Programmable switches have revolutionized networking by enabling a new spectrum of applications, such as network telemetry, in-network computation, and machine learning. These applications heavily utilize register memory but their performance is significantly constrained by the scarcity of on-chip resources, such as the 15 MB of SRAM available on a Tofino switch. To effectively accommodate increasingly memorydemanding applications, we aim to pool register resources across multiple switches, creating a larger unified register memory space. This resource pooling approach addresses the limitations of existing single-switch Virtual Register Memory (VRM) solutions, which cannot meet the demands of these applications in distributed environments. To achieve this, we propose dVRM, a distributed VRM deployment framework that enables crossswitch memory sharing and self-dynamic memory allocation on the data plane. dVRM introduces three innovations: (1) Grouped Multi-Switch Registers (GMRs), virtualizing distributed pipeline stages into a unified memory pool; (2) a self-adaptive, bit-width allocation mechanism driven by real-time data-plane feedback; and (3) lightweight heuristics for concurrent application deployment with distributed VRM, formulated as a mixed-integer linear programming (MILP) problem. We have implemented dVRM on both P4 hardware switches (with Intel Tofino ASIC) and BMv2. Experimental results show that dVRM significantly reduces hash unit consumption by up to 26% and achieves an improvement in accuracy (ARE) of up to 57.3% across various workloads.

Communication overhead is a critical challenge in federated learning, particularly in bandwidth-constrained networks. Although many methods have been proposed to reduce communication overhead, most focus solely on compressing individual gradients, overlooking the temporal correlations among them. Prior studies have shown that gradients exhibit spatial correlations, typically reflected in low-rank structures. Through empirical analysis, we further observe a strong temporal correlation between client gradients across adjacent rounds. Based on these observations, we propose GradESTC, a compression technique that exploits both spatial and temporal gradient correlations. GradESTC exploits spatial correlations to decompose each full gradient into a compact set of basis vectors and corresponding combination coefficients. By exploiting temporal correlations, only a small portion of the basis vectorsneed to be dynamically updated in each round. GradESTC significantly reduces communication overhead by transmitting lightweight combination coefficients and a limited number of updated basis vectors instead of the full gradients. Extensive experiments show that, upon reaching a target accuracy level near convergence, GradESTC reduces uplink communication by an average of 39.79% compared to the strongest baseline, while maintaining comparable convergence speed and final accuracy to uncompressed FedAvg. By effectively leveraging spatio-temporal gradient structures, GradESTC offers a practical and scalable solution for communication-efficient federated learning.

Bit-alterable flash-based SSDs, offering page-level erase operation, allows individual flash pages in a block to be erased independently. The page-level erase operation alleviates the overhead of page migration during garbage collection and improves the SSD lifetime. However, when the number of invalid pages within a block exceeds a certain threshold, the latency of page-level garbage collections using page-level erase may exceed that of block-level garbage collections. In bit-alterable flash memory, existing garbage collection strategies dynamically choose between page-level and block-level garbage collections based on their latency. This often fails to fully exploit the advantage of page-level garbage collection in reducing write amplification under low-load conditions.To address this limitation, we propose an adaptive garbage collection strategy called AGCB to dynamically adjust garbage collection operations by the runtime workload of flash channels, thereby enhancing SSD performance and lifetime. Specifically, AGCB classifies flash channels as busy or idle by monitoring the depth of the transaction queue in cache. According to this classification, AGCB selectively applies page-level or block-level garbage collection operations, aiming to minimize the impact of garbage collections with host I/O requests. Meanwhile, we introduce a staged victim block selection scheme to further improve garbage collection efficiency and wear leveling. The experimental results unveil that compared with the existing schemes, AGCB reduces the number of garbage collection operations, average response time, and blocked user requests by an average of 14.6%, 14.7%, and 17.3%, respectively.

