This paper introduces NeCTAr (Near-Cache Transformer Accelerator), a 16nm heterogeneous multicore RISC-V SoC for sparse and dense machine learning kernels with both near-core and near-memory accelerators. A prototype chip runs at 400MHz at 0.85V and performs matrix-vector multiplications with 109 GOPs/W. The effectiveness of the design is demonstrated by running inference on a sparse language model, ReLU-Llama.

This paper presents non-cooperative game-theoretic transmission power control algorithm to optimize network-wide communication performances for homogeneous competitive vehicle nodes in interference-limited vehicular ad hoc network (VANET) with limited inter-node feedback signaling channels. Based on a class of self-incentive convex payoff functions, vehicular nodes employ the proposed Feedback-free Adaptive Self-regulating (FAS) control algorithm to select optimal strategies that converge to Nash equilibrium (NE) whenever system changes occur. The control algorithm is characterized with the following features: (1) it does not require inter-node feedback signaling and messaging controls; (2) each transmitting node only needs to observe the aggregate interference in the environment; (3) the game is formulated to enable self-enforcement guided by rational self-incentive. System simulation of an interference-limited VANET with a realistic vehicular mobility model is developed to evaluate NE convergence performance of the game-theoretic FAS control algorithm.

The Optical Burst-Train Switching (OBTS) network efficiently transports multi-granular subwavelength constant-bit-rate (CBR) and variable-bitrate (VBR) traffic streams over an all-optical network with sparse wavelength convertibility while minimizing access blocking of traffic streams. Sub-wavelengt

This paper considers the problem of competitive sharing of open spectrum resources between collocated spread spectrum based secondary systems. The problem is formulated as a strategic form game where the objective of each player (secondary system) is to maximize its own payoff defined in terms of re

This paper presents the dynamic integrated-services MAC (DIS-MAC) protocol designed to enable service differentiation in CDMA-based hybrid wireless mesh networks (WMNs) through an integrated medium access that allows both contention-based and contention-free scheduled access modes. DIS-MAC utilizes dynamic capacity partitioning that adjusts CDMA capacity limits assigned to the two access modes in response to changes in network traffic loading. The contention-based access is made compatible with the IEEE 802.11 standard, while the contention-free scheduled access provides differentiated delay guarantees for real-time traffic over single-hops connecting among mesh routers, clients and gateways. The scheduler enables feedback-enhanced delay target-tracking and delay-bandwidth decoupling. Numerical results show that DIS-MAC provides high resource utilizations and differentiated delay bounds in a WMN environment characterized with heterogeneous real-time and non-real-time traffic sources.

This paper focuses on sharing of open spectrum under a new sharing policy of prioritized dynamic spectrum leasing, where the designated spectrum broker dynamically allocates time-varying unused licensed spectrum resources to secondary networks of various access priorities. The proposed optimal stochastically controlled dynamic guard bandwidth (OSC-DGB) scheme enables the spectrum broker to adapt differentiated capacity limits of open spectrum for prioritized connection accesses by secondary networks in anticipation of their access request rates, which are expected to be opportunistic and non-stationary. The differences in capacity limits determine the amount of guard band reserved for the higher priority secondary networks. This spectrum adaptation problem is formulated as a finite-horizon Markov decision process to minimize service denials of secondary networks according to their spectrum access priorities, which is solved through dynamic programming. Numerical results show that OSC-DGB enables close tracking of connection blocking targets while optimizing resource utilizations.

This paper presents a novel architecture of multi-granular stream optical burst switching (MGS-OBS), which enables an all-optical network to support sub-wavelength streaming traffic of diverse bandwidth granularities with guaranteed end-to-end delay and data loss rate. It shapes a sub-wavelength stream into a periodic burst train; and allows variable bit rate (VBR) streaming traffic with same source and destination to share the reserved bandwidth through statistical multiplexing. By adopting periodic advanced scheduling of burst trains, MGS-OBS substantially decreases recursive signaling processing in supporting sub-wavelength streaming traffic. MGS-OBS employs burst train reservation blocking recovery by time-shifting a blocked burst train. Performance of MGS-OBS is analyzed in terms of wavelength utilization and data loss rate, and the stream transmission failure rate. It is found that the performance of the MGS-OBS is compared favorably with other OBS schemes that target for efficient support of bursty traffic.

