We propose AQM-LLM, a framework that distills Large Language Models for Active Queue Management in modern networks. Unlike conventional learning-based AQMs that require extensive feature engineering and struggle with dynamic conditions, our approach leverages the contextual reasoning capability of LLMs to enhance the Low Latency, Low Loss, and Scalable Throughput (L4S) architecture with minimal manual intervention. The L4S-LLM design introduces three key components: (i) a state encoder that transforms heterogeneous network telemetry into token embeddings, (ii) a specialized L4S-LLM head that produces congestion actions in a single inference step, and (iii) a data-driven Low-Rank Adaptation scheme that drastically reduces trainable parameters while preserving accuracy. Our open-source FreeBSD-14 implementation demonstrates improved queue delay stability and higher bandwidth utilization for both DCTCP and UDP Prague traffic. We emphasize that this work demonstrates architectural feasibility through controlled experiments; deployment on router-class hardware will require additional model optimization (e.g., compression, pruning, or quantization) and is left for future work.

Large language models (LLMs) are computationally intensive. The computation workload and the memory footprint grow quadratically with the dimension (layer width). Most of LLMs' parameters come from the linear layers of the transformer structure and are highly redundant. These linear layers contribute more than 80% of the computation workload and 99% of the model size. To pretrain and finetune LLMs efficiently, there are three major challenges to address: 1) reducing redundancy of the linear layers; 2) reducing GPU memory footprint; 3) improving GPU utilization when using distributed training. Prior methods, such as LoRA and QLoRA, utilized low-rank structure and quantization to reduce the number of trainable parameters and model size, respectively. However, the resulting model still consumes a large amount of GPU memory. In this paper, we present high-performance GPU-based methods for both pretraining and finetuning quantized LLMs with low-rank structures. We replace a single linear layer in the transformer structure with two narrower linear layers, significantly reducing the number of parameters by several orders of magnitude. By quantizing the pretrained parameters into low precision (8-bit and 4-bit), the memory consumption of the resulting model is further reduced. Compared with existing LLMs, our methods achieve a speedup of 1.3x and a model compression ratio of 2.64 x for pretraining without accuracy drop. For finetuning, our methods achieve an average accuracy score increase of 6.3 and 24.0 in general tasks and financial tasks, respectively, and GPU memory consumption is reduced by 6.3x. The sizes of our models are smaller than 0.59 GB, allowing inference on a smartphone.

Snapshot compressive imaging (SCI) cameras compress high-speed videos or hyperspectral images into measurement frames. However, decoding the data frames from measurement frames is compute-intensive. Existing state-of-the-art decoding algorithms suffer from low decoding quality or heavy running time or both, which are not practical for real-time applications. In this article, we exploit the powerful learning ability of deep neural networks (DNN) and propose a novel tensor fast iterative shrinkage-thresholding algorithm net (Tensor FISTA-Net) as a real-time decoder for SCI cameras. Since SCI cameras have an accurate physical model, we can trade training time for the decoding time by generating abundant synthetic data and training a decoder on the cloud. Tensor FISTA-Net not only learns a sparse representation of the frames through convolution layers but also reduces the decoding time and memory consumption significantly through tensor operations, which makes Tensor FISTA-Net an appropriate approach for a real-time decoder. Our proposed Tensor FISTA-Net obtains an average PSNR improvement of 0.79-2.84 dB (video images) and 2.61-4.43 dB (hyperspectral images) over the state-of-the-art algorithms, along with more clear and detailed visual results on real SCI datasets, Hammer and Wheel, respectively. Our Tensor FISTA-Net reaches 45 frames per second in video datasets and 70 frames per second in hyperspectral datasets, meeting the real-time requirement. Besides, the trained model occupies only a 12 -MB memory footprint, making it applicable to real-time Internet of Things (IoT) applications.

The Smart Grid (SG) infrastructure plan offers growth opportunities for the electric vehicle (EV) industry and aims to reduce dependence on fossil fuels. Surprisingly, the literature lacks comprehensive research on data privacy issues within the EV-SG ecosystem. In response, this paper presents an efficient federated reinforcement learning (FRL) framework tailored to cost-effectively preserve privacy in wireless vehicle-to-grid (V2G) systems. Our approach involves the use of a small auxiliary battery to generate noise, conceal the true energy demand of electric vehicles, and learn the time-varying dynamics of energy usage for wireless EV charging through a federated process. Within this framework, we employ deep Q-learning to concurrently minimize costs and maximize privacy rewards, while exploring innovative techniques to enhance learning speed and communication efficiency through a global FRL approach.

Extracting information from large-scale high-dimensional data is a fundamentally important task in high performance computing, where the hierarchical Tucker (HT) tensor learning approach (learning a tensor-tree structure) has been widely used in many applications. However, HT tensor learning algorithms are compute-intensive due to the “curse of dimensionality,” i.e., the time complexity grows exponentially with the order of the data tensor. The computation of HT tensor learning algorithms boils down to tensor primitives, which are amenable to computing on GPU tensor cores. Existing work does not support HT tensor learning using GPU tensor cores. There are three main challenges to address: 1) to accelerate tensor learning primitives using GPU tensor cores; 2) to implement the tensor learning algorithms using GPU tensor cores and multiple GPUs; 3) to support large-scale data tensors exceeding the GPU memory capacity. In this paper, we present efficient HT tensor learning primitives using GPU tensor cores and demonstrate three applications. First, we utilize GPU tensor cores to optimize HT tensor learning primitives, including tensor contractions, tensor matricizations and tensor singular value decomposition (SVD). We employ the optimized primitives to optimize HT tensor decomposition algorithms for Big Data analysis. Second, we propose a novel HT tensor layer for deep neural networks, whose training process only involves a forward pass without back propagation. The forward pass consists of tensor operations, thus further exploiting the computing power of GPU tensor cores. Third, we apply the optimized primitives to develop a tensor-tree structured quantum machine learning algorithmtree-tensor network (TTN). Compared with TensorLy and TensorNetwork on NVIDIA A100 GPUs, our third-order HT tensor decomposition algorithm achieves up to$8.92 \times$and$6.42 \times$speedups, respectively, and our high-order case achieves up to$32.67 \times$and$23.97 \times$speedups, respectively. Our HT tensor layer for a fully connected neural network achieves$49.2 \times$compression at the cost of 0.5% drops in accuracy and$1.42 \times$speedup compared with the implementation on CUDA cores; for the AlexNet, our HT tensor layer achieves$9.45 \times$compression at the cost of 0.8% drops in accuracy and$1.87 \times$speedup compared with the implementation on CUDA cores. Our TTN algorithm achieves up to$11.17\times$speedup compared with TensorNetwork, indicating the potential of optimized tensor learning primitives for the classical simulation of quantum machine learning algorithms.