In the realm of Mobile Edge Computing (MEC) networks, mobile interactive applications, such as multiplayer online games, federated learning, and interactive multi-view video, are becoming increasingly popular. The embedding of Bidirectional Service Function Chains (BSFCs) for these applications has been studied. However, existing BSFC embedding research only considers static users, and the prevailing Virtual Network Functions (VNFs) placement methods do not account for the impact of individual node resources on the path, thus failing to maximize node resource utilization at the network level. Considering the aforementioned issues, we introduce a framework for BSFC embedding tailored for mobile users, named Path as a Point (PAAP). This framework integrates the resources of the paths and then globally considers the impact of the computing resources of edge nodes on the paths connecting them, maximizing edge node resource utilization across the entire network. We propose a three-phase algorithm within the PAAP framework. In the first phase, we introduce a singleuser algorithm for VNFs deployment during BSFC embedding. In the second phase, this optimization is extended to multiple users by establishing inter-user association rules. In the third phase, candidate application placement positions are evaluated across the entire network, culminating in the determination of optimal placement strategies. Empirical evaluations confirm the effectiveness of the proposed algorithm, demonstrating significant performance improvements over baseline methods.

Mobile Edge Computing (MEC) is an emerging network architecture designed to enhance Quality of Service (QoS) by bringing computational resources and storage closer to end users. In MEC environments, strategic placement of edge servers is crucial to minimize costs and optimize QoS. The Edge Server Placement (ESP) problem has been effectively addressed by using Multi-Armed Bandit (MAB) algorithms, known for their efficiency and adaptability, with the Upper Confidence Bound (UCB) algorithm being particularly prominent due to its stable performance and low dependency on parameters. However, UCB suffers excessive exploration and high uncertainty in reward estimation during its initial phase, leading to slow convergence. To overcome these challenges, we propose a novel Preference-based UCB (P-UCB) algorithm, which integrates the preference function into the UCB framework, drawing inspiration from the Gradient Bandit (GB) method. This modification not only accelerates convergence, but also improves overall efficiency. Furthermore, to address the Base Station Allocation (BSA) issue within the ESP context, we introduce a Weighted Base Station Allocation (WBSA) algorithm, which helps better manage access delay and workload balance. The P-UCB algorithm is evaluated through a comprehensive metric that based on access delay and workload balance, showing significant improvements over existing methods such as Multiple Choice (MC)-UCB, Q- Particle Swarm Optimization (QPSO), and Genetic Algorithm (GA). Experimental results on a real-world dataset show that P-UCB achieves a notable performance increase of at least 12.9%, effectively optimizing access delay and workload balance across various experimental settings, including the number of base stations, edge servers, and other relevant system parameters.

Virtual machine (VM) consolidation strategies are widely deployed in Cloud Data Centers (CDCs) to optimize resource utilization and improve the Quality of Service (QoS). However, the host overload detection algorithms in current VM consolidation strategies are static. That means, once the overload threshold is calculated, it will not change until the next recalculation. The current algorithms are not suitable for the environment of highly dynamic workloads which results in additional energy consumption and potential Service Level Agreement Violations (SLAVs) which will affect the QoS of CDC. In PMPHD, a novel host dynamic threshold adjustment algorithm is proposed. In the proposed algorithm, the PMs are classified into mildly overloaded, normal, and severely overloaded based on the resource utilization. If the PM is predicted to be severely overloaded in the next moment, the threshold of this PM will be proactively reduced. The PM is determined to be overloaded, and some VMs in this PM will be migrated in advance. Thus, this PM will be in normal in the next moment, and the VM performance degradation resulting from SLAV and VM migration overlap in the next moment will be avoided. If the PM is predicted to be mildly overloaded, the threshold will be appropriately increased to transit it to be in normal state in the next moment, and the VM in the PM will not be migrated. Since the PMs’ workloads are dynamic, the PMPHD overload algorithm predicts the resource utilization rate of PM continuously, and adjusts the overload threshold of PM. Compared with other algorithms, PMPHD maintains high efficiency while having lower ESV (a combination metric for balancing energy consumption and SLAV).

In-band network telemetry (INT) enables real-time network monitoring by embedding telemetry data into packets. The advent of programmable switches further enhances the flexibility of INT by enabling dynamic customization of telemetry collection at the hardware level. However, the practical application of INT is hampered by significant challenges arising from data noise (due to packet loss, delay, and measurement inaccuracies) and network dynamics (such as changing INT paths and feature requirements). These issues severely degrade the performance of machine learning models used for analyzing INT data, hindering the accurate capture of spatio-temporal network characteristics. This paper presents Planner, a novel generative graph learning framework designed to address these limitations. Planner enables the collection of network features at various levels of granularity on programmable switches. Crucially, it constructs dynamic graphs representing evolving INT paths and employs a hybrid Graph Neural Network (GNN) and Recurrent Neural Network (RNN) architecture to effectively learn spatial and temporal dependencies. Furthermore, Planner incorporates variational inference to generate robust latent representations, mitigating the detrimental effects of noise and instability in INT data. We have implemented a testbed prototype of Planner using Intel Tofino ASIC switches. Extensive experiments demonstrate the performance superiority and robustness of Planner over the baseline methods, achieving a 23.2% improvement in F1 score.