To enable combined connection-level and application-level QoS support in future multi-class CDMA-based system characterized with highly non-stationary loading at both levels, this paper integrates two adaptive control mechanisms of prioritized admission and rate scheduling. The buffered adjusted multimode dynamic guard bandwidth (BAM-DGB) admission control enables connection-level prioritization via dynamic guard method with differentiated congestion controls that minimizes handoff droppings in non-stationary environment. The adaptive differentiated services medium access control (ADS-MAC) complements the admission control with the target-tracking weighted fair queuing (TT-WFQ) scheduler that adapts allocated transmission rates subject to constraints imposed by the admission control. It enables delay-bandwidth decoupling for maximum resource utilizations, and absolute delay target- tracking for balanced degradations during system congestions. Numerical results indicate that the proposed mechanisms are able to maintain consistent connection dropping probabilities and scheduling delay targets under non-stationary loading and congestion conditions.

Groupcast communication is gaining momentum due to the increasing demands of bandwidth-intensive parallel computing and collaborations. Groupcast is a multipoint-to-multipoint communication pattern, where one or more member nodes in a group are allowed to send traffic with various rates to all other members within the group. Efficient implementation of groupcast requires in-group or cross-group traffic grooming. This paper studies the application of traffic grooming to support dynamic groupcast provisioning in WDM networks with ring and mesh topologies. The design issues such as routing, in-group and cross-group traffic grooming and traffic sub-grouping are addressed.

Guard-based call admission control schemes support admission priorities based on resources sharing with differentiated resource capacity limits. To minimize deviation from call blocking/dropping targets due to nonstationary call arrival condition, dynamic guard-based schemes with predictive adaptation control adjust differentiated capacity limits according to predicted future arrival rates based on specified estimation algorithms. Existing dynamic guard admission schemes are developed under the assumption of perfect estimation, which may not be possible in a highly nonstationary environment and, thus resulting in failures to maintain targeted blocking/dropping probabilities. This paper presents the fairly adjusted multimode-dynamic guard bandwidth scheme, which is a dynamic-guard-based scheme over code-division multiple-access systems with predictive adaptation control to adapt interference-based guard loading-limits under nonstationary call arrival condition; and reactive adaptation control to counteract arrival rate estimation errors. When the predictive adaptation control policy mode is not able to maintain long-term call blocking or dropping targets due to estimation error, this will trigger reactive adaptation control policy modes that include temporary blocking (preemption) of one or more lower priority classes subject to fairness constraints to ensure lower priority classes are not preempted at all costs during estimation error recovery. Analytical and simulation results show that proposed scheme is able to provide performance guarantees in terms of dropping probabilities under nonstationary traffic arrival and imperfect arrival rate estimation.

In wavelength-division multiplexing (WDM) optical networks, multicast is implemented by constructing a light-forest, which is a set of light-trees with each light-tree rooted from the multicast source and terminated at a partition subset of the destination nodes. Quality of optical signal at each destination node is greatly affected by multicast routing scenario. To guarantee that different destinations in a multicast session can receive relatively equivalent optical signal quality, it is desirable to deliver the multicast traffic over a loss-balanced light-forest. A loss-balanced light-forest is composed of a set of light-trees with each light-tree having both bounded destination size and bounded source-destination distance. This paper studies multicast routing and wavelength assignment (MC-RWA) problem under loss-balanced light-forest constraint. The problem is formulated as an optimization model using integer linear programming (ILP). Numerical solutions to the optimization model can supply useful performance benchmarks for loss-balance-constrained optical multicast in WDM networks.

Group multicast is a generalized multicast pattern important to future high-bandwidth collaborative applications. In this paper, we investigate group multicast routing and wavelength assignment (GMC-RWA) problem in WDM networks. We state GMC-RWA problem as in-group optical approach and multicasting problem, and give its mathematic formulation. Two heuristics based on traffic source grooming are proposed with bounded costs and polynomial-time complexities. Simulations on various network topologies show the efficiency of the heuristics.

The TDMA-based dynamic guard channel scheme supports prioritized call admission control and adapts new call and handoff request admission priorities in response to nonstationary request arrivals or loading. We present an adjusted multimode dynamic guard bandwidth (AM-DGB) scheme that extends the blocked-calls-dropped dynamic guard bandwidth concept to CDMA-based systems, supporting multiple admission priority classes, based on request types and traffic classes. AM-DGB employs operation modes that temporarily block one or more lower-priority classes completely when pure dynamic guard operation is unable to keep long-term blocking rates under predefined targeted thresholds. Analytical and simulation results show that AM-DGB is able to maintain consistent performance guarantees in terms of dropping probabilities.