Tensor learning is a powerful tool for big data analytics and machine learning, e.g., gene analysis and deep learning. However, tensor learning algorithms are compute-intensive since their time and space complexities grow exponentially with the order of tensors, which hinders their application. In this paper, we exploit the parallelism of tensor learning primitives using GPU tensor cores and develop high-performance tensor learning algorithms. First, we propose novel hardware-oriented optimization strategies for tensor learning primitives on GPU tensor cores. Second, for big data analytics, we employ the optimized tensor learning primitives to accelerate the CP tensor decomposition and then apply it for gene analysis. Third, we optimize the Tucker tensor decomposition and propose a novel Tucker tensor layer to compress deep neural networks. We employ natural gradients to train the neural networks, which only involve a forward pass without backpropagation and thus are suitable for GPU computations. Compared with TensorLab and TensorLy libraries on an A100 GPU, our third-order CP tensor decomposition achieves up to$16.32\times$and$32.25\times$speedups; and$6.09\times$and$6.72\times$speedups for our third-order Tucker tensor decomposition. The proposed fourth-order CP and Tucker tensor decompositions achieve up to$30.65\times$and$5.41\times$speedups over the TensorLab. Our CP tensor decomposition for gene analysis achieves up to$5.88\times$speedup over TensorLy. Compared with a conventional fully connected neural network, our Tucker tensor layer neural network achieves an accuracy of$97.9\%$, a speedup of$4.47\times$, and a compression ratio of$2.92$at the cost of$0.4\%$drop in accuracy.

Multipath TCP (MPTCP) has emerged as a facilitator for harnessing and pooling available bandwidth in wireless/wireline communication networks and in data centers. Existing implementations of MPTCP such as, Linked Increase Algorithm (LIA), Opportunistic LIA (OLIA) and BAlanced LInked Adaptation (BALIA) include separate algorithms for congestion control and packet scheduling, with pre-selected control parameters. We propose a Deep Q-Learning (DQL) based framework for joint congestion control and packet scheduling for MPTCP. At the heart of the solution is an intelligent agent for interface, learning and actuation, which learns from experience optimal congestion control and scheduling mechanism using DQL techniques with policy gradients. We provide a rigorous stability analysis of system dynamics which provides important practical design insights. In addition, the proposed DQL-MPTCP algorithm utilizes the ‘recurrent neural network’ and integrates it with ‘long short-term memory’ for continuously i) learning dynamic behavior of subflows (paths) and ii) responding promptly to their behavior using prioritized experience replay. With extensive emulations, we show that the proposed DQL-based MPTCP algorithm outperforms MPTCP LIA, OLIA and BALIA algorithms. Moreover, the DQL-MPTCP algorithm is robust to time-varying network characteristics, and provides dynamic exploration and exploitation of paths.

The bottleneck of distributed edge learning (DEL) over wireless has shifted from computing to communication, primarily the aggregation-averaging (Agg-Avg) process of DEL. The existing transmission control protocol (TCP)-based data networking schemes for DEL are application-agnostic and fail to deliver adjustments according to application layer requirements. As a result, they introduce massive excess time and undesired issues such as unfairness and stragglers. Other prior mitigation solutions have significant limitations as they balance data flow rate from workers across paths but often incur imbalanced backlogs when the paths exhibit variance, causing stragglers. To facilitate a more productive DEL, we develop a hybrid multipath TCP (MPTCP) by combining model-based and deep reinforcement learning (DRL) based MPTCP for DEL that strives to realize quicker iteration of DEL and better fairness (by ameliorating stragglers). Hybrid MPTCP essentially integrates two radical TCP developments: i) successful existing model-based MPTCP control strategies and ii) advanced emerging DRL-based techniques, and introduce a novel hybrid MPTCP data transport for easing the communication of Agg-Avg process. Extensive emulation results demonstrate that the proposed hybrid MPTCP can overcome excess time consumption and ameliorate the application layer unfairness of DEL effectively without injecting additional inconstancy and stragglers.

Recommendations based on prediction of user preferences from partial information are widely used in various applications. However, recommendations using smart devices have some challenges related to limited data device resources, data sparsity, and data privacy. Since there are many multidimensional data in smart devices, recommendations may collect a large amount of user private data. In this article, we study privacy-preserving recommendations with high-dimensional tensor data in smart devices. First, we propose a federated tensor completion scheme to infer the user’s preferences and we prove that this scheme satisfies the differential privacy guarantee. Our scheme consists of a global update and a local update, which reduce information exposure and guarantee local data privacy. Second, we mathematically analyze the privacy and utility of the proposed algorithm. Third, we provide empirical evaluations on synthetic data sets and real-world data sets. Results show that our scheme has a low recovery error and provides strong privacy protection.

We consider few-shot learning (FSL), where a model learns from very few labeled examples such that it can generalize to unseen examples. Model-agnostic meta-learning (MAML) has been proposed to solve FSL. However, the low performance of MAML suggests its difficulty in tackle diverse tasks, due to the restriction of sharing a single model initialization for fast adaptation. In this paper, we propose meta-learning with attention mechanisms. Our method meta-learns attention modules to instantiate task-specific model initialization for fast adaptation, which can obtain high-quality solution to a new task using few gradient descent steps. To further improve generalization during inference, we propose to incorporate an entropy regularizer into the adaptation objective to penalize the Shannon entropy of prediction probability. Extensive experiments under various FSL scenarios show that our method achieves state-of-the-art performance on the mini-ImageNet and tiered-ImageNet.

Video frame synthesis is an important task in computer vision and has drawn great interests in wide applications. However, existing neural network methods do not explicitly impose tensor low-rankness of videos to capture the spatiotemporal correlations in a high-dimensional space, while existing iterative algorithms require hand-crafted parameters and take relatively long running time. In this paper, we propose a novel multi-phase deep neural network Transform-Based Tensor-Net that exploits the low-rank structure of video data in a learned transform domain, which unfolds an Iterative Shrinkage-Thresholding Algorithm (ISTA) for tensor signal recovery. Our design is based on two observations: (i) both linear and nonlinear transforms can be implemented by a neural network layer, and (ii) the soft-thresholding operator corresponds to an activation function. Further, such an unfolding design is able to achieve nearly real-time at the cost of training time and enjoys an interpretable nature as a byproduct. Experimental results on the KTH and UCF-101 datasets show that compared with the state-of-the-art methods, i.e., DVF and Super SloMo, the proposed scheme improves Peak Signal-to-Noise Ratio (PSNR) of video interpolation and prediction by 4.13 dB and 4.26 dB, respectively.

