High Bandwidth Utilization Research Articles

Optimized Tiny Machine Learning and Explainable AI for Trustable and Energy-Efficient Fog-Enabled Healthcare Decision Support System

The Internet of things (IoT)-based healthcare decision support system plays a crucial role in modern medicine, especially with the rise in chronic illnesses and an aging population necessitating continuous remote health monitoring. Current healthcare decision support systems struggle to deliver timely and accurate decisions with minimal latency due to limited real-time healthcare data and inefficient computational resources. There is a critical need for an optimized, energy-efficient machine learning model that reliably supports remote health monitoring within IoT and fog computing environments. Our study proposes an Optimized Tiny Machine Learning (TinyML) and Explainable AI (XAI) binary classification model for a trustable and energy-efficient healthcare decision support system, leveraging fog computing to optimize performance. The fog-based approach improves response times and enhances bandwidth usage, addressing critical needs such as reduced latency, higher bandwidth utilization, and decreased packet loss. To further improve efficiency, we incorporate the innovative mLZW data compression technique, significantly enhancing data communication efficiency and reducing response time to critical health alerts. However, limited real-time healthcare data records challenge machine learning classification performance. By implementing a TinyML algorithm, our system demonstrates superior performance to other machine learning models. The proposed optimized TinyML model achieves an impressive F1 score of 0.93 for health abnormalities detection, emphasizing its robustness and effectiveness. This paper highlights the potential of TinyML and XAI in delivering robust, trustworthy, and energy-aware healthcare solutions, making significant contributions toward effective remote health monitoring and decision support in fog-enabled IoT networks.Graphical abstract

Performance Evaluation of TCP BBRv3 in Networks with Multiple Round Trip Times

The Transmission Control Protocol (TCP) serves as a cornerstone mechanism for implementing Congestion Control (CC) across the Internet. Designing a solution that provides high bandwidth utilization and mitigates the phenomenon of bufferbloat across a spectrum of diverse scenarios poses a considerable challenge. The introduction of Bottleneck Bandwidth and Round Trip propagation time (BBR) in 2016 marked a significant shift in congestion control methodology. Its improved performance and adaptability contributed to the initial acclaim and widespread interest that it received.. Unlike most currently used CCs, it operates around Kleinrock’s optimal point, thus offering high throughput even in lossy networks while preventing buffer saturation. Unfortunately, it quickly became evident that BBR was unable to fairly share bandwidth with flows characterized by different path delays, as well as loss-based CCs. In response, Google recently introduced a third iteration to address these shortcomings. This study explores the performance of BBRv3 across a wide range of scenarios, thereby considering different buffer sizes and paths with varying Round Trip Times (RTTs), and it evaluates its superiority over its predecessors. Through extensive simulations, this work assesses whether BBRv3 can finally play fair with other bandwidth contenders, which is a critical consideration given the widespread deployment of BBR. The framework is publicly available to facilitate additional validation and ensure the reproducibility of the study’s findings. The results indicate that while BBRv3 demonstrates enhanced fairness towards loss-based CC algorithms, it struggles when competing against other BBR flows, especially in multi-RTT networks, thus falling short even when compared to the initial version.

Hierarchical Cross Traffic Scheduling Based on Time-Aware Shapers for Mobile Time-Sensitive Fronthaul Network

To solve the problem of jitter and low network throughput caused by the impact of background flows on IQ traffic in mobile fronthaul network, this paper proposed a new scheduling model for background flows, named hierarchical crossover traffic scheduling mechanism based on time-aware shaper (HC-TAS) by improving the traditional counterpart. Then, in this new model, we designed an inbound scheduling algorithm based on frame length matching and an outbound scheduling algorithm based on queue status, making sure that smaller data frames will not be blocked by large data frames. This greatly improves the utilization of timeslots in the scheduling process and reduces the jitter impact of background flows. To verify its performance, we conducted experiments in a simulated fronthaul network conforming to IEEE 802.1CM. The experimental results show that, under the condition that the jitter is guaranteed to be zero, compared with two mainstream scheduling schemes, Comb-FITting and TAS + Preemption, our proposed scheme can achieve lower maximum end-to-end delay and higher link utilization. The proposed HC-TAS meets the requirements of low jitter and high bandwidth utilization in 5G fronthaul network, and the research results provide a technical basis for the application and development of general time-sensitive networks as well.