This paper presents the Photonic Interdomain Negotiator (PIN) as the secure global optical control plane architecture to interoperate multiple wavelength-routed network domains with incompatible local control planes. The PIN interdomain routing and wavelength assignment is characterized by constraint-based optical path-vector routing to determine end-to- end lightpath across multiple domains, and by conversion-aware wavelength discovery to resolve wavelength blocking during discovery phase. The PIN interdomain optical signaling is supported by the Robust Fast Optical Reservation Protocol (RFORP) to minimize reservation delay by employing parallel reservation processing in multiple domains along a domain-path, and to minimize wavelength blocking during reservation phase by employing localized rerouting.

Emerging lambda-Grid systems employ wavelength-routed network with optical switches to enable dynamic on-demand lightpaths with multi-gigabit rate bandwidth to interconnect shared computing clusters of user domains. User-centric lambda-Grid systems enable user domains to act as distributed connectivity providers of shared wavelength resources; and as distributed network control service providers of brokering shared wavelength resources and of provisioning scheduled lightpaths. To support user-centric lambda-Grid systems, the proposed Dynamic User-centric Switched Optical Network (DUSON) architecture employs an Intelligent Control Service Proxy (ICSP) at each user domain to support wavelength connectivity brokerage and lightpath provisioning/restoration middleware for user applications to control on-demand lightpaths. Furthermore, ICSP employs the proposed Robust Fast Optical Reservation Protocol (RFORP) during lightpath provisioning/restoration to minimize wavelength reservation blocking and delay via localized rerouting and parallel concurrent wavelength reservation processing respectively.

Existing bandwidth broker approach for centralized or hierarchical reservation control in differentiated service (DiffServ) network is appropriate for long-lived flows, and requires out-of-band signaling transport mechanism and per-flow state maintenance within the core. While persistent congestion may be minimized with bandwidth broker, transient congestion is unavoidable as per-flow reservation control is not feasible in DiffServ network. This paper proposes a novel distributed-buffer core-stateless weighted fair queuing (DC- WFQ) network architecture to support: (1) a distributed core-stateless reservation control scheme that is appropriate for short-lived and long-lived flows, and does not require the support of an explicit signaling transport mechanism; (2) a transient congestion control scheme to prevent packet loss by absorbing transient congestion with distributed buffering in upstream routers associated with the potential congestion-causing traffic. The proposed DC-WFQ network is based on extension and DiffServ adaptation of the core-stateless fair queuing (CSFQ) network, which is enabled with dynamic packet state encoding and probabilistic packet dropping to approximate fair queuing among flows. The CSFQ network is adapted to support weighted fair queuing among flow-aggregates.

For 3G mobile radio systems, the combined time and code division multiple access (TD-CDMA) is well suited for wireless local area networks. An optimized-capacity medium access control (OC-MAC) protocol is proposed to maximize transport capacity by supporting per-slot bit error rate (BER) and traffic rate scheduling while satisfying multimedia quality-of-service (QoS). An interference-based prioritized call admission control (IP-CAC) is proposed to maintain QoS under the effect of mobile handoffs by supporting adaptive admission priorities for handoffs and new calls of multiple traffic classes.

High speed bulk data transfer is an important part of many data-intensive scientific applications. This paper describes an aggressive bulk data transfer scheme, called Reliable Blast UDP (RBUDP), intended for extremely high bandwidth, dedicated- or Quality-of-Service-enabled networks, such as optically switched networks. This paper also provides an analytical model to predict RBUDP's performance and compares the results of our model against our implementation of RBUDP. Our results show that RBUDP performs extremely efficiently over high speed dedicated networks and our model is able to provide good estimates of its performance.

The General Packet Radio Service (GPRS) offers performance guaranteed packet data services to mobile users. A dynamic adaptive guaranteed quality-of-service (QoS) provisioning scheme is proposed over GPRS wireless mobile links via the guaranteed QoS media access control (GQ-MAC) protocol and the accompanying adaptive prioritized-handoff call admission control (AP-CAC) protocol to maintain QoS guarantees under the effect of mobile handoffs. The GQ-MAC protocol supports bounded access delay and packet-loss probability for respective delay and loss sensitive traffic, and dynamic adaptive resource allocation for bursty traffic. The AP-CAC protocol provides dynamic adaptive prioritized admission by differentiating handoff requests of different traffic classes with higher admission priorities over new calls via the dynamic multiple guard channels scheme, which adapts the channel capacity limits reserved for the multiple handoff request classes in each radio cell based on the current estimates of their arrival rates derived from the current number of ongoing calls in neighboring radio cells and the mobility pattern.