We consider the auto-scaling problem for application hosting in a cloud, where applications are elastic and the number of requests changes over time. The application requests are serviced by Virtual Machines (VMs), which reside on Physical Machines (PMs) in a cloud. We aim to minimize the number of hosting PMs by intelligently packing VMs into PMs, while the VMs are auto-scaled, i.e., dynamically acquired and released, to accommodate varying application needs. We consider a shadow routing based approach for this problem. The proposed shadow algorithm employs a specially constructed virtual queueing system to dynamically produce an optimal solution that guides the VM auto-scaling and the VM-to-PM packing. The proposed algorithm runs continuously without the need to re-solve the underlying optimization problem “from scratch”, and adapts automatically to the changes in the application demands. We prove the asymptotic optimality of the shadow algorithm. The simulation experiments further demonstrate the algorithm's good performance and high adaptivity.

With the rapid development and proliferation of multimedia systems and applications, there is a growing need to handle multimedia heterogeneous data safely and efficiently on the cloud. Tensor models are effective in representing multimedia multidimensional data, and the tensor decomposition is one of the basic building blocks of data analysis and learning models. In this article, we propose a secure tensor singular value decomposition (${S}$-tSVD), in which the time-domain operation is converted into a scheme featuring frequency-domain multilinear circular unfolding–folding. First, we represent various multimedia data as cipher subtensors, using fully homomorphic encryption. We then take the fast Fourier transform (FFT) approach to launch a new multiplication operation along the tubal fibers of a unified high-order tensor. Second, relying on the homomorphism of addition and multiplication theory, we prove the fully homomorphic consistency of the proposed${S}$-tSVD algorithm. Third, we provide an elegant solution to tackle the typical dimensionality inconsistency problem while working with multiple subtensors. Finally, we carry out theoretical analyses with respect to dimensionality reduction, reconstruction error of${S}$-tSVD, running time, and data security. We use real unstructured video data and semistructured XML documents, integrating them within a unified tensor model for decomposition. We demonstrate that the error ratio of the${S}$-tSVD is lower than the same compression ratio compared to the SVD decomposition and tSVD-slice approaches. Moreover, the specific${S}$-tSVD decomposition not only enables effective data mining and dimensionality reduction but also ensures the accuracy of the decomposition result and data privacy protection.

Tensors are the cornerstone data structures in high-performance computing, big data analysis and machine learning. However, tensor computations are compute-intensive and the running time increases rapidly with the tensor size. Therefore, designing high-performance primitives on parallel architectures such as GPUs is critical for the efficiency of ever growing data processing demands. Existing GPU basic linear algebra subroutines (BLAS) libraries (e.g., NVIDIA cuBLAS) do not provide tensor primitives. Researchers have to implement and optimize their own tensor algorithms in a case-by-case manner, which is inefficient and error-prone. In this paper, we develop the cuTensor-tubal library of seven key primitives for the tubal-rank tensor model on GPUs: t-FFT, inverse t-FFT, t-product, t-SVD, t-QR, t-inverse, and t-normalization. cuTensor-tubal adopts a frequency domain computation scheme to expose the separability in the frequency domain, then maps the tube-wise and slice-wise parallelisms onto the single instruction multiple thread (SIMT) GPU architecture. To achieve good performance, we optimize the data transfer, memory accesses, and design the batched and streamed parallelization schemes for tensor operations with data-independent and data-dependent computation patterns, respectively. In the evaluations oft-product, t-SVD, t-QR, t-inverse and t-normalization, cuTensor-tubal achieves maximum 16.91x, 27.03x, 38.97x, 22.36x,15.43x speedups respectively over the CPU implementations running on dual 10-core Xeon CPUs. Two applications, namely, t-SVD-based video compression and low-tubal-rank tensor completion, are tested using our library and achieve maximum 9.80x and 269.26x speedups over multi-core CPU implementations.

Efficient network slicing is vital to deal with the highly variable and dynamic characteristics of traffic in 5G networks. Network slicing addresses a challenging dynamic network resource allocation problem where a single network infrastructure is divided into (virtual) multiple slices to meet the demands of different users with varying requirements, the main challenges being - the traffic arrival characteristics and the job resource requirements (e.g., compute, memory and bandwidth resources) for each slice can be highly dynamic. Traditional model-based optimization or queueing theoretic modeling becomes intractable with the high reliability, and stringent bandwidth and latency requirements imposed by 5G. We propose a deep reinforcement learning approach to address this dynamic coupled resource allocation problem. Model evaluation using synthetic and real workload data demonstrates that our deep reinforcement learning solution improves overall resource utilization, latency performance, and demands satisfied as compared to a baseline equal slicing strategy.

By abstracting network resources and functions, network slicing allows to partition a single physical infrastructure into multiple logical independent networks, called slices, that can be tailored to the diverse needs of a wide range of application domains. In order to be independently managed by different players (tenants), slices need to be logically isolated from each other. As some of the applications envisioned for 5G networks are quite critical from performance point of view, a static slicing of resources, where isolation is not only logical but also physical, may appear a natural solution. However, the price to pay in terms of resource overprovisioning can be high, and therefore dynamic allocation strategies, based on optimized resource sharing, are alternative solutions. In this work, we propose an optimization framework to compare the two resource allocation schemes and define a metric to quantify the slicing gain. We show that the dynamic slicing outperforms the static approach and that the gain is significant even for stringent performance requirements, making dynamic slicing a viable solution also for critical applications.

As we enter the Internet of Things (IoT) era in which the communication network is becoming increasingly dynamic, heterogeneous, and complex, it is desirable to have cognitive communication systems and networks that possess multiple interacting capabilities for situation assessment, resource management, online/distributed learning, big-data processing, and intelligent decision making. AI techniques, such as deep learning, probabilistic graph model, and reinforcement learning, aided with big data and IoT, provide a wide variety of tools and solutions to many new problems encountered in the design, operation, and optimization of cognitive communication systems and networking, including resource management, situation assessment, channel identification, anomaly detection, root cause analysis, and online/distributed learning.

Multiple distributed controllers have been used in software-defined networks (SDNs) to improve scalability and reliability, where each controller manages one static partition of the network. In this paper, we show that dynamic mapping between switches and controllers can improve efficiency in managing traffic load variations. In particular, we propose balanced controller (BalCon) and BalConPlus, two SDN switch migration schemes to achieve load balance among SDN controllers with small migration cost. BalCon is suitable for the scenarios where the network does not require a serial processing of switch requests. For other scenarios, BalConPlus is more suitable, as it is immune to the switch migration blackout and does not cause any service disruption. Simulations demonstrate that BalCon and BalConPlus significantly reduce the load imbalance among SDN controllers by migrating only a small number of switches with low computation overhead. We also build a prototype testbed based on the open-source SDN framework RYU to verify the practicality and effectiveness of BalCon and BalConPlus. Experiment confirms the results of the simulations. It also shows that BalConPlus is immune to switch migration blackout, an adverse effect in the baseline BalCon.