SDN-Based Congestion Control and Bandwidth Allocation Scheme in 5G Networks

5G cellular networks are already more than six times faster than 4G networks, and their packet loss rate, especially in the Internet of Vehicles (IoV), can reach 0.5% in many cases, such as when there is high-speed movement or obstacles nearby. In such high bandwidth and high packet loss network environments, traditional congestion control algorithms, such as CUBIC and bottleneck bandwidth and round-trip propagation time (BBR), have been unable to balance flow fairness and high performance, and their flow rate often takes a long time to converge. We propose a congestion control algorithm based on bottleneck routing feedback using an in-network control mode called bottleneck routing feedback (BRF). We use SDN technology (OpenFlow protocol) to collect network bandwidth information, and BRF controls the data transmission rate of the sender. By adding the bandwidth information of the bottleneck in the option field in the ACK packet, considering the flow fairness and the flow convergence rate, a bandwidth allocation scheme compatible with multiple congestion control algorithms is proposed to ensure the fairness of all flows and make them converge faster. The performance of BRF is evaluated via Mininet. The experimental results show that BRF provides higher bandwidth utilization, faster convergence rate, and fairer bandwidth allocation than existing congestion control algorithms in 5G cellular networks.

Design Synthesis and Optimization Strategy for Low Delay and High Bandwidth Utilization in Time-Sensitive Networking

Design Synthesis and Optimization Strategy for Low Delay and High Bandwidth Utilization in Time-Sensitive Networking

FMPTCP: Achieving High Bandwidth Utilization and Low Latency in Data Center Networks

FMPTCP: Achieving High Bandwidth Utilization and Low Latency in Data Center Networks

DECC: Achieving Low Latency in Data Center Networks With Deep Reinforcement Learning

Data Center Networks (DCNs) suffer from synchronized bursts for segmentation and aggregation modes, leading to buffer overflows at switches and increasing network delay. To overcome this problem, some congestion control algorithms like DCTCP use Explicit Congestion Notification (ECN) to notify in-network congestion and reduce switch buffer occupancy. However, the traditional Additive Increase Multiplicative Decrease (AIMD) method causes high fluctuation of round-trip time (RTT) in DCNs. Some intelligent congestion control algorithms designed for Internet can achieve great flexibility, but are not applicable in DCNs for a lack of accurate congestion feedback. In this paper, we analyze the deficiencies of utilizing RTT as congestion signals and the applicability of learning algorithms in DCNs. Then, we propose DECC, a smart TCP congestion control algorithm for DCNs, which combines Deep Reinforcement Learning (DRL) with ECN to achieve high bandwidth utilization as well as low queuing delay. DECC fully utilizes precise in-network feedback and formulates several QoS requirements to a multi-objective function. Meanwhile, it decouples cwnd adjustment with DRL decision making to gradually learn the optimal congestion control policy in real-time. We evaluate the performance of DECC in various scenarios. Simulation results show that DECC can reduce the queue length at bottleneck switches by more than 50% compared to DCTCP, while maintaining high bandwidth utilization and reducing Flow Completion Time (FCTs) under burst traffic.

Design analysis and optimization of chaotic systems for duration and carrier parallel index modulation

With the rapid advancement in communication technology, the demand for high speed and efficient transmission has increased significantly. Traditional communication systems are unable to meet these demands, and thus, this paper proposes a new correlated delay shift keying system for duration and carrier parallel index modulation (DCIM-CDSK) that employs duration and carrier parallel index modulation. The DCIM-CDSK system utilizes index mapping to transmit extra bits, while group mapping the timeslots on the carrier. By using spare timeslots in each group to transmit modulated information bits, the system achieves high-speed transmission with high efficiency. Additionally, the system introduces a noise reduction method to optimize the information-carrying signal to improve its noise immunity. The theoretical bit error rate (BER) equations is derived in the paper for the DCIM-CDSK system under Gaussian and Rayleigh fading channels, and compared them with practical simulations to validate the theoretical derivation. The paper also analyzes the BER of the system for a single longest timeslot length and mixed longest timeslot length, respectively. The simulation results show that the proposed DCIM-CDSK system has better noise immunity and higher transmission rates than other similar systems. In today's increasingly stringent requirements for communication systems, the DCIM-CDSK system, with its high spectral and energy efficiency, provides a feasible solution for the current communication system industry to realize large user group services and high bandwidth utilization efficiency. At the same time, because of the flexibility of the system between the transmission rate and BER performance, the communication system can get a better trade-off and balance between the two contradictions of effectiveness and reliability.