As an effective alternative to deep neural networks, broad learning system (BLS) has attracted more attention due to its efficient and outstanding performance and shorter training process in classification and regression tasks. Nevertheless, the performance of BLS will not continue to increase, but even decrease, as the number of nodes reaches the saturation point and continues to increase. In addition, the previous research on neural networks usually ignored the reason for the good generalization of neural networks. To solve these problems, this article first proposes the Extreme Fuzzy BLS (E-FBLS), a novel cascaded fuzzy BLS, in which multiple fuzzy BLS blocks are grouped or cascaded together. Moreover, the original data is input to each FBLS block rather than the previous blocks. In addition, we use residual learning to illustrate the effectiveness of E-FBLS. From the frequency domain perspective, we also discover the existence of the frequency principle in E-FBLS, which can provide good interpretability for the generalization of the neural network. Experimental results on classical classification and regression datasets show that the accuracy of the proposed E-FBLS is superior to traditional BLS in handling classification and regression tasks. The accuracy improves when the number of blocks increases to some extent. Moreover, we verify the frequency principle of E-FBLS that E-FBLS can obtain the low-frequency components quickly, while the high-frequency components are gradually adjusted as the number of FBLS blocks increases.

Serverless computing has been widely regarded as an ideal computing paradigm, enabling edge servers to host serverless functions. Due to its high scalability and usage-based pricing model, it provides efficient services across various applications. However, in the deployment process of serverless applications, past works lack considerations for the parallel relationships between stateful functions, which increases end to end latency. To leverage the parallel dependencies between functions, we propose a strategy for dependent function parallelization deployment, named LPCD (Low latency Parallel Candidate Deployment strategy). By partitioning the problem into inter-layer function deployment and analyzing optimal substructures, a heuristic algorithm is introduced to determine candidate deployment strategies for each layer of the users, which aims at identifying the optimal edge server for each function instance during deployment to enhance user satisfaction. Through simulation experiments, we evaluate the performance of the strategy. The experiments results indicate that the average latency was reduced by at least 41% compared with the state-of-the-art strategies.

Programmable data plane (PDP) has emerged as a powerful platform for line-rate packet processing, utilizing on-chip register memory to execute stateful applications. Yet most existing efforts concentrate on static approaches for allocating register memory, necessitating switch restarting and service interruption. Despite the availability of research on sharing memory for concurrent applications, the rigid requirement of limiting memory sharing to the same pipeline stages hampers application flexibility and poses scalability challenges. To address this limitation, we presentFlxVRM,a flexible register memory virtualization layerfor data plane P4 programs which supports high-flexibility sharing of register memory for concurrent applications on PDP.FlxVRMenables memory allocation at any stage and location of the pipeline on PDP for each application at run time. To reduce resource usage during virtualization in the data plane pipeline,FlxVRMfurther merges different tables and actions with similar structures within P4 programs. Additionally,FlxVRMprovides a compiler to generate data plane programs for virtualization as well as the control plane API configuration. A prototype ofFlxVRMis implemented based on P4 hardware switches with Intel Tofino ASIC. Our experiment results show thatFlxVRMsignificantly improves the allocatable memory space for applications by up to 50%, while reducing the resource of the table up to 68%.