Radio-frequency (RF) tomographic imaging is a promising technique for inferring multi-dimensional physical space by processing RF signals traversed across a region of interest. However, conventional RF tomography schemes are generally based on vector compressed sensing, which ignores the geometric structures of the target spaces and leads to low recovery precision. The recently proposed transform-based tensor model is more appropriate for sensory data processing, as it helps exploit the geometric structures of the three-dimensional target and improve the recovery precision. In this paper, we propose a novel tensor sensing approach that achieves highly accurate estimation for real-world three-dimensional spaces. First, we use the transform-based tensor model to formulate a tensor sensing problem, and propose a fast alternating minimization algorithm called Alt-Min. Secondly, we drive an algorithm which is optimized to reduce memory and computation requirements. Finally, we present evaluation of our Alt-Min approach using IKEA 3D data and demonstrate significant improvement in recovery error and convergence speed compared to prior tensor-based compressed sensing.

In this work, we investigate the problem of joint optimization over placement and routing of network function chains in data centers. In the offline case, we demonstrate that a classical randomization algorithm works well and we derive a new bound on the performance sub-optimality gap. In the online case, we prove a fundamental lower bound in resource violation and propose a new algorithm that combines techniques from multiplicative weight update and primal-dual update paradigms. This online algorithm asymptotically achieves the best possible performance in terms of resource allocation among all online algorithms. We demonstrate the applicability of our solutions to address practical problems by conducting simulation-based evaluations over different data center architectures using data generated from real trace distribution.

The Blockchain paradigm provides a popular mechanism for establishing trust and consensus in distributed environments. While Blockchain technology is currently primarily deployed in crypto-currency systems like Bitcoin, the concept is also expected to emerge as a key component of the Internet-of-Things (IoT), enabling novel applications in digital health, smart energy, asset tracking and smart transportation. As Blockchain networks evolve to industrial deployments with large numbers of geographically distributed nodes, the block transfer and processing delays arise as a critical issue which may create greater potential for forks and vulnerability to adversarial attacks. Motivated by these issues, we develop stochastic network models to capture the Blockchain evolution and dynamics and analyze the impact of the block dissemination delay and hashing power of the member nodes on Blockchain performance in terms of the overall block generation rate and required computational power for launching a successful attack. The results provide useful insight in crucial design issues, e.g., how to adjust the `difficulty-of-work' in the presence of delay so as to achieve a target block generation rate or appropriate level of immunity from adversarial attacks. We employ a combination of analytical calculations and simulation experiments to investigate both stationary and transient performance features, and demonstrate close agreement with measurements on a wide-area network testbed running the Ethereum protocol.

We consider a shadow routing based approach to the problem of real-time adaptive placement of virtual machines (VM) in large data centers (DC) within a network cloud. Such placement in particular has to respect vector packing constraints on the allocation of VMs to host physical machines (PM) within a DC, because each PM can potentially serve multiple VMs simultaneously. Shadow routing is attractive in that it allows a large variety of system objectives and/or constraints to be treated within a common framework (as long as the underlying optimization problem is convex). Perhaps even more attractive feature is that the corresponding algorithm is very simple to implement, it runs continuously, and adapts automatically to changes in the VM demand rates, changes in system parameters, etc., without the need to re-solve the underlying optimization problem “from scratch”. In this paper we focus on the min-max-DC-load problem. Namely, we propose a combined VM-to-DC routing and VM-to-PM assignment algorithm, referred to as Shadow scheme, which minimizes the maximum of appropriately defined DC utilizations. We prove that the Shadow scheme is asymptotically optimal (as one of its parameters goes to 0). Simulation confirms good performance and high adaptivity of the algorithm. Favorable performance is also demonstrated in comparison with a baseline algorithm based on VMware implementation [7], [8]. We also propose a simplified - “more distributed” - version of the Shadow scheme, which performs almost as well in simulations.

Scalability and reliability are among the main concerns in large-scale Software Defined Networking (SDN) application scenarios. A common approach is to use multiple distributed controllers, each managing one static partition of the network. In this paper, we show that dynamic mapping can improve efficiency in managing traffic load variations. We then propose BalCon (Balanced Controller): an algorithmic solution designed to tackle and reduce the load imbalance among SDN controllers through proper SDN switch migrations. Simulations demonstrate that BalCon is lightweight from the computational point of view and reduces the load imbalance among SDN controllers (expressed as variance) by 40% by migrating only a small number of switches. We also built a realistic prototype of SDN controller, BalConController, based on the open-source SDN framework RYU.

Network functions typically need to be visited in a specific order to meet certain objectives, giving rise to the notion of Service Function Chaining. Software-Defined-Networking enables fine-grained traffic routing optimization while satisfying correct traversal of network functions. In this work, we investigate the problem of maximizing throughput in SDN-enabled networks with respect to service chaining specifications under both traditional and new constraints. Besides the algorithm design, we also derive rigorous performance bounds. In the offline traffic routing case, we propose Traffic-Merging-Algorithm and prove that, although the underlying optimization problem is generally NP-hard, our algorithm can efficiently compute the optimal solution in practical settings. In the online traffic routing case, we propose the Primal-Dual-Update-Algorithm, which comes with a system parameter that trades off the algorithm's throughput competitiveness and its meeting of QoS requirements, and prove that our online algorithm achieves optimal tradeoff. We demonstrate that our solutions can be used to address practical problems by conducting simulation-based evaluation over backbone and data center topologies.

Multipath TCP (MP-TCP) has the potential to greatly improve application performance by using multiple paths transparently. We propose a fluid model for a large class of MP-TCP algorithms and identify design criteria that guarantee the existence, uniqueness, and stability of system equilibrium. We clarify how algorithm parameters impact TCP-friendliness, responsiveness, and window oscillation and demonstrate an inevitable tradeoff among these properties. We discuss the implications of these properties on the behavior of existing algorithms and motivate our algorithm Balia (balanced linked adaptation), which generalizes existing algorithms and strikes a good balance among TCP-friendliness, responsiveness, and window oscillation. We have implemented Balia in the Linux kernel. We use our prototype to compare the new algorithm to existing MP-TCP algorithms.

Upstream dynamic bandwidth allocation (DBA) for Gigabit Passive Optical Networks (GPON) is an important design problem since the shared uplink carries many bursty streams of different QoS requirements and there are various practical system constraints on processing loads in the OLT (Optical Line Terminal) and the ONUs (Optical Network Units). We propose self-adaptive bandwidth allocation solutions, where the OLT allocates bandwidth among the ONUs in an efficient and responsive manner, and importantly, without requiring buffer status reports from ONUs. These solutions could allow design of simpler ONUs. Our first solution is based on simple estimation by the OLT of the sample gradient of the ONU buffer content and application of stochastic approximation for sequential updates of bandwidth allocations that minimize buffer contents. The second approach is a closed form solution that minimizes the sum of weighted throughput across all ONUs. We compare our solutions with DBA algorithms that require buffer status reporting and empirically demonstrate their advantages.