A comprehensive survey of free-space optical communication – modulation schemes, advantages, challenges and mitigations

Abstract The demand for large bandwidth and high data rates in communication systems has become the main cause of the upgrade of traditional networks into free space optical (FSO) technology. FSO technology has gained significant popularity due to its easy deployment, high data rates, abundant bandwidth, enhanced security, and license-free spectrum utilization. However, the performance of FSO communication systems can be affected by certain limiting factors, such as changes in weather conditions during data transmission. To overcome these challenges and improve FSO performance, various modulation techniques are employed. This article presents a concise overview of the FSO communication system, highlighting different modulation techniques used to enhance its performance, as well as discussing its advantages, applications, and existing challenges. Some advanced modulation formats which are recently introduced in the field of FSO communication such as QPSK, DP-QPSK, QAM, and OFDM are also made part of this paper.

Dependable Virtualized Fabric on Programmable Data Plane

In modern multi-tenant data centers, each tenant desires reassuring dependability from the virtualized network fabric – bandwidth guarantee with work conservation, bounded tail latency and resilient reachability. However, the slow convergence of prior works under network dynamics and uncertainties can hardly provide the dependability for tenants. Further, state-of-the-art load balance schemes are guarantee-agnostic and bring great risks on breaking bandwidth guarantee, which is overlooked in prior works. In this paper, we propose, a dependable virtualized fabric framework which can (1) quickly detect network failure in data plane, (2) explicitly select proper paths for all flows, and (3) converge to ideal bandwidth allocation at sub-millisecond. The core idea of is to leverage the programmable data plane to build a fusion of an active edge (e.g., NIC) and an informative core (e.g., switch), where the core sends link status and tenant information to the edge via telemetry to help the latter make a timely and accurate decision on path selection and traffic admission. We fully implement with commodity SmartNICs and programmable switches. Extensive evaluations show that can keep bandwidth guarantee with high bandwidth utilization, low and bounded latency, and resilient reachability under various network scenarios with limited overhead. Application-level experiments show that can improve QPS by 2.4 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\times$</tex-math> </inline-formula> and cut tail latency by 10 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\times$</tex-math> </inline-formula> compared to the alternatives.

Efficient and fast contention resolution for AWGR based optical data center networks

A packet contention resolution scheme featuring efficient bandwidth utilization (CRSebu) is proposed for arrayed waveguide grating router (AWGR)-based optical data centers (DCNs). Optical bandwidth can be efficiently provided in this proposed scheme to adapt the burst network traffic by flexibly arranging the timeslot and wavelength. Moreover, the CRSebu scheme implements a “Dislocation Pipeline” strategy to prevent the conventional “Request-Grant” scheduling caused extra traffic waiting delay. Experimental investigations prove the CRSebu scheme can completely prevent packet contention and also improves the network performance in terms of packet loss and server-to-server latency due to the high bandwidth utilization. Numerical assessments also validate the excellent scalability of this proposed contention resolution scheme in large-scale AWGR-based DCNs. Compared with the typical OPSquare structure deploying an Optical Flow Control scheme to prevent packet contention, the CRSebu-equipped optical DCN eliminates the retransmission of conflicted packets, which significantly improves the network performance. Experimental investigations prove the proposed CRSebu strategy can completely prevent packet contention. Numerical assessments show that the network deployed with the CRSebu scheme decreases 73.6% packet loss, and 11.9% server-to-server latency compared with the static scheme.

A Multipath Data-Scheduling Strategy Based on Path Correlation for Information-Centric Networking

Information-Centric Networking (ICN) has revolutionized the manner of content acquisition by shifting the communication mode from host-centric to information-centric. Considering the existing, large amount of IP infrastructure in current networks, the new ICN architecture is proposed to be compatible with existing networks in order to reduce deployment cost. However, due to compatibility with IP networks, ICN data packets must be transmitted through the default path provided by IP routing regulations, which also limits the transmission efficiency and reliability of ICN. In order to address this issue, this paper introduces a multipath transmission method applied in ICN which takes full advantage of the functions and characteristics of ICN and builds multiple end-to-end relay paths by using the ICN routers as relay nodes. We then propose a relay-node-selection algorithm based on path correlation to minimize the impact of overlapping links. Moreover, we comprehensively calculate the path state value by combining the round-trip time and packet loss rate and propose a multipath data-scheduling algorithm based on the path state value. Simulation experiments show that the proposed method can maintain high bandwidth utilization while reducing the number of out-of-order packets.

