Federated learning (FL) is a promising paradigm that enables collaboratively learning a shared model across massive clients while keeping the training data locally. However, for many existing FL systems, clients need to frequently exchange model parameters of large data size with the remote cloud server directly via wide-area networks (WAN), leading to significant communication overhead and long transmission time. To mitigate the communication bottleneck, we resort to the hierarchical federated learning paradigm of HiFL, which reaps the benefits of mobile edge computing and combines synchronous client-edge model aggregation and asynchronous edge-cloud model aggregation together to greatly reduce the traffic volumes of WAN transmissions. Specifically, we first analyze the convergence bound of HiFL theoretically and identify the key controllable factors for model performance improvement. We then advocate an enhanced design of HiFlash by innovatively integrating deep reinforcement learning based adaptive staleness control and heterogeneity-aware client-edge association strategy to boost the system efficiency and mitigate the staleness effect without compromising model accuracy. Extensive experiments corroborate the superior performance of HiFlash in model accuracy, communication reduction, and system efficiency.

The accelerating convergence of artificial intelligence and edge computing has sparked a recent wave of interest in edge intelligence. While pilot efforts focused on edge DNN inference serving for a single user or DNN application, scaling edge DNN inference serving to multiple users and applications is however nontrivial. In this paper, we propose an online optimization framework EdgeAdaptor for multi-user and multi-application edge DNN inference serving at scale, which aims to navigate the three-way trade-off between inference accuracy, latency, and resource cost via jointly optimizing the application configuration adaption, DNN model selection and edge resource provisioning on-the-fly. The underlying long-term optimization problem is difficult since it is NP-hard and involves future uncertain information. To address these dual challenges, we fuse the power of online optimization and approximate optimization into a joint optimization framework, via i) decomposing the long-term problem into a series of single-shot fractional problems with a regularization technique, and ii) rounding the fractional solution to a near-optimal integral solution with a randomized dependent scheme. Rigorous theoretical analysis derives a parameterized competition ratio of our online algorithms, and extensive trace-driven simulations verify that its empirical value is no larger than 1.4 in typical scenarios.

Auction design has recently been studied for dynamic resource bundling and virtual machine (VM) provisioning in IaaS clouds, but is mostly restricted to one-shot or offline setting. This paper targets a more realistic case of online VM auction design, where: 1) cloud users bid for resources into the future to assemble customized VMs with desired occupation durations, possibly located in different data centers; 2) the cloud provider dynamically packs multiple types of resources on heterogeneous physical machines (servers) into the requested VMs; 3) the operational costs of servers are considered in resource allocation; and 4) both social welfare and the cloud provider's net profit are to be maximized over the system running span. We design truthful, polynomial time auctions to achieve social welfare maximization and/or the provider's profit maximization with good competitive ratios. Our mechanisms consist of two main modules: 1) an online primal-dual optimization framework for VM allocation to maximize the social welfare with server costs, and for revealing the payments through the dual variables to guarantee truthfulness and 2) a randomized reduction algorithm to convert the social welfare maximizing auctions to ones that provide a maximal expected profit for the provider, with competitive ratios comparable to those for social welfare. We adopt a new application of Fenchel duality in our primal-dual framework, which provides richer structures for convex programs than the commonly used Lagrangian duality, and our optimization framework is general and expressive enough to handle various convex server cost functions. The efficacy of the online auctions is validated through careful theoretical analysis and trace-driven simulation studies.

Machine learning (ML) tasks are one of the major workloads in today's edge computing networks. Existing edge-cloud schedulers allocate the requested amounts of resources to each task, falling short of best utilizing the limited edge resources for ML tasks. This paper proposes TapFinger , a distributed scheduler for edge clusters that minimizes the total completion time of ML tasks through co-optimizing task placement and fine-grained multi-resource allocation. To learn the tasks’ uncertain resource sensitivity and enable distributed scheduling, we adopt multi-agent reinforcement learning (MARL) and propose several techniques to make it efficient, including a heterogeneous graph attention network as the MARL backbone, a tailored task selection phase in the actor network, and the integration of Bayes’ theorem and masking schemes. We first implement a single-task scheduling version, which schedules at most one task each time. Then we generalize to the multi-task scheduling case, in which a sequence of tasks is scheduled simultaneously. Our design can mitigate the expanded decision space and yield fast convergence to optimal scheduling solutions. Extensive experiments using synthetic and test-bed ML task traces show that TapFinger can achieve up to 54.9% reduction in the average task completion time and improve resource efficiency as compared to state-of-the-art schedulers.

As mobile edge computing (MEC) emerges as a paradigm to meet the ever-increasing computation demands from real-time Internet of Things (IoT) applications in 5 G era, the development trends of which are mainly divided into two, with one being MEC with advanced computing architectures, and the other being MEC with high efficiency for sustainable operations. We are committed to taking advantage of these two trends to explore a novel multi-tier edge computing scenario with hierarchical task offloading and green energy provisioning via leveraging the energy harvesting (EH) technique. Specifically, we focus on the key problem of joint task offloading and energy scheduling in such green multi-tier edge computing systems. We aim to minimize the task execution cost by jointly considering the system cost that covers latency, energy consumption, and cloud rental fees. By formulating the problem as a stochastic optimization problem, we invoke the Lyapunov technique to decompose the long-term optimization problem into a series of one-slot optimization problems which only use the current system information. To solve the one-slot optimization problem which is a mixed-integer linear problem (MILP) proved to be NP-hard , we first relax the integer variables into real ones to obtain the optimal fractional solutions. Considering the capacity of the physical resources of each edge server, we propose a resource-constrained randomized dependent rounding algorithm to properly round up or down the fractional variables to get a feasible yet near-optimal solution. We conduct rigorous theoretical analysis and extensive simulations to verify the superior performance of the proposed schemes.

Provisioning dynamic machine learning (ML) inference as a service for artificial intelligence (AI) applications of edge devices faces many challenges, including the tradeoff among accuracy loss, carbon emission, and unknown future costs. Besides, many governments are launching carbon emission rights (CERs) for operators to reduce carbon emissions further to reverse climate change. Facing these challenges, to achieve carbon-aware ML task offloading under limited CERs thus to achieve green edge AI, we establish a joint ML task offloading and CER purchasing problem, intending to minimize the accuracy loss under the long-term time-averaged cost budget of purchasing the required CER. However, considering the uncertainty of the resource prices, the CER purchasing prices, the carbon intensity of sites, and ML tasks’ arrivals, it is hard to decide the optimal policy online over a long-running period time. To overcome this difficulty, we leverage the two-timescale Lyapunov optimization technique, of which the $T$ -slot drift-plus-penalty methodology inspires us to propose an online algorithm that purchases CER in multiple timescales (on-preserved in carbon future market and on-demanded in the carbon spot market) and makes decisions about where to offload ML tasks. Considering the NP-hardness of the $T$ -slot problems, we further propose the resource-restricted randomized dependent rounding algorithm to help to gain the near-optimal solution with no help of any future information. Our theoretical analysis and extensive simulation results driven by the real carbon intensity trace show the superior performance of the proposed algorithms.