Most mobile devices today come with multiple access interfaces, \emph{e.g.}, 4G and WiFi. Multipath TCP (MP-TCP) can greatly improve network performance by exploiting the connection diversity of multiple access interfaces, at the expense of higher energy consumption. In this paper, we design MP-TCP algorithms for mobile devices by jointly considering the performance and energy consumption. We consider two main types of mobile applications: realtime applications that have a fixed duration and file transfer applications that have a fixed data size. For each type of applications, we propose a two-timescale algorithm with theoretical guarantee on the performance. We present simulation results that show that our algorithms can reduce energy consumption by up to 22$\%$ without sacrificing throughput compared to a baseline MP-TCP algorithm.

We develop shadow routing based online algorithms for the joint problem of application-to-VM and VM-to-PM assignments in a cloud environment. The asymptotic optimality of the shadow algorithm is proved and the performance is evaluated by simulations.

Recent standardization efforts to reduce power consumption in optical access networks include the use of cyclic sleep modes to more effectively scale energy consumption with activity. This paper analyzes a probing-based cyclic sleep mode mechanism that is applied to the downstream of a passive optical network. The analysis studies the effectiveness of the sleep mode mechanism and provides explicit mathematical expressions for the power consumption and packet response delay in the optical network unit (ONU) at the customer premises, and the amount of buffered traffic per ONU in the optical line terminal at the central office. The analytical results are confirmed and complemented by numerical simulations, and used to provide guidelines for selecting sleep mode configuration parameters.

We consider a shadow routing based approach to the problem of real-time adaptive placement of virtual machines (VM) in large data centers (DC) within a network cloud. Such placement in particular has to respect vector packing constraints on the allocation of VMs to host physical machines (PM) within a DC, because each PM can potentially serve multiple VMs simultaneously. Shadow routing is attractive in that it allows a large variety of system objectives and/or constraints to be treated within a common framework (as long as the underlying optimization problem is convex). Perhaps even more attractive feature is that the corresponding algorithm is very simple to implement, it runs continuously, and adapts automatically to changes in the VM demand rates, changes in system parameters, etc., without the need to re-solve the underlying optimization problem “from scratch”. In this paper we focus on the minmax-DC-load problem. Namely, we propose a combined VM-toDC routing and VM-to-PM assignment algorithm, referred to as Shadow scheme, which minimizes the maximum of appropriately defined DC utilizations. We prove that the Shadow scheme is asymptotically optimal (as one of its parameters goes to 0). Simulation confirms good performance and high adaptivity of the algorithm. Favorable performance is also demonstrated in comparison with a baseline algorithm based on VMware implementation [7], [8]. We also propose a simplified - “more distributed” - version of the Shadow scheme, which performs almost as well in simulations.

Saving power in datacenter networks has become a pressing issue. While in operation, ElasticTree and CARPO can save power consumed by a fat-tree network by using sleep mode where some components such as ports and switches are turned off when traffic demand in the network is relatively moderate. In this paper, we propose a new approach by exploring the design stage of a datacenter network and focus on how to choose the right switch size that can potentially save the most power during the expected operation of the network. We also consider speed scaling where the power of a switch can be varied by adjusting its processing rate according to its traffic demand. We use analysis and simulation to investigate the power-saving performance of different switch sizes, power-saving modes and traffic demand patterns. Our findings with sleep mode reveal that deploying a large number of small switches is more power-efficient than a small number of large switches when the traffic demand is relatively moderate or when servers exchanging traffic are in close proximity. With speed scaling, the reverse is generally true.

Multi-path TCP (MP-TCP) has the potential to greatly improve application performance by using multiple paths transparently. We propose a fluid model for a large class of MP-TCP algorithms and identify design criteria that guarantee the existence, uniqueness, and stability of system equilibrium. We characterize algorithm parameters for TCP-friendliness and prove an inevitable tradeoff between responsiveness and friendliness. We discuss the implications of these properties on the behavior of existing algorithms and motivate a new design that generalizes existing algorithms. We use ns2 simulations to evaluate the proposed algorithm and illustrate its superior overall performance.

Reducing power consumption in access networks has become an increasingly important design goal due to increased concerns for both global warming and network operation costs. Although passive optical networks (PONs) consume the least power among all access network technologies, it is desirable to further reduce PON's power consumption, especially when the line rate is being increased to 10Gbps (10G) and wider deployment is underway. Generally, an effective approach to reduce power consumption is to use sleep-mode operation where functionality is powered off during periods of idleness. This paper therefore proposes a sleep-mode based power saving protocol applied to an optical line terminal (OLT) and optical network units (ONUs) to implement ONU power saving in 10G-Ethernet-PON (10G-EPON) systems. The proposed protocol explicitly distinguishes sleep-, listen-, and awake- states with different power consumption levels. In addition, the paper proposes analytical models for ONU power consumption and response delay to evaluate performance of the proposed power saving protocol. The simulations show that the proposed protocol can significantly improve ONU power saving. The validity of the numerical models is also confirmed by extensive simulations.

Various link bandwidth adjustment mechanisms are being developed to save network energy. However, their interaction with congestion control can significantly reduce network throughput, and is not well understood. We firstly put forward a framework to study this interaction, and then propose an easily implementable dynamic bandwidth adjustment (DBA) mechanism for the links. In DBA, each link updates its bandwidth according to an integral control law to match its average buffer size with a target buffer size. We prove that DBA reduces link bandwidth without sacrificing throughput---DBA only turns off excess bandwidth---in the presence of congestion control. Preliminary ns2 simulations confirm this result.

There have been several approaches to provisioning traffic between core network nodes in Internet Service Provider networks. Such approaches aim to minimize network delay, increase capacity, and enhance security services. MATE (Multipath Adaptive Traffic Engineering) has been proposed for multipath adaptive traffic engineering between an ingress node (source) and an egress node (destination). Its novel idea is to avoid network congestion and attacks that might exist in edge and node disjoint paths between two core network nodes. This paper builds an adaptive, robust, and reliable traffic engineering scheme for better performance and operation of communication networks. This will also provision quality of service (QoS) and protection of traffic engineering to maximize network efficiency. Specifically, we present a new approach, S-MATE (secure MATE) is developed to protect the network traffic between two core nodes (routers, switches, etc.) in a cloud network. S-MATE secures against a single link attack/failure by adding redundancy in one of the operational paths between the sender and receiver. The proposed scheme can be built to secure core networks such as optical and IP networks.