The general packet radio service (GPRS) offers performance guaranteed packet data services to mobile users. This paper presents a guaranteed quality-of-service (QoS) provisioning scheme over GPRS wireless mobile links by proposing a guaranteed QoS media access control (GQ-MAC) protocol and an accompanying prioritized handoff call admission control (PH-CAC) protocol to maintain QoS guarantees under the effect of mobile handoffs. The GQ-MAC protocol supports bounded channel access delay for delay-sensitive traffic, bounded packet loss probability for loss-sensitive traffic, and dynamic adaptive resource allocation for bursty traffic. The PH-CAC protocol provides prioritized admission by differentiating handoff requests with different higher admission priorities over new calls via a multiple guard channel scheme.

The growing demand for real-time broadband multimedia communications and Internet access has accelerated the evolution to the next generation Internet which support real-time performance guarantees through the integrated service architecture, and fast packet forwarding through multiprotocol label switching. On the other hand, there is also a growing demand to run familiar Internet application in the wireless mobile environment. This paper proposes mobile integrated services consisting of mobile guaranteed and controlled-load services to support intolerant and tolerant real-time multimedia applications, and a wireless-aware middleware to counteract the bandwidth and latency limitation of the wireless link.

The paper discusses the quality of service (QoS) requirements for each traffic class in a wireless environment and problems in implementing the QoS requirements in the CDMA based Universal Mobile Telecommunication Service (UMTS) network. In UMTS networks, the medium access control (MAC)/link access control (LAC) sublayer of the OSI layer 2 will be used for priority handling, QoS monitoring, and radio resource management for traffic flows from several users. The paper presents some suggestions to realize QoS requirements in the MAC/LAC sublayer along with restrictions imposed by the W-CDMA air interface.

The next generation Internet based on multiprotocol label switching (MPLS) and integrated services architecture will likely emerge as a prominent core network for future generation wireless mobile systems. This paper proposes a lightweight near-seamless handoff scheme to support terminal mobility over a MPLS-based Internet supporting performance guarantees. The handoff scheme is supported by: (1) a novel mobile label switched tree (MLST) architecture which provides dynamic mobility tracking and adaptive traffic rerouting during handoffs; and (2) an extended RSVP to enable fast and robust resource reservation during handoffs.

In future personal communications networks (PCNs) supporting network-wide handoffs, new and handoff requests will compete for connection resources in both the mobile and backbone networks. Forced call terminations due to handoff call blocking are generally more objectionable than new call blocking. The previously proposed guard channel scheme for radio channel allocation in cellular networks reduces handoff call blocking probability substantially at the expense of slight increases in new call blocking probability by giving resource access priority to handoff calls over new calls in call admission control. While the effectiveness of a fixed number of guard channels has been demonstrated under stationary traffic conditions, with nonstationary call arrival rates in a practical system, the achieved handoff call blocking probability may deviate significantly from the desired objective. We propose a novel dynamic guard channel scheme which adapts the number of guard channels in each cell according to the current estimate of the handoff call arrival rate derived from the current number of ongoing calls in neighboring cells and the mobility pattern, so as to keep the handoff call blocking probability close to the targeted objective while constraining the new call blocking probability to be below a given level. The proposed scheme is applicable to channel allocation over cellular mobile networks, and is extended to bandwidth allocation over the backbone network to enable a unified approach to prioritized call admission control over the ATM-based PCN.

In future B-ISDN and PCN environments, an increasing number of network control services, such as dynamic multiparty conferences, mobile call handoff, etc., will require realtime connection rerouting services not currently supported by the B-ISDN signaling network architecture based on the SS7 CCS network and its transaction protocols (TCAP). To support these services, we propose a novel signaling network architecture employing the following new signaling transport protocols in the signaling ATM adaptation layer: (1) associated-CCS employing source-routing unicast and multicast to provide fast distributed signaling transport for realtime distributed transactions, and (2) inband signaling required for synchronization data transport, and the following new transaction protocols in the TCAP layer: (1) robust fast reservation of communication resources, and (2) packet-ordering synchronization to minimize traffic descriptions.

The next generation personal communications network (PCN) will likely consist of an ATM-based B-ISDN backbone interconnecting wireless mobile networks. Support of user terminal mobility, particularly the capability for fast and seamless handoffs, over the B-ISDN is an expected requirement that is not currently met. We propose extensions to the B-ISDN user transport and signaling network architectures and signaling protocols to meet these requirements. The proposed extensions optimizes the handoff delay over the B-ISDN while minimizing the resources allocation, chances of handoff failure, and the call processing load in the intelligent network (IN). The network architecture extensions are backward compatible to current B-ISDN standards and implementations.