Virtual machine (VM) consolidation strategies are widely used in cloud data centers (CDC) to optimize resource utilization and reduce total energy consumption. Although existing strategies consider current and future resource utilization, the impact of sudden bursts in historical resource utilization on the hosts has been underestimated in uncertain future periods. Insufficient analysis of historical resource utilization may increase the risk of host overloading and Service Level Agreement Violation (SLAV). By defining historical and future trends based on resource utilization, we propose a novel combined trend VM consolidation (CTVMC) strategy which can effectively reduce energy consumption and SLAV. The VMs with the largest combined trend are selected for migration to prevent host overloading. Based on the temporal locality and prediction technique, CTVMC then employs the past, present, and future resource utilization to filter candidate hosts, and identifies the most complementary host to place VM using combined trends. We conduct extensive simulation experiments with PlanetLab Trace and Google Cluster Trace in the CloudSim simulator. Compared with the well-known strategies, CTVMC strategy using the PlanetLab Trace can reduce the number of migrations by over 72.39%, SLAV by over 75.85%, and ESV (a combined metric that judges the trade-off between energy consumption and SLAV) by over 81.54%. According to the Google Cluster Trace, our strategy can reduce the number of migrations by over 61.51%, SLAV by over 37.37%, and ESV by over 35.30%.

Cloud computing has emerged as an infrastructure in the era of digital economy and has been widely applied in various fields. Virtual Machine(VM) management is the key mechanism in a Cloud Data Center(CDC). A typical VM management system is responsible for VM allocation and VM reallocation, and it is usually designed to optimize specific objectives, especially for the metrics of energy consumption, resource wastage, communication cost, and Service Level Agreement Violations (SLAV). However, it is greatly challenging to optimize these metrics at the same time, and most existing VM management strategies focus on optimizing part of the above four metrics. In this paper, we propose a Holistic Virtual Machine Management (HVMM) strategy to optimize the energy consumption, resource wastage, communication cost, and SLAV simultaneously. First, we define two parameters, the Compatibility and the Performance-to-Power Ratio (PPR), for VM allocation to optimize energy consumption and resource wastage. Then, we propose a reallocation approach based on spectral clustering that can handle dynamic traffic between VMs without a priori knowledge of the traffic between VMs, and it takes a slight expense of resource wastage and energy consumption to reduce communication cost between VMs and ensures low SLAV risk. To evaluate the performance of HVMM, we compared the proposed strategy with the state-of-the-art strategies in various experiments on real-world traces. Compared with the other strategies, the resource wastage of HVMM is reduced by 62%. Simultaneously, the communication cost is reduced by about 26%, the energy consumption is reduced by about 5%, and SLAV is much lower than that of the others.

Increasing workload conditions lead to a significant surge in power consumption and computing node failures in data centers. The existing workload distribution strategies focused on either thermal awareness or failure mitigation, overlooking the impact of node failures on the energy efficiency of cloud data centers. To address this issue, a new holistic model is built to characterize the impacts of workloads, computing and cooling costs, heat recirculation, and node failure on the energy efficiency of cloud data centers. Leveraging such a holistic model, we propose a novel thermal-aware workload distribution strategy calledHGSAthat takes node failure into accountand can improve the energy efficiency of cloud data centers. Our empirical findings confirm that (i) faulty nodes lead to a large rise in power consumption, and (ii) failure locations play a vital role in the power consumption of data centers. Experimental results unveil that HGSA is adroit at making near-optimal decisions in workload distribution strategies. In particular, HGSA cuts down the minimum inlet temperature by 5.2$\%$-15$\%$, improves the maximum air temperature of a Computer Room Air Conditioner (CRAC) model by 4.2$\%$-26.5$\%$, lowers the cooling cost by 15.4$\%$-50$\%$compared to the existing solutions. Furthermore, HGSA cuts back the total power consumption by 0.65$\%$-78$\%$.

Instantiated containers of an application are distributed across multiple Physical Machines (PMs) to achieve high parallel performance. Container placement plays a vital role in network traffic and the performance of containerized data centers. Existing container placement techniques are inadequate due to the ignorance of container traffic patterns. To solve this issue, we first investigate the network traffic between containers and observe that it exhibits a Zipf-like distribution. Motivated by this finding, we propose a novel container placement approach-Blender-by taking into account the Zipf-like distribution. Blender employs two algorithms calledRefineAlgandSplitAlgto divide containers of applications into blocks, and place these blocks across Virtual Machines (VMs). Blender exhibits two salient features: (i) it minimizes inter-block traffic by arranging the containers that communicate frequently in the same block. (ii) it achieves good load balancing by combining complementary blocks that request different resource types (e.g.,CPU-intensiveandmemory-intensiveblocks) and distributing these blocks across multiple VMs. The experimental results show that Blender significantly reduces communication traffic and network latency. In particular, Blender reduces the traffic of SBP and CA-WFD by 22% and 32%, respectively. Blender decreases network latency by 16% and 26% compared to SBP and CA-WFD. Furthermore, with Blender in place, the physical resources of hosting PMs are well balanced and utilized.