The delivery of video content is expected to gain huge momentum, fueled by the popularity of user-generated clips, growth of VoD libraries, and wide-spread deployment of IPTV services with features such as CatchUp/PauseLive TV and NPVR capabilities. The `time-shifted' nature of these personalized applications defies the broadcast paradigm underlying conventional TV networks, and increases the overall bandwidth demands by orders of magnitude. Caching strategies provide an effective mechanism for mitigating these massive bandwidth requirements by replicating the most popular content closer to the network edge, rather than storing it in a central site. The reduction in the traffic load lessens the required transport capacity and capital expense, and alleviates performance bottlenecks. In the present paper, we develop light-weight cooperative cache management algorithms aimed at maximizing the traffic volume served from cache and minimizing the bandwidth cost. As a canonical scenario, we focus on a cluster of distributed caches, either connected directly or via a parent node, and formulate the content placement problem as a linear program in order to benchmark the globally optimal performance. Under certain symmetry assumptions, the optimal solution of the linear program is shown to have a rather simple structure. Besides interesting in its own right, the optimal structure offers valuable guidance for the design of low-complexity cache management and replacement algorithms. We establish that the performance of the proposed algorithms is guaranteed to be within a constant factor from the globally optimal performance, with far more benign worst-case ratios than in prior work, even in asymmetric scenarios. Numerical experiments for typical popularity distributions reveal that the actual performance is far better than the worst-case conditions indicate.

One of the key questions in dimensioning a hybrid P2P content distribution system is that of the required infrastructure support in terms of server bandwidth. In this paper, we develop and propose simple mathematical models for analyzing and dimensioning hybrid peer-to-peer content distribution networks. We first use a deterministic fluid model to capture the essential peer and server dynamics within a single swarm, and subsequently derive a stochastic fluid model to capture the dynamics in the case of multiple swarms, i.e., concurrent swarms of a number of content objects. Based on the models, we derive solutions for estimating the server capacity required to support a single swarm as well as a number of concurrent file swarms at a given level of service quality. Numerical results demonstrate how a hybrid P2P approach can yield substantial performance gains and capacity savings compared to a pure client/server system, with churn rate and upload bandwidth being critical factors. Compared to a pure peer-to-peer scenario, the hybrid approach can dramatically boost the performance and improve reliability.

We present a novel approach for joint optical network provisioning and Internet protocol (IP) traffic engineering, in which the IP and optical networks collaboratively optimize a combined objective of network performance and lightpath provisioning cost. We develop a framework for distributed multilayer optimization. Our framework is built upon the IP-over-optical (IPO) overlay model, where each network domain has a limited view of the other. Our formulation allows the two domains to communicate and coordinate their decisions through minimal information exchange. Our solution is based on a novel application of Generalized Bender's Decomposition, which divides a difficult global optimization problem into tractable subproblems, each solved by a different domain. The procedure is iterative and converges to the global optimum. We present case studies to demonstrate the efficiency and applicability of our approach in various networking scenarios. Our work builds a foundation for "multilayer" grooming, which extends traditional grooming in the optical domain to include data networks. The data networks are active participants in the grooming process with intelligent homing of data traffic to optical gateways

The Internet transport infrastructure is moving towards a model of high-speed router networks that are directly interconnected by re-configurable optical core networks. This IP over optical (IPO) architecture, when coupled with the emerging IP-based generalized multi-protocol label switching (GMPLS) control plane, offers network operators opportunities for dynamic, multi-layer optimization of traffic engineering. We present a novel approach for joint optical network provisioning and IP traffic engineering, in which the IP and optical network domains collaboratively achieve the objective of maximizing the benefit of carrying end-to-end IP traffic at the minimum lightpath design cost. Our approach is developed under the overlay inter-networking model with each network having a limited view of the other. The cooperation is achieved through minimal information exchange. Our solution is based on a novel application of generalized bender's decomposition to implement distributed, multi-layer optimization. The process converges to the optimal value of the objective function. A case study is presented to illustrate the proposed method.

We consider meshed optical transport networks having multiple levels of hierarchy whereby cross-connects in different levels perform multiplexing and demultiplexing functions at different granularities. We present a traffic grooming approach that can be implemented in a centralized or distributed fashion, based on the novel design that takes advantage of algorithm efficiency and a simple threshold mechanism to decide when grooming is economical. Our approach allows nodes to perform grooming and degrooming automatically, which is desired in the next-generation optical network where circuits are setup and torn-down dynamically through signaling. We present several experiments using different network topologies. The results indicate that the flow pattern influences the port requirement behavior, and that the flow thickness influences the optimal value of the threshold.

Link bundling is a way to increase routing scalability whenever a pair of label switching routers in MPLS are connected by multiple parallel links. However, link bundling can be inefficient as a label switched path (LSP) has to be associated with a particular link. In this paper, we show that the efficiency of link bundling can be significantly improved if traffic can be effectively distributed across the parallel links. We propose an IP switch architecture that is capable of distributing flows both inside the switch and among the parallel links based on operations that are relatively simple to implement. The switch requires no speedup, guarantees in-sequence packet delivery for a given flow, avoids complex coordination algorithms, and can achieve LSP throughput higher than the line rate. By means of simulation using IP traces, we investigate the performance of the proposed switch, and show that the switch achieves good load-balancing performance. We describe extensions to the basic architecture which allows for very large bundle size, handles incremental upgrade strategy, improves reliability, and accommodates nonIP traffic.

Generalized multi-protocol label switching (GMPLS) is currently being specified as the control plane for next-generation optical networks. This paper investigates how GMPLS affects the performance of lambda-router networks. Our performance study reveals the advantage of using a fully tunable laser at each source in the presence of a limited number of wavelength converters. We propose an enhancement to GMPLS which allows a lightpath request to crankback to a lambda router that can use a wavelength converter to complete the lightpath setup successfully. This effectively makes the wavelength converters shared across the network. It is shown that crankback provides an appreciable improvement in performance when fully tunable lasers at the sources are not universally available.

Using measurements collected at network monitors to infer network conditions is a promising approach for network-centric monitoring. In this context, an important question arises: given the number and locations of network monitors, how much network management resources (e.g., the number of measurements) are needed to obtain an accurate estimate of network states? We define the scalability of measurement-based network monitoring as the growth rate of the number of measurements required for accurate network monitoring/inference with respect to the size of a network. We develop a framework for investigating the scalability in the context of multicast inference with the monitors at the edges of a network. In such a framework, network monitoring/inference can be formulated as probability density estimation of network states. The growth rate is characterized through the sample complexity, which is the number of measurements needed to accurately estimate the density. The missing data framework is introduced to estimate the growth rate, where the missing data reflect unavailable measurements at the unobservable nodes without resident monitors, and the underlying nodal packet losses. We show that when the missing information is mainly due to the number of unobservable nodes, the number of measurements needed grows linearly with the size of the network, and the measurement-based inference approach is, thus, scalable. When the missing information is mainly due to the underlying nodal packet losses, the number of measurements needed grows faster than linear with the size of the network, and the measurement-based inference approach is, thus, not scalable. Our results provide guidelines for accessing feasibility of the measurement-based inference approach, and the number of probes required. We give numerical examples to illustrate some of our results.