Machine Learning in Visible Light Communication System: A Survey

With the widespread adoption of high bandwidth utilisation, visible light communication (VLC) has emerged as a potential solution to meet the demands for high-speed data communication due to its simultaneous illumination and transmission. However, numerous nonlinear distortions in VLC cause substantial signal processing challenges and diminish the system’s efficacy. VLC communication based on machine learning (ML) approaches provides a greater ability to offset the negative impacts of transceiver nonlinearity. ML is applicable to a variety of VLC challenges, including channel estimation, jitter compensation, position tracking, modulation detection, phase estimation, and security. This study provides a detailed review of several machine learning (ML) algorithms to reduce the design complexity of indoor VLC transmission, as well as ML applications in different design aspects to improve system performance. Furthermore, various applications, challenges, and future research directions based on machine learning algorithms in VLC are addressed.

Parallel Training of Pre-Trained Models via Chunk-Based Dynamic Memory Management

The pre-trained model (PTM) is revolutionizing Artificial Intelligence (AI) technology. However, the hardware requirement of PTM training is prohibitively high, making it a game for a small proportion of people. Therefore, we proposed PatrickStar system to lower the hardware requirements of PTMs and make them accessible to everyone. PatrickStar uses the CPU-GPU heterogeneous memory space to store the model data. Different from existing works, we organize the model data in memory chunks and dynamically distribute them in the heterogeneous memory. Guided by the runtime memory statistics collected in a warm-up iteration, chunks are orchestrated efficiently in heterogeneous memory and generate lower CPU-GPU data transmission volume and higher bandwidth utilization. Symbiosis with the Zero Redundancy Optimizer, PatrickStar scales to multiple GPUs on multiple nodes. % using data parallelism. The system can train tasks on bigger models and larger batch sizes, which cannot be accomplished by existing works. Experimental results show that PatrickStar extends model scales 2.27 and 2.5 times of DeepSpeed, and consistently exhibits significantly higher execution speed. PatricStar also successfully runs the 175B GPT3 training task on a 32 GPU cluster. Our code is publicly available at https://github.com/Tencent/PatrickStar.

PID Controller for IEEE 802.17 with Dead Time Compensation

ABSTRACT Resilient Packet Ring (RPR) is defined by IEEE 802.17 to achieve fairness and high bandwidth utilization. Fairness is important in RPR where fairness algorithm has two types of operation, aggressive and conservative. The current algorithm is simple; however, it becomes unstable, if the traffic is unbalanced which cause resources loss. In this paper, the control theory is applied by designing a PID controller to calculate the fair rate. The controller operates well in resolving the fairness issue, but becomes unstable if the system’s delay is huge. Therefore, the delay influence on the stability is studied. Furthermore, a Smith predictor is suggested to compensate the delay. The enhanced controller reassures ring utilization and stability.

MiCS

Existing general purpose frameworks for gigantic model training, i.e., dense models with billions of parameters, cannot scale efficiently on cloud environment with various networking conditions due to large communication overheads. In this paper, we propose MiCS, which Minimizes the Communication Scale to bring down communication overhead. Specifically, by decreasing the number of participants in a communication collective, MiCS can utilize heterogeneous network bandwidth, reduce network traffic over slower links, reduce the latency of communications for maintaining high network bandwidth utilization, and amortize expensive global gradient synchronization overhead. Our evaluation on AWS shows that the system throughput of MiCS is up to 2.89× that of the state-of-the-art large model training systems. MiCS achieves near-linear scaling efficiency, which is up to 1.27× that of DeepSpeed. MiCS allows us to train a proprietary model with 100 billion parameters on 512 GPUs with 99.4% weak-scaling efficiency, and it is able to saturate over 54.5% theoretical computation power of each GPU on a public cloud with less GPU memory and more restricted networks than DGX-A100 clusters.

Collision-free distributed MAC protocol for passive optical intra-rack data center networks

In this paper, we present a distributed medium access control (MAC) protocol and a network architecture suitable for optical intra-rack data center networks (DCNs). The intra-rack communication is performed using passive optical components, over four data wavelength division multiplexing (WDM) channels of either 40 or 100 Gbps each, keeping low power consumption. On the other hand, the inter-rack communication is performed over a separate network through upper layer routers. In this study, we focus only on the intra-rack communication. We introduce an intra-rack DCN (IR-DCN) architecture that works in the optical domain, and two IR-DCN configurations with different total nominal capacity: 160 and 400 Gbps, respectively. Also, we propose a synchronous pre-transmission coordination fair access intra-rack MAC (intra-MAC) protocol taking into account the traffic characteristics and priority classes within existing DCNs. The proposed intra-MAC protocol totally eliminates packet collisions, achieving high performance. Particularly, it reaches high bandwidth utilization even under heavy loads: 90% and 87.5% for the two IR-DCN configurations of 160 and 400 Gbps total capacity, respectively. Also, it achieves low mean end-to-end (e2e) packet delay, lower than 0.25 and 0.12 ms, respectively, providing a reliable solution for time-sensitive DCN traffic. Specifically, simulation results demonstrate that the highest priority traffic experiences e2e delay lower than 1.9 and 1.1 µs, respectively, which is sufficient for the service of the strictest delay requirements of time-sensitive cloud applications. The intra-MAC protocol is decentralized, without the need for a network controller, providing high flexibility. Our IR-DCN proposal is studied in comparison to other currently dominant intra-rack/cluster DCNs, and it achieves from 6% to 57% higher throughput and from 20% to 99% lower e2e delay at high loads. Comparatively, it is on average 80% and 68% more energy and cost efficient, respectively.