Many applications operating in the Internet of Things (IoT) require timely and fair data collection from devices. This has motivated research into a new metric called Age of Information (AoI). This paper contributes to this effort by proposing to minimize the maximum average AoI (min-max AoI) in a multi-hop IoT network comprising of solar-powered Power Beacons (PBs). It outlines a Mixed Integer Linear Program (MILP) that jointly optimizes: (i) the beamforming vector used by PBs to charge devices, and (ii) routing, which determines how samples from devices are forwarded to a sink node, and (iii) the sampling time of sources. It also presents two protocols: Centralized Linear Relaxation (CLR) and Distributed Path Selection (DPS), respectively. CLR is run by the sink to determine the transmit power of PBs and the path of each source using two Linear Programs (LPs). On the other hand, DPS is a distributed approach whereby PBs and sources make their own decisions using local information. Our simulation results show that min-max AoI increases with the number of sources, but reduces with increasing number of PBs. The number of paths available to a source, the number of frames, and solar panel size have limited impact on performance. The min-max AoI of CLR and DPS is$1.60\times $and$1.95\times $higher than that of MILP.

This article considers a novel Internet of Things network comprising of sensor devices and power beacons (PBs); both types of nodes are equipped with a cognitive radio (CR). In addition, these sensor devices are powered by radio-frequency signals from PBs. Our aim is to maximize the minimum rate of devices acting as sources. We outline the first mixed integer linear program (MILP) that jointly optimizes the channel assignment of PBs and devices, beamforming vector of PBs, data routing over multiple hops and link activation schedule of devices. We also design a distributed protocol called distributed max–min rate with CR (D-MRCR) for use by devices and PBs. Devices set their operation mode using local information and use a game theory-based approach to iteratively adjust their transmit power. On the other hand, each PB employs a linear program to determine its beamforming vector. Our results show that the max–min rate of D-MRCR is within 51.84% that of MILP.

As the volume of data keeps growing rapidly, more and more storage devices, servers and network devices are continuously added into data centers to store, manage and analyze the data. The industry experience indicates that, instead of a huge number of servers added at a time, the data center network also expands gradually by adding a small number of servers from time to time. As a result, how to achieve an incremental scalability is becoming a very important challenge in designing modern data center network architectures in order to maintain the topological properties unchanged when the size of data centers grows. In this paper, we propose a new type of data center network architecture named HSDC (High Scalability Data Center Network Architecture) based on the hypercube network. The HSDC is constructed by using $m$m-port switches and 2-port servers. The fault-tolerant routing algorithm designed in this paper for HSDC can be executed on any vertex and is able to construct a path between any pair of vertices. In order to achieve an incremental scalability, we further propose three types of incomplete HSDC structures that allow gradually adding servers into the structures, while maintaining all the topological properties. The simulation experiments and performance results demonstrate that the throughput of HSDC is comparable to that of Fat-Tree, BCube and DCell. Furthermore, the analysis results indicate that HSDC strikes a good balance among diameter, bisection width, incremental scalability, cost and energy consumption in contrast to the state-of-the-art data center network architectures.

Over the past decade, many data centers have been constructed around the world due to the explosive growth of data volume and type. The cost and energy consumption have become the most important challenges of building those data centers. Data centers today use commodity computers and switches instead of high-end servers and interconnections for cost-effectiveness. In this paper, we propose a new type of interconnection networks called Exchanged Cube-Connected Cycles (ExCCC). The ExCCC network is an extension of Exchanged Hypercube (EH) network by replacing each node with a cycle. The EHnetwork is based on link removal from a Hypercube network, which makes the EHnetwork more cost-effective as it scales up. After analyzing the topological properties of ExCCC, we employ commodity switches to construct a new class of data center network models, namely ExCCC-DCN, by leveraging the advantages of the ExCCC architecture. The analysis and experimental results demonstrate that the proposed ExCCC-DCN models significantly outperform four state-of-the-art data center network models in terms of the total cost, power consumption, scalability, and other static characteristics. It achieves the goals of low cost, low energy consumption, high network throughput, and high scalability simultaneously.