Destination-based forwarding in traditional IP routers has not been able to take full advantage of multiple paths that frequently exist in Internet service provider networks. As a result, the networks may not operate efficiently, especially when the traffic patterns are dynamic. This paper describes a multipath adaptive traffic engineering mechanism, called MATE, which is targeted for switched networks such as multiprotocol label switching (MPLS) networks. The main goal of MATE is to avoid network congestion by adaptively balancing the load among multiple paths based on measurement and analysis of path congestion. MATE adopts a minimalist approach in that intermediate nodes are not required to perform traffic engineering or measurements besides normal packet forwarding. Moreover MATE does not impose any particular scheduling, buffer management, or a priori traffic characterization on the nodes. This paper presents an analytical model, derives a class of MATE algorithms, and proves their convergence. Several practical design techniques to implement MATE are described. Simulation results are provided to illustrate the efficacy of MATE under various network scenarios.

Generalized processor sharing (GPS) is the basis for the packet scheduler of choice in IP routers and ATM switches of the future. The currently accepted approach for the design of GPS schedulers is based on deterministic QoS guarantees, which, it is generally accepted, is overly conservative and leads to limitations on capacity. We develop a framework for GPS scheduling which is based on statistical QoS guarantees and statistical multiplexing. We give the design of GPS weights which maximize the coverage of operating points, and also the design of the connection admission control (CAC). The general framework is end-to-end, with two heterogeneous QoS classes coexisting with a third, best effort class. Each QoS class has a specified delay bound together with a bound on the probability of its violation. An important objective is to maximize the bandwidth available to best effort traffic, while just satisfying the guarantees of the QoS classes. To this end, we consider output regulated GPS scheduling, which limits each connection's share of the bandwidth to a designed value. The sources are subject to standard dual leaky bucket regulation. For the design of the GPS weights we give procedures based on two key concepts, the realizable set and the critical weights. The realizable set is the union of all admissible sets of connections of both classes over all weights. One of the main contributions is a pragmatic design process by which most of the realizable set is realized by only two critical weights. The numerical results show that there are substantial capacity gains from statistical multiplexing.

In a multiprotocol label switching (MPLS) domain, ATM label-switching routers (LSRs) are potentially capable of providing the highest forwarding capacity in the backbone network. Virtual circuit (VC) merging is a mechanism in an ATM-LSR that allows many IP routes to be mapped to the same VC label and provides a scalable mapping method that can support thousands of destinations. VC merging requires reassembly buffers so that cells belonging to different packets intended for the same destination do not interleave with each other. In this paper, the impact of VC merging on the buffering requirement for the reassembly buffers is investigated. We propose a realistic architecture that supports VC merging. We study the performance of this architecture using an analytic approach and using simulation driven by empirical Internet packet-size distribution. At the cell level, our main finding indicates that VC merging incurs a minimal overhead compared to non-VC merging, in terms of additional buffering. Moreover, the overhead decreases as utilization increases or as the traffic becomes more bursty with longer dependence. The finding has important practical consequences since routers and switches are dimensioned for high utilization and stressful traffic conditions. At the packet level, VC merging generally achieves a higher goodput than non-VC merging with EPD for the same buffer size. We also study the delay performance and find that the additional delay due to VC merging is insignificant at high speed.

VC merging allows many routes to be mapped to the same VC label, providing a scalable mapping method that can support tens of thousands of edge routers. VC merging requires reassembly buffers so that cells belonging to different packets intended for the same destination do not interleave with each other. The impact of VC merging on the additional buffer required for the reassembly buffers and other buffers due to the perturbation in the traffic process is investigated. We propose a realistic output-buffered ATM switch architecture that supports VC merging capability. We analyze the performance of the switch using a decomposition approach, and verify the results using simulation. We investigate the impact of VC merging on loss and delay performance for realistic traffic scenarios. The main result indicates that VC merging incurs a minimal overhead compared to non-VC merging in terms of additional buffering. Moreover, the overhead decreases as utilization increases, or as the traffic becomes more bursty. The finding has important implication since practical ATM switches are dimensioned for high utilization and stressful traffic conditions. We also study the delay performance and find that the additional delay due to VC merging is insignificant for most applications.

This paper develops a framework of traffic shaping which has the goal of increasing the nodal connection-carrying capacity. The shaping filters are designed to interwork with statistical multiplexers which use FIFO buffers. Differences in delay tolerances between traffic classes are exploited to shape and smooth the less delay sensitive traffic. The source model used in the analysis is asynchronous, worst-case subject to dual leaky bucket regulation, where the asynchrony is due to the assumption of independent sources. This, together with the allowance of small loss probabilities, makes statistical multiplexing feasible. We require the shapers to be lossless so that losses, if any, occur only at the multiplexer. We obtain the optimal design of the shaper such that a given delay tolerance for the connection is satisfied. For purposes of admission control we obtain the admissible region of combinations of sources of various types such that the QoS requirements in loss and delay are satisfied. By examining the regions obtained with and without shaping we quantify the capacity gain from shaping. Our numerical results show that the gain can be substantial.

There has been a growing concern about the potential impact of long-term correlations (second-order statistic) in variable-bit-rate (VBR) video traffic on ATM buffer dimensioning. Previous studies have shown that video traffic exhibits long-range dependence (LRD) (Hurst parameter large than 0.5). We investigate the practical implications of LRD in the context of realistic ATM traffic engineering by studying ATM multiplexers of VBR video sources over a range of desirable cell loss rates and buffer sizes (maximum delays). Using results based on large deviations theory, we introduce the notion of Critical Time Scale (CTS). For a given buffer size, link capacity, and the marginal distribution of frame size, the CTS of a VBR video source is defined as the number of frame correlations that contribute to the cell loss rate. In other words, second-order behavior at the time scale beyond the CTS does not significantly affect the network performance. We show that whether the video source model is Markov or has LRD, its CTS is finite, attains a small value for small buffer, and is a non-decreasing function of buffer size. Numerical results show that (i) even in the presence of LRD, long-term correlations do not have significant impact on the cell loss rate; and (ii) short-term correlations have dominant effect on cell loss rate, and therefore, well-designed Markov traffic models are effective for predicting Quality of Service (QOS) of LRD VBR video traffic. Therefore, we conclude that it is unnecessary to capture the long-term correlations of a real-time VBR video source under realistic ATM buffer dimensioning scenarios as far as the cell loss rates and maximum buffer delays are concerned.

