this observation, we present a coordination mechanism that can be employed by the infrastructure provider to maximize the social welfare of the system. Finally, we demonstrate the effectiveness of our solutions using realistic simulations.

Large-scale online service providers have been increasingly relying on geographically distributed cloud infrastructures for service hosting and delivery. In this context, a key challenge faced by service providers is to determine the locations where service applications should be placed such that the hosting cost is minimized while key performance requirements (e.g. response time) are assured. Furthermore, the dynamic nature of both demand pattern and infrastructure cost favors a dynamic solution to this problem. Currently most of the existing solutions for service placement have either ignored dynamics, or provided inadequate solutions that achieve both objectives at the same time. In this paper, we present a framework for dynamic service placement problems based on control- and game-theoretic models. In particular, we present a solution that optimizes the desired objective dynamically over time according to both demand and resource price fluctuations. We further consider the case where multiple service providers compete for resource in a dynamic manner, and show that there is a Nash equilibrium solution which is socially optimal. Using simulations based on realistic topologies, demand and resource prices, we demonstrate the effectiveness of our solution in realistic settings.