Delay-reliability-aware protocol adaption and quality of service guarantee for message queuing telemetry transport-empowered electric Internet of things

Message queuing telemetry transport has emerged as a promising communication protocol for resource-constrained electric Internet of things due to high bandwidth utilization, simple implementation, and various quality of service levels. Enabled by message queuing telemetry transport, electric Internet of things gateways adopt dynamic protocol adaptation, conversion, and quality of service level selection to realize bidirectional communication with massive devices and platforms based on heterogeneous communication protocols. However, protocol adaptation and quality of service guarantee in message queuing telemetry transport-empowered electric Internet of things still faces several challenges, such as unified communication architecture, differentiated quality of service requirements, lack of quality of service metric models, and incomplete information. In this paper, we first establish a unified communication architecture for message queuing telemetry transport-empowered electric Internet of things for adaptation and conversion of heterogeneous protocols. Second, we formulate the quality of service level selection optimization problem to minimize the weighted sum of packet-loss ratio and delay. Then, a delay-reliability-aware message queuing telemetry transport quality of service level selection algorithm based on upper confidence bound is proposed to learn the optimal quality of service level through dynamically interacting with the environment. Compared with single and fixed quality of service level selection strategies, delay-reliability-aware message queuing telemetry transport quality of service level selection can effectively reduce the weighted sum of delay and packet-loss ratio and satisfy the differentiated quality of service requirements of electric Internet of things.

Accelerating CPU-Based Sparse General Matrix Multiplication With Binary Row Merging

Sparse general matrix multiplication (SpGEMM) is a fundamental building block for many real-world applications. Since SpGEMM is a well-known memory-bounded application with vast and irregular memory accesses, considering the memory access efficiency is of critical importance for optimizing SpGEMM. Yet, the existing methods put less consideration into the memory subsystem and achieved suboptimal performance. In this paper, we thoroughly analyze the memory access patterns of SpGEMM and their influences on the memory subsystem. Based on the analysis, we propose a novel and more efficient accumulation method named BRMerge for the multi-core CPU architectures. The BRMerge accumulation method follows the row-wise dataflow. It first accesses the <i>B</i> matrix, generates the intermediate lists for one output row, and stores these intermediate lists in a consecutive memory space, which is implemented by a ping-pong buffer. It then immediately merges these intermediate lists generated in the previous phase two by two in a tree-like hierarchy between two ping-pong buffers. The architectural benefits of BRMerge are 1) streaming access patterns, 2) minimized TLB cache misses, and 3) reasonably high L1/L2 cache hit rates, which result in both low access latency and high bandwidth utilization when performing SpGEMM. Based on the BRMerge accumulation method, we propose two SpGEMM libraries named BRMerge-Upper and BRMerge-Precise, which use different allocation methods. Performance evaluations with 26 commonly used benchmarks on two CPU servers show that the proposed SpGEMM libraries significantly outperform the state-of-the-art SpGEMM libraries.

High Bandwidth Utilization DWBA Algorithm for Upstream Channel in NG-EPON

In this paper, an (ONU-load and RTTs)-based (OLR) dynamic wavelength and bandwidth allocation (DWBA) algorithm for upstream channel in next-generational Ethernet passive optical network (NG-EPON) is proposed. By proposing adaptive threshold grouping algorithm, the waste of bandwidth resources caused by massive guard timeslots and mismatch between assigned transmission window and frame size is reduced. By adjusting the Optical Network Unit (ONU) scheduling order, the idle time caused by Round-Trip Times (RTTs) is reduced. By proposing joint ONU grouping and RTT scheduling mechanism, load balance among wavelengths is achieved. By proposing a fair allocation scheme, the fairness of bandwidth granting for each ONU is ensured. Finally, by the simulation, the effectiveness of proposed algorithm is demonstrated. The simulation indicates that the rate of bandwidth utilization, the average package delay, the scheduling cycle and the network throughput in the system based on proposed algorithm have better performance.