The main contributions of this paper are two-fold. First, we prove fundamental, similarly behaving lower and upper bounds, and give an approximation based on the bounds, which is effective for analyzing ATM multiplexers, even when the traffic has many, possibly heterogeneous, sources and their models are of high dimension. Second, we apply our analytic approximation to statistical models of video teleconference traffic, obtain the multiplexing system's capacity as determined by the number of admissible sources for given cell loss probability, buffer size and trunk bandwidth, and, finally, compare with results from simulations, which are driven by actual data from coders. The results are surprisingly close. Our bounds are based on Large Deviations theory. Our approximation has two easily calculated parameters, one is from Chernoff's theorem and the other is the system's dominant eigenvalue. A broad range of systems are analyzed and the time for analysis in each case is a fraction of a second.

The main contributions of this paper are two-fold. First, we prove fundamental, similarly behaving lower and upper bounds, and give an approximation based on the bounds, which is effective for analyzing ATM multiplexers, even when the traffic has many, possibly heterogeneous, sources and their models are of high dimension. Second, we apply our analytic approximation to statistical models of video teleconference traffic, obtain the multiplexing system's capacity as determined by the number of admissible sources for given cell-loss probability, buffer size and trunk bandwidth, and, finally, compare with results from simulations, which are driven by actual data from coders. The results are surprisingly close. Our bounds are based on large deviations theory. The main assumption is that the sources are Markovian and time-reversible. Our approximation to the steady-state buffer distribution is called Chenoff-dominant eigenvalue since one parameter is obtained from Chernoffs theorem and the other is the system's dominant eigenvalue. Fast, effective techniques are given for their computation. In our application we process the output of variable bit rate coders to obtain DAR(1) source models which, while of high dimension, require only knowledge of the mean, variance, and correlation. We require cell-loss probability not to exceed 10/sup -6/, trunk bandwidth ranges from 45 to 150 Mb/s, buffer sizes are such that maximum delays range from 1 to 60 ms, and the number of coder-sources ranges from 15 to 150. Even for the largest systems, the time for analysis is a fraction of a second, while each simulation takes many hours. Thus, the real-time administration of admission control based on our analytic techniques is feasible.>

A new approach to determining the admissibility of variable bit rate (VBR) traffic in buffered digital networks is developed. In this approach all traffic presented to the network is assumed to have been subjected to leaky-bucket regulation, and extremal, periodic, on-off regulated traffic is considered; the analysis is based on fluid models. Each regulated traffic stream is allocated bandwidth and buffer resources which are independent of other traffic. Bandwidth and buffer allocations are traded off in a manner optimal for an adversarial situation involving minimal knowledge of other traffic. This leads to a single-resource statistical-multiplexing problem which is solved using techniques previously used for unbuffered traffic. VBR traffic is found to be divisible into two classes, one for which statistical multiplexing is effective and one for which statistical multiplexing is ineffective in the sense that accepting small losses provides no advantage over lossless performance. The boundary of the set of admissible traffic sources is examined, and is found to be sufficiently linear that an effective bandwidth can be meaningfully assigned to each VBR source, so long as only statistically-multiplexable sources are considered, or only nonstatistically-multiplexable sources are considered. If these two types of sources are intermixed, then nonlinear interactions occur and fewer sources can be admitted than a linear theory would predict. A qualitative characterization of the nonlinearities is presented. The complete analysis involves conservative approximations; however, admission decisions based on this work are expected to be less overly conservative than decisions based on alternative approaches.>

Statistical multiplexing with loss priorities is a central element in ATM-based B-ISDN. Cell priorities arise from the marking schemes employed by the access regulators to identify excess cells, which are dropped during periods of congestion, Also, in real time applications, such as hierarchically coded voice and video, cells are assigned priorities which correspond to their importance to service quality, so that when congestion occurs only the least important are dropped. The authors present a stochastic fluid model of statistical multiplexing with loss priorities. Each Markov modulated fluid source generates streams of different priorities. The burstiness of each stream and the correlation between the priority streams are captured in the mode. The loss priority is implemented by selectively discarding cells of certain priority classes when the buffer content exceeds a corresponding threshold. To handle high dimensional source models, the authors develop an algebraic theory for the efficient computation of the spectrum of the statistical multiplexing system, which generalizes previous results for on-off sources. It is shown that to obtain the solution of the statistical multiplexing problem with J priority classes, J different 1-class problems need to be solved, together with a system of linear equations which describe the behavior of the stationary distribution at the thresholds. The numerical results demonstrate the manner in which i) the threshold level controls the tradeoff between delay of higher priority cells and the loss probability of lower priority cells, and ii) the buffer size controls the loss probability of higher priority cells.>

A prime instrument for controlling congestion in high-speed broadband ISDN (BISDN) networks is admission control, which limits call and guarantees a grade of service determined by delay and loss probability in the multiplexer. It is shown, for general Markovian traffic sources, that it is possible to assign a notational effective bandwidth to each source which is an explicitly identified, simply computing quantity with provably correct properties in the natural asymptotic regime of small loss probabilities. It is the maximal real eigenvalue of a matrix which is directly obtained from the source characteristics and the admission criterion, and for several sources it is simply additive. Both fluid and point process models are considered, and parallel results are obtained. Numerical results show that the acceptance set for heterogeneous classes of sources is closely approximated and conservatively bounded by the set obtained from the effective bandwidth approximation.>

A prime instrument for controlling congestion in a high-speed network is admission control, which limits calls and guarantees a grade of service determined by delay and loss probability in the multiplexer. It is shown that for general Markovian traffic sources it is possible to assign a notional effective bandwidth to each source that is an explicitly identified, simply computed quantity with provably correct properties in the natural asymptotic regime of small loss probabilities. It is the maximal real eigenvalue of a matrix that is directly obtained from the source characteristics and the admission criterion, and for several sources it is simply additive. Both fluid and point process models are considered. Numerical results show that the acceptance set for heterogeneous classes of sources is closely approximated and conservatively bounded by the set obtained from the effective bandwidth approximation. The bandwidth-reducing properties of the leaky bucket regulator are exhibited numerically.>

The authors give the complete solution for a stochastic fluid model of statistical multiplexing with loss priorities in asynchronous transfer mode (ATM)-based broadband ISDN. In this model each Markov modulated fluid source generates priority and marked cell streams which are bursty, mutually correlated, and periodic during bursts. The output of many such sources is buffered and multiplexed for transmission. The loss priority is implemented by selectively discarding marked cells when the buffer content exceeds a threshold level. The equilibrium state distribution exhibits jumps, a feature not existent in prior fluid models. The computational complexity for two-state sources is dominated by a single system of linear equations of dimension equal to twice the number of sources, in particular, the complexity is independent of buffer size. The complete delay distribution for each traffic class is obtained. Numerical results are given. The analysis is generalized to several priority classes of